Showing 1–50 of 144 results for author: Andreas, J

Modular Cognitive Architecture Emerges in Large Language Models

Authors: Pengrui Han, Jacob Andreas, Evelina Fedorenko, Andrea Gregor de Varda

Abstract: The human brain exhibits a striking degree of functional specialization, with distinct networks supporting language, formal reasoning, reasoning about other minds, and reasoning about the physical world. Is this modular organization a fundamental principle of how intelligent systems must be built, or an evolutionary accident specific to biological brains? Here, we test whether a similar organizati… ▽ More The human brain exhibits a striking degree of functional specialization, with distinct networks supporting language, formal reasoning, reasoning about other minds, and reasoning about the physical world. Is this modular organization a fundamental principle of how intelligent systems must be built, or an evolutionary accident specific to biological brains? Here, we test whether a similar organization emerges in Large Language Models--another class of intelligent systems created through a very different optimization process. Using circuit analyses across N=46 tasks spanning four cognitive domains (language, formal reasoning, social reasoning, physical reasoning), we find that LLMs develop a modular architecture that mirrors the human brain: tasks drawing on the same network in humans recruit overlapping neurons in LLMs, whereas tasks drawing on different networks recruit distinct neurons. The convergent emergence of modularity in brains and neural networks suggests that it may be a fundamental property of intelligent systems. △ Less

Submitted 26 June, 2026; originally announced August 2026.

Comments: https://pengrui-han.github.io/LLM_Modularity_Page/

Position: It's Time to Optimize LLMs for Self-Consistency

Authors: Itamar Pres, Belinda Z. Li, Laura Ruis, Zifan Carl Guo, Keya Hu, Mehul Damani, Isha Puri, Ekdeep Singh Lubana, Jacob Andreas

Abstract: Despite ever-increasing sophistication in language model (LM) pre- and post-training pipelines, many important failures persist: models overcondition on user framing ("sycophancy"), exhibit incomplete logical generalization, and produce confident but incorrect responses. We argue that these failures arise from a modeling assumption permeating all aspects of the pipeline: that behavior can be speci… ▽ More Despite ever-increasing sophistication in language model (LM) pre- and post-training pipelines, many important failures persist: models overcondition on user framing ("sycophancy"), exhibit incomplete logical generalization, and produce confident but incorrect responses. We argue that these failures arise from a modeling assumption permeating all aspects of the pipeline: that behavior can be specified and evaluated independently on single-output pairs. Many model failures are difficult, if not impossible, to detect without reasoning about relationships between a model's responses across inputs. In this position paper, we propose self-consistency as a framework for understanding these failures. We first observe that a wide variety of techniques designed to improve specific aspects of LM behavior-targeting properties as diverse as adversarial robustness and factual coherence-can be understood as special cases of a common "consistency optimization" procedure and addressed with a standard set of optimization tools. We next outline a set of new model properties that could be achieved by optimizing for consistency, and conclude with a discussion of what it would mean to develop generally consistent LMs, including the capabilities they would enable and the objections they raise. △ Less

Submitted 31 July, 2026; originally announced August 2026.

Comments: Accepted at the 43rd International Conference on Machine Learning (ICML 2026), Position Paper Track

A Roadmap to Impactful Pluralistic Alignment Research

Authors: Elinor Poole-Dayan, Jillian Fisher, Atoosa Kasirzadeh, Jacob Andreas, Mitchell Gordon, Michiel A. Bakker

Abstract: Pluralistic value alignment---the goal of building AI systems that represent and serve diverse human values and perspectives---has emerged as an active research agenda. Yet, there's no public evidence that it has shaped the training or evaluation of the AI systems people actually use. We audit the public behavior documents and evaluations of frontier labs, finding none name pluralism as a goal, an… ▽ More Pluralistic value alignment---the goal of building AI systems that represent and serve diverse human values and perspectives---has emerged as an active research agenda. Yet, there's no public evidence that it has shaped the training or evaluation of the AI systems people actually use. We audit the public behavior documents and evaluations of frontier labs, finding none name pluralism as a goal, and as of this writing, no clear indication that production models are explicitly trained or tested for it. This goes against the primary motivations and goals of pluralistic alignment, which revolve around making a positive difference in the models serving billions of users worldwide. We argue that the pluralistic alignment research community should focus on supporting impact and adoption in deployed, widely-used AI systems. We provide evidence for the adoption problem, present three main reasons behind it, and discuss three corresponding areas for future research to address it: 1. The primary justifications for pluralistic alignment so far have been normative or speculative. We need studies showing empirically how pluralistic AI benefits users or society. 2. The pluralistic alignment research community has not settled when pluralistic behavior is warranted or what pluralism ideally looks like in practice. We need to establish a concrete goal for developers to operationalize. 3. Current methods trade off against other desiderata of LLMs in ways that are largely unmeasured, and existing metrics are not "hill-climbable." We need trade-off-aware evaluations and methods that meet the requirements of production systems. This paper serves as a collective call to action for the pluralistic alignment researchers: progress requires moving beyond normative justification toward empirical foundations, a concrete account of ideal pluralistic behavior, and practical methods and evaluations built for adoption. △ Less

Submitted 24 July, 2026; originally announced July 2026.

From Weights to Words: Expressing and Editing Preference Model Inferences in Natural Language

Authors: Zachary Wojtowicz, Ayush Nayak, Jacob Andreas

Abstract: The growing use of statistical learning algorithms to infer human preferences from high-dimensional choice data runs up against a fundamental challenge: choice alternatives typically differ in many ways simultaneously, so it is generally unclear which factors actually drove an observed decision and should be credited as preferences. Compounding this problem, the opacity of these methods leaves hum… ▽ More The growing use of statistical learning algorithms to infer human preferences from high-dimensional choice data runs up against a fundamental challenge: choice alternatives typically differ in many ways simultaneously, so it is generally unclear which factors actually drove an observed decision and should be credited as preferences. Compounding this problem, the opacity of these methods leaves human operators unable to inspect, contest, or correct models when they err. We introduce \emph{weights to words}, a method that takes a dataset of choice problems as input and automatically discovers a collection of domain-relevant preference dimensions, each described in natural language and paired with a vector in the model's representational space. These dimensions address both under-determination and opacity: they can be applied to concentrate attribution on a small set of meaningful factors, and they can externalize the model's inferences in natural language so that users can inspect and edit them in real time. We first qualitatively illustrate the method's versatility on four diverse domains: moral dilemmas, movies, wines, and free-form LLM responses. We then report two pre-registered human-subjects experiments, on moral dilemmas ($N=450$) and movie selection ($N=449$), that demonstrate its benefits for learning preference models: (1) regularizing a preference model toward the learned basis increases prediction accuracy on held-out choices, and (2) incorporating participants' structured edits further improves accuracy. In head-to-head comparisons, participants prefer the method's inferred preference profiles and endorse its predictions as more accurate. △ Less

Submitted 24 June, 2026; originally announced July 2026.

Comments: 42 pages, 22 figures, 14 tables; main text 11 pages, remainder appendices

Right in the Right Way: LM Training with Verifiable Rewards and Human Demonstrations

Authors: Mehul Damani, Isha Puri, Idan Shenfeld, Jacob Andreas

Abstract: RL with verifiable rewards (RLVR) has emerged as a powerful paradigm for training LMs on tasks with well-defined success metrics, such as code generation and mathematical reasoning. However, current RLVR methods optimize only what can be objectively scored, often neglecting subjective, non-verifiable aspects of human-like outputs, such as style and structure. This limitation leads to well-document… ▽ More RL with verifiable rewards (RLVR) has emerged as a powerful paradigm for training LMs on tasks with well-defined success metrics, such as code generation and mathematical reasoning. However, current RLVR methods optimize only what can be objectively scored, often neglecting subjective, non-verifiable aspects of human-like outputs, such as style and structure. This limitation leads to well-documented failure modes such as diversity collapse, unnatural-sounding responses, and reward hacking. We propose an adversarial generator-discriminator framework that augments verifiable rewards with a learned signal from human demonstrations. A generator model is trained using RL to maximize both task accuracy and an adversarial reward derived from a discriminator. The discriminator, trained alongside the generator policy, learns to distinguish human-written outputs from model-generated ones. The discriminator serves as a learned proxy for the human output distribution, providing feedback on aspects of generation that are difficult to formalize as scalar rewards. Across diverse domains, including bug fixing and open-ended generation, our approach consistently improves non-verifiable properties while preserving the accuracy gains of RLVR. In bug fixing, our method produces solutions with significantly lower edit distance compared to RLVR baselines while matching end performance. In story generation, our method significantly improves win rate while producing stories that are diverse and more human-like. And in a simple reward hacking benchmark, our method nearly eliminates model misbehavior while maintaining high benchmark scores. Together, these results show that our approach bridges RL and SFT, offering a scalable path toward jointly optimizing the verifiable and non-verifiable properties of a task. △ Less

Submitted 1 July, 2026; originally announced July 2026.

Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervision

Authors: Zifan Carl Guo, Laura Ruis, Jacob Andreas, Belinda Z. Li

Abstract: When does training language models (LMs) to generate explanations of their predictions yield faithful introspection, rather than superficial imitation? We study LMs trained to explain which features of their inputs influenced their behavior, using models' counterfactual behavior on modified inputs as supervision. Surprisingly, we find that LMs trained on fixed counterfactual explanations derived f… ▽ More When does training language models (LMs) to generate explanations of their predictions yield faithful introspection, rather than superficial imitation? We study LMs trained to explain which features of their inputs influenced their behavior, using models' counterfactual behavior on modified inputs as supervision. Surprisingly, we find that LMs trained on fixed counterfactual explanations derived from earlier checkpoints of themselves, or even from behaviorally similar models in different families, frequently produce explanations more faithful to their own current behaviors than to those of their training targets. This "introspective" coupling between LM explanations and behaviors occurs when training explanations remain sufficiently correlated with current behaviors over the course of training, even as behaviors themselves shift. We also show that introspective coupling tracks behavior shifts: when explanation training is provided concurrently with other post-training objectives, explanations track those shifts without requiring updated supervision. This phenomenon appears in multiple tasks, including sycophancy and refusal, and is robust to label noise. Overall, our results show that even fixed datasets of counterfactual explanations can provide scalable and generalizable post-training signal for introspection. △ Less

Submitted 30 June, 2026; originally announced June 2026.

Comments: 32 pages, 19 figures

Explaining Attention with Program Synthesis

Authors: Amiri Hayes, Belinda Z Li, Jacob Andreas

Abstract: A longstanding goal of research on interpretable deep learning is to replace opaque neural computations with human-meaningful symbolic descriptions. In this paper, we propose an approach for approximating the behavior of components of deep networks with executable programs. We focus on attention heads in transformer language models. For a given head, we first compute its associated attention matri… ▽ More A longstanding goal of research on interpretable deep learning is to replace opaque neural computations with human-meaningful symbolic descriptions. In this paper, we propose an approach for approximating the behavior of components of deep networks with executable programs. We focus on attention heads in transformer language models. For a given head, we first compute its associated attention matrices on a collection of randomly selected training examples. Next, we prompt a pre-trained language model with a summary of these matrices, and instruct it to generate a set of Python programs that can reproduce the associated attention patterns given only text from the input sentence. Finally, we re-rank programs according to how well our final set of programs predict behavior on held-out inputs. We demonstrate that a set of fewer than 1,000 such generated programs can reproduce the attention patterns of heads in GPT-2, TinyLlama-1.1B, and Llama-3B, achieving an average Intersection-over-Union similarity above 75% on TinyStories. Moreover, the best-fit programs can replace neural attention heads without substantially affecting model behavior: replacing 25% of attention heads with programmatic surrogates across the three models incurs only a 16% average perplexity increase, while maintaining performance on a variety of downstream question answering benchmarks. This work contributes a scalable pipeline for reverse-engineering attention heads in transformer models using human-readable, executable code, advancing a path toward symbolic transparency in neural models. △ Less

Submitted 29 June, 2026; v1 submitted 17 June, 2026; originally announced June 2026.

VISUALSKILL: Multimodal Skills for Computer-Use Agents

Authors: Ziyan Jiang, Li An, Yujian Liu, Jiabao Ji, Qiucheng Wu, Jacob Andreas, Yang Zhang, Shiyu Chang

Abstract: Computer-use agents (CUAs) approach human-level performance on standardised benchmarks but still struggle on long-horizon tasks and unseen software. Existing skill libraries address this with reusable skills, but represent the skill artifact as text only, despite the visual nature of GUI interaction. We propose VISUALSKILL: a hierarchical multimodal skill, tailored to each target application and o… ▽ More Computer-use agents (CUAs) approach human-level performance on standardised benchmarks but still struggle on long-horizon tasks and unseen software. Existing skill libraries address this with reusable skills, but represent the skill artifact as text only, despite the visual nature of GUI interaction. We propose VISUALSKILL: a hierarchical multimodal skill, tailored to each target application and organised as a central index over per-topic files, which the agent consumes through a load_topic MCP tool that fetches the relevant topic's text and figures on demand. We construct each skill with a two-stage pipeline that combines authored documentation with live-application UI exploration. On two CUA benchmarks, CUA-World and OSExpert-Eval, a Claude Code CLI agent backed by Claude Opus 4.6 reaches an average score of 0.456 with VISUALSKILL, a +15.3 point absolute lift over the no-skill baseline (0.303). Against a matched text-only skill that is generated from the same source content and differs from VISUALSKILL only in modality, VISUALSKILL yields a further +8.3 point absolute gain over the matched text-only skill (0.373 vs. 0.456), providing direct evidence that retaining visual figures in the skill artifact, rather than verbalizing them away, helps the agent both identify UI elements and verify workflow state after each action. Our code is available at https://github.com/XMHZZ2018/VisualSkills. △ Less

Submitted 16 June, 2026; originally announced June 2026.

Self-CTRL: Self-Consistency Training with Reinforcement Learning

Authors: Itamar Pres, Laura Ruis, Melat Ghebreselassie, Belinda Z. Li, Jacob Andreas

Abstract: Language models (LMs) that faithfully describe their own behavior can more easily be audited, understood, and trusted by users. This paper describes Self-Consistency Training with Reinforcement Learning (Self-CTRL), a method that optimizes for consistency between a LM's self-explanations and behavior on related inputs by updating explanations to better predict behavior or updating behavior to bett… ▽ More Language models (LMs) that faithfully describe their own behavior can more easily be audited, understood, and trusted by users. This paper describes Self-Consistency Training with Reinforcement Learning (Self-CTRL), a method that optimizes for consistency between a LM's self-explanations and behavior on related inputs by updating explanations to better predict behavior or updating behavior to better match explanations. We apply our method in two domains. First, we study a formal probabilistic reasoning task in which LMs must learn to imitate a family of biased samplers and evaluated on their ability to report the associated biases. We find that consistency training improves the correlation between self-reported and behaviorally-measured latent biases from $R^2=0.24$ to $R^2=0.64$ on a set of held-out distributions, matching the generalization of direct ground-truth supervision. Second, we study a constitutional AI domain in which LMs must describe when they will refuse or comply with user requests. Here, Self-CTRL produces rules that faithfully describe the model's behavior on held-out requests, improving the refusal predictions of a third-party auditor model from $36\%$ to $92\%$. In the other direction, behavior updates improve alignment, reducing HarmBench failure rate from $15.0\%$ to $0.5\%$ without substantially increasing refusal on harmless prompts. By aligning explanations and behavior, our work provides a general recipe for training AI models to be safer, more transparent, and more controllable. △ Less

Submitted 16 June, 2026; originally announced June 2026.

Comments: 34 pages, 12 figures, includes appendices

ELF: Embedded Language Flows

Authors: Keya Hu, Linlu Qiu, Yiyang Lu, Hanhong Zhao, Tianhong Li, Yoon Kim, Jacob Andreas, Kaiming He

Abstract: Diffusion and flow-based models have become the de facto approaches for generating continuous data, e.g., in domains such as images and videos. Their success has attracted growing interest in applying them to language modeling. Unlike their image-domain counterparts, today's leading diffusion language models (DLMs) primarily operate over discrete tokens. In this paper, we show that continuous DLMs… ▽ More Diffusion and flow-based models have become the de facto approaches for generating continuous data, e.g., in domains such as images and videos. Their success has attracted growing interest in applying them to language modeling. Unlike their image-domain counterparts, today's leading diffusion language models (DLMs) primarily operate over discrete tokens. In this paper, we show that continuous DLMs can be made effective with minimal adaptation to the discrete domain. We propose Embedded Language Flows (ELF), a class of diffusion models in continuous embedding space based on continuous-time Flow Matching. Unlike existing DLMs, ELF predominantly stays within the continuous embedding space until the final time step, where it maps to discrete tokens using a shared-weight network. This formulation makes it straightforward to adapt established techniques from image-domain diffusion models, e.g., classifier-free guidance (CFG). Experiments show that ELF substantially outperforms leading discrete and continuous DLMs, achieving better generation quality with fewer sampling steps. These results suggest that ELF offers a promising path toward effective continuous DLMs. △ Less

Submitted 25 June, 2026; v1 submitted 11 May, 2026; originally announced May 2026.

Comments: Tech report. arXiv v2: add distillation results in Appendix B. https://linlu-qiu.github.io/assets/html/elf_pd.html

Reaching Beyond the Mode: RL for Distributional Reasoning in Language Models

Authors: Isha Puri, Mehul Damani, Idan Shenfeld, Marzyeh Ghassemi, Jacob Andreas, Yoon Kim

Abstract: Given a question, a language model (LM) implicitly encodes a distribution over possible answers. In practice, post-training procedures for LMs often collapse this distribution onto a single dominant mode. While this is generally not a problem for benchmark-style evaluations that assume one correct answer, many real-world tasks inherently involve multiple valid answers or irreducible uncertainty. E… ▽ More Given a question, a language model (LM) implicitly encodes a distribution over possible answers. In practice, post-training procedures for LMs often collapse this distribution onto a single dominant mode. While this is generally not a problem for benchmark-style evaluations that assume one correct answer, many real-world tasks inherently involve multiple valid answers or irreducible uncertainty. Examples include medical diagnosis, ambiguous question answering, and settings with incomplete information. In these cases, we would like LMs to generate multiple plausible hypotheses, ideally with confidence estimates for each one, and without computationally intensive repeated sampling to generate non-modal answers. This paper describes a multi-answer reinforcement learning approach for training LMs to perform distributional reasoning over multiple answers during inference. We modify the RL objective to enable models to explicitly generate multiple candidate answers in a single forward pass, internalizing aspects of inference-time search into the model's generative process. Across question-answering, medical diagnostic, and coding benchmarks, we observe improved diversity, coverage, and set-level calibration scores compared to single answer trained baselines. Models trained with our approach require fewer tokens to generate multiple answers than competing approaches. On coding tasks, they are also substantially more accurate. These results position multi-answer RL as a principled and compute-efficient alternative to inference-time scaling procedures such as best-of-k. Code and more information can be found at https://multi-answer-rl.github.io/. △ Less

Submitted 25 March, 2026; originally announced March 2026.

Pitfalls in Evaluating Interpretability Agents

Authors: Tal Haklay, Nikhil Prakash, Sana Pandey, Antonio Torralba, Aaron Mueller, Jacob Andreas, Tamar Rott Shaham, Yonatan Belinkov

Abstract: Automated interpretability systems aim to reduce the need for human labor and scale analysis to increasingly large models and diverse tasks. Recent efforts toward this goal leverage large language models (LLMs) at increasing levels of autonomy, ranging from fixed one-shot workflows to fully autonomous interpretability agents. This shift creates a corresponding need to scale evaluation approaches t… ▽ More Automated interpretability systems aim to reduce the need for human labor and scale analysis to increasingly large models and diverse tasks. Recent efforts toward this goal leverage large language models (LLMs) at increasing levels of autonomy, ranging from fixed one-shot workflows to fully autonomous interpretability agents. This shift creates a corresponding need to scale evaluation approaches to keep pace with both the volume and complexity of generated explanations. We investigate this challenge in the context of automated circuit analysis -- explaining the roles of model components when performing specific tasks. To this end, we build an agentic system in which a research agent iteratively designs experiments and refines hypotheses. When evaluated against human expert explanations across six circuit analysis tasks in the literature, the system appears competitive. However, closer examination reveals several pitfalls of replication-based evaluation: human expert explanations can be subjective or incomplete, outcome-based comparisons obscure the research process, and LLM-based systems may reproduce published findings via memorization or informed guessing. To address some of these pitfalls, we propose an unsupervised intrinsic evaluation based on the functional interchangeability of model components. Our work demonstrates fundamental challenges in evaluating complex automated interpretability systems and reveals key limitations of replication-based evaluation. △ Less

Submitted 20 March, 2026; originally announced March 2026.

MSC Class: 68T50 ACM Class: I.2.7

Do LLMs Benefit From Their Own Words?

Authors: Jenny Y. Huang, Leshem Choshen, Wei Sun, Omar Khattab, Ramón Fernandez Astudillo, Mehul Damani, Tamara Broderick, Jacob Andreas

Abstract: In multi-turn conversations, large language models typically condition on the full conversation history: both past user prompts and assistant responses. We revisit this design choice by comparing full-context prompting to four alternative, substantially-reduced context configurations. Analyzing in-the-wild multi-turn conversations across three open reasoning and one state-of-the-art model, we find… ▽ More In multi-turn conversations, large language models typically condition on the full conversation history: both past user prompts and assistant responses. We revisit this design choice by comparing full-context prompting to four alternative, substantially-reduced context configurations. Analyzing in-the-wild multi-turn conversations across three open reasoning and one state-of-the-art model, we find that response quality is largely preserved under aggressive context filtering: replacing all prior assistant turns with one-sentence summaries or keeping only the most recent user--assistant exchange often matches storing full context in performance while using roughly 8x less context. To understand this result, we observe that a substantial fraction of user turns (36.4%) in multi-turn conversations are self-contained and that many follow-up turns can be addressed by seeing only the immediately preceding user--assistant exchange. Furthermore, we find that when models condition on their own past responses, this can lead to context pollution, a phenomenon in which reasoning errors, hallucinations, or stylistic artifacts propagate across turns. Motivated by these findings, we design a context-filtering approach that selectively omits the assistant-side history. Taken together, these findings suggest moving away from storing full dialogue transcripts and instead retaining only what is relevant. △ Less

Submitted 11 August, 2026; v1 submitted 27 February, 2026; originally announced February 2026.

CONCUR: A Framework for Continual Constrained and Unconstrained Routing

Authors: Peter Baile Chen, Weiyue Li, Dan Roth, Michael Cafarella, Samuel Madden, Jacob Andreas

Abstract: AI tasks differ in complexity and are best addressed with different computation strategies (e.g., combinations of models and decoding methods). Hence, an effective routing system that maps tasks to the appropriate strategies is crucial. Most prior methods build the routing framework by training a single model across all strategies, which demands full retraining whenever new strategies appear and l… ▽ More AI tasks differ in complexity and are best addressed with different computation strategies (e.g., combinations of models and decoding methods). Hence, an effective routing system that maps tasks to the appropriate strategies is crucial. Most prior methods build the routing framework by training a single model across all strategies, which demands full retraining whenever new strategies appear and leads to high overhead. Attempts at such continual routing, however, often face difficulties with generalization. Prior models also typically use a single input representation, limiting their ability to capture the full complexity of the routing problem and leading to sub-optimal routing decisions. To address these gaps, we propose CONCUR, a continual routing framework that supports both constrained and unconstrained routing (i.e., routing with or without a budget). Our modular design trains a separate predictor model for each strategy, enabling seamless incorporation of new strategies with low additional training cost. Our predictors also leverage multiple representations of both tasks and computation strategies to better capture overall problem complexity. Experiments on both in-distribution and out-of-distribution, knowledge- and reasoning-intensive tasks show that our method outperforms the best single strategy and strong existing routing techniques with higher end-to-end accuracy and lower inference cost in both continual and non-continual settings, while also reducing training cost in the continual setting. △ Less

Submitted 10 December, 2025; originally announced December 2025.

ARC Is a Vision Problem!

Authors: Keya Hu, Ali Cy, Linlu Qiu, Xiaoman Delores Ding, Runqian Wang, Yeyin Eva Zhu, Jacob Andreas, Kaiming He

Abstract: The Abstraction and Reasoning Corpus (ARC) is designed to promote research on abstract reasoning, a fundamental aspect of human intelligence. Common approaches to ARC treat it as a language-oriented problem, addressed by large language models (LLMs) or recurrent reasoning models. However, although the puzzle-like tasks in ARC are inherently visual, existing research has rarely approached the probl… ▽ More The Abstraction and Reasoning Corpus (ARC) is designed to promote research on abstract reasoning, a fundamental aspect of human intelligence. Common approaches to ARC treat it as a language-oriented problem, addressed by large language models (LLMs) or recurrent reasoning models. However, although the puzzle-like tasks in ARC are inherently visual, existing research has rarely approached the problem from a vision-centric perspective. In this work, we formulate ARC within a vision paradigm, framing it as an image-to-image translation problem. To incorporate visual priors, we represent the inputs on a "canvas" that can be processed like natural images. It is then natural for us to apply standard vision architectures, such as a vanilla Vision Transformer (ViT), to perform image-to-image mapping. Our model is trained from scratch solely on ARC data and generalizes to unseen tasks through test-time training. Our framework, termed Vision ARC (VARC), achieves 60.4% accuracy on the ARC-1 benchmark, substantially outperforming existing methods that are also trained from scratch. Our results are competitive with those of leading LLMs and close the gap to average human performance. △ Less

Submitted 18 November, 2025; originally announced November 2025.

Comments: Technical Report. Project webpage: https://github.com/lillian039/VARC

Training Language Models to Explain Their Own Computations

Authors: Belinda Z. Li, Zifan Carl Guo, Vincent Huang, Jacob Steinhardt, Jacob Andreas

Abstract: Can language models (LMs) learn to faithfully describe their internal computations? Are they better able to describe themselves than other models? We study the extent to which LMs' privileged access to their own internals can be leveraged to produce new techniques for explaining their behavior. Using existing interpretability techniques as a source of ground truth, we fine-tune LMs to generate nat… ▽ More Can language models (LMs) learn to faithfully describe their internal computations? Are they better able to describe themselves than other models? We study the extent to which LMs' privileged access to their own internals can be leveraged to produce new techniques for explaining their behavior. Using existing interpretability techniques as a source of ground truth, we fine-tune LMs to generate natural language descriptions of (1) the information encoded by LM features, (2) the causal structure of LMs' internal activations, and (3) the influence of specific input tokens on LM outputs. When trained with only tens of thousands of example explanations, explainer models exhibit non-trivial generalization to new queries. This generalization appears partly attributable to explainer models' privileged access to their own internals: using a model to explain its own computations generally works better than using a *different* model to explain its computations (even if the explainer model is significantly more capable than the target). Our results suggest not only that LMs can learn to reliably explain their internal computations, but that such explanations offer a scalable complement to existing interpretability methods. Code and data at https://github.com/TransluceAI/introspective-interp △ Less

Submitted 9 February, 2026; v1 submitted 11 November, 2025; originally announced November 2025.

Comments: 23 pages, 8 tables, 7 figures. Code and data at https://github.com/TransluceAI/introspective-interp

Automated Detection of Visual Attribute Reliance with a Self-Reflective Agent

Authors: Christy Li, Josep Lopez Camuñas, Jake Thomas Touchet, Jacob Andreas, Agata Lapedriza, Antonio Torralba, Tamar Rott Shaham

Abstract: When a vision model performs image recognition, which visual attributes drive its predictions? Detecting unintended reliance on specific visual features is critical for ensuring model robustness, preventing overfitting, and avoiding spurious correlations. We introduce an automated framework for detecting such dependencies in trained vision models. At the core of our method is a self-reflective age… ▽ More When a vision model performs image recognition, which visual attributes drive its predictions? Detecting unintended reliance on specific visual features is critical for ensuring model robustness, preventing overfitting, and avoiding spurious correlations. We introduce an automated framework for detecting such dependencies in trained vision models. At the core of our method is a self-reflective agent that systematically generates and tests hypotheses about visual attributes that a model may rely on. This process is iterative: the agent refines its hypotheses based on experimental outcomes and uses a self-evaluation protocol to assess whether its findings accurately explain model behavior. When inconsistencies arise, the agent self-reflects over its findings and triggers a new cycle of experimentation. We evaluate our approach on a novel benchmark of 130 models designed to exhibit diverse visual attribute dependencies across 18 categories. Our results show that the agent's performance consistently improves with self-reflection, with a significant performance increase over non-reflective baselines. We further demonstrate that the agent identifies real-world visual attribute dependencies in state-of-the-art models, including CLIP's vision encoder and the YOLOv8 object detector. △ Less

Submitted 18 November, 2025; v1 submitted 24 October, 2025; originally announced October 2025.

Comments: 32 pages, 10 figures, Neurips 2025

Code-enabled language models can outperform reasoning models on diverse tasks

Authors: Cedegao E. Zhang, Cédric Colas, Gabriel Poesia, Joshua B. Tenenbaum, Jacob Andreas

Abstract: Reasoning models (RMs), language models (LMs) trained with reinforcement learning to produce long-form natural language reasoning, have been remarkably successful, but they still require large amounts of computation and data to train, and can be slow and expensive to run. In this paper, we show that standard instruct LMs can already be elicited to be strong reasoners at a level comparable to or ev… ▽ More Reasoning models (RMs), language models (LMs) trained with reinforcement learning to produce long-form natural language reasoning, have been remarkably successful, but they still require large amounts of computation and data to train, and can be slow and expensive to run. In this paper, we show that standard instruct LMs can already be elicited to be strong reasoners at a level comparable to or even surpassing their corresponding RMs (e.g., DeepSeek V3 vs R1) without finetuning, across diverse domains from instruction following and creative generation to mathematical reasoning. This is achieved by CodeAdapt, our simple recipe that combines the CodeAct framework, where LMs interleave natural language reasoning with code execution in a multi-step fashion, with few-shot bootstrap in-context learning from as few as five training problems. Analyzing four matched pairs of LMs and RMs, we find that CodeAdapt enables three LMs to outperform the corresponding RMs on average over eight tasks (up to 22.9%) while being 10-81% more token efficient, and delivers superior performance on six tasks when averaged over the four models (up to 35.7%). Furthermore, the code-augmented reasoning traces display rich and varied problem-solving strategies. Our findings support that (1) CodeAdapt-style learning and reasoning may be robust and domain general and (2) code-enabled LMs are cognitively grounded and powerful systems, potentially providing a strong foundation for in-weight reinforcement learning. △ Less

Submitted 23 October, 2025; originally announced October 2025.

Shoot First, Ask Questions Later? Building Rational Agents that Explore and Act Like People

Authors: Gabriel Grand, Valerio Pepe, Jacob Andreas, Joshua B. Tenenbaum

Abstract: Many emerging applications of AI--from scientific discovery to medical diagnosis--require agents to seek information strategically: forming hypotheses, asking targeted questions, and making decisions under uncertainty. In high-stakes settings with limited resources, do language models (LMs) behave like rational agents? Drawing on insights from human cognition, we develop methods to evaluate and en… ▽ More Many emerging applications of AI--from scientific discovery to medical diagnosis--require agents to seek information strategically: forming hypotheses, asking targeted questions, and making decisions under uncertainty. In high-stakes settings with limited resources, do language models (LMs) behave like rational agents? Drawing on insights from human cognition, we develop methods to evaluate and enhance agentic information-seeking. First, we introduce a decision-oriented dialogue task called Collaborative Battleship, in which a Captain must balance exploration (asking questions) and action (taking shots), while a Spotter must supply accurate, contextually-grounded answers. Compared to human players (N=42), we find that many LM agents struggle to ask informative questions, produce accurate answers, and identify high-utility actions. To address these gaps, we develop novel Monte Carlo inference strategies for LMs inspired by Bayesian Experimental Design (BED). For Spotter agents, our approach boosts accuracy by up to 14.7% absolute over LM-only baselines; for Captain agents, it raises expected information gain (EIG) by up to 0.227 bits (94.2% of the achievable noise ceiling). Combined, these components yield sharper targeting (+0.303-0.374 F1), and enable weaker LMs, such as Llama-4-Scout, to outperform both humans (8% -> 82% win rate) and frontier models (0% -> 67% win rate vs. GPT-5) at ~1% of GPT-5's cost. We replicate these findings on Guess Who?, where our methods significantly boost accuracy (+28.3-42.4 p.p.), demonstrating their general applicability for building information-seeking agents. △ Less

Submitted 6 March, 2026; v1 submitted 23 October, 2025; originally announced October 2025.

Comments: ICLR 2026

OpenEstimate: Evaluating LLMs on Reasoning Under Uncertainty with Real-World Data

Authors: Alana Renda, Jillian Ross, Michael Cafarella, Jacob Andreas

Abstract: Real-world settings where language models (LMs) are deployed -- in domains spanning healthcare, finance, and other forms of knowledge work -- require models to grapple with incomplete information and reason under uncertainty. Yet most LM evaluations focus on problems with well-defined answers and success criteria. This gap exists in part because natural problems involving uncertainty are difficult… ▽ More Real-world settings where language models (LMs) are deployed -- in domains spanning healthcare, finance, and other forms of knowledge work -- require models to grapple with incomplete information and reason under uncertainty. Yet most LM evaluations focus on problems with well-defined answers and success criteria. This gap exists in part because natural problems involving uncertainty are difficult to construct: given that LMs have access to most of the same knowledge as humans, it is non-trivial to design questions for which LMs will struggle to produce correct answers, but which humans can answer reliably. As a result, LM performance on reasoning under uncertainty remains poorly characterized. To address this gap, we introduce OpenEstimate, an extensible, multi-domain benchmark for evaluating LMs on numerical estimation tasks that require models to synthesize significant amounts of background information and express predictions as probabilistic priors. We assess these priors for accuracy and calibration, quantifying their usefulness relative to samples from the true distribution of interest. Across six frontier LMs, we find that LM-elicited priors are often inaccurate and overconfident. Performance improves modestly depending on how uncertainty is elicited from the model, but is largely unaffected by changes in sampling strategy, reasoning effort, or prompt design. The OpenEstimate benchmark thus offers a challenging evaluation for frontier LMs and a platform for developing models that are better at probabilistic estimation and reasoning under uncertainty. △ Less

Submitted 22 April, 2026; v1 submitted 16 October, 2025; originally announced October 2025.

Learning to Make MISTAKEs: Modeling Incorrect Student Thinking And Key Errors

Authors: Alexis Ross, Jacob Andreas

Abstract: Research on reasoning in language models (LMs) predominantly focuses on improving the correctness of their outputs. But some important applications require modeling reasoning patterns that are incorrect. For example, automated systems that can reason about and simulate student errors are useful for providing real-time feedback in the classroom or offline practice for educators-in-training. This pa… ▽ More Research on reasoning in language models (LMs) predominantly focuses on improving the correctness of their outputs. But some important applications require modeling reasoning patterns that are incorrect. For example, automated systems that can reason about and simulate student errors are useful for providing real-time feedback in the classroom or offline practice for educators-in-training. This paper presents a new method, MISTAKE, that (1) constructs high-quality synthetic examples of reasoning errors by leveraging cycle consistency between incorrect answers and latent misconceptions; and (2) uses the generated data to learn models for student simulation, misconception classification, and answer generation. We evaluate MISTAKE on three educational tasks and find that it results in (1) higher accuracy when simulating incorrect student answers based on specific misconceptions, (2) increased performance inferring latent misconceptions from observed incorrect answers, and (3) higher alignment with expert-written distractor answers when generating incorrect answers (e.g., for multiple-choice tests). △ Less

Submitted 13 October, 2025; originally announced October 2025.

Sample-Efficient Online Learning in LM Agents via Hindsight Trajectory Rewriting

Authors: Michael Y. Hu, Benjamin Van Durme, Jacob Andreas, Harsh Jhamtani

Abstract: Language model (LM) agents deployed in novel environments often exhibit poor sample efficiency when learning from sequential interactions. This significantly hinders the usefulness of such agents in environments where interaction is costly (for example, when they interact with humans or reset physical systems). While a number of existing LM agent architectures incorporate various mechanisms for ex… ▽ More Language model (LM) agents deployed in novel environments often exhibit poor sample efficiency when learning from sequential interactions. This significantly hinders the usefulness of such agents in environments where interaction is costly (for example, when they interact with humans or reset physical systems). While a number of existing LM agent architectures incorporate various mechanisms for experience storage and reflection, they make limited use of LMs' abilities to directly generate or reason about full counterfactual trajectories. We introduce ECHO (Experience Consolidation via Hindsight Optimization), a prompting framework that adapts hindsight experience replay from reinforcement learning for language model agents. ECHO generates optimized trajectories for alternative goals that could have been achieved during failed attempts, effectively creating synthetic positive examples from unsuccessful interactions. Our approach consists of two components: a hindsight rule that uses the language model itself to identify relevant subgoals and generate optimized trajectories, and an update rule that maintains compressed trajectory representations in memory. We evaluate ECHO on stateful versions of XMiniGrid, a text-based navigation and planning benchmark, and PeopleJoinQA, a collaborative information-gathering enterprise simulation. Across both domains, ECHO outperforms vanilla language agent baselines by up to 80%; in XMiniGrid, it also outperforms a number of sophisticated agent architectures including Reflexion and AWM, demonstrating faster adaptation to novel environments through more effective utilization of past experiences. △ Less

Submitted 2 January, 2026; v1 submitted 11 October, 2025; originally announced October 2025.

Modeling Student Learning with 3.8 Million Program Traces

Authors: Alexis Ross, Megha Srivastava, Jeremiah Blanchard, Jacob Andreas

Abstract: As programmers write code, they often edit and retry multiple times, creating rich "interaction traces" that reveal how they approach coding tasks and provide clues about their level of skill development. For novice programmers in particular, these traces reflect the diverse reasoning processes they employ to code, such as exploratory behavior to understand how a programming concept works, re-stra… ▽ More As programmers write code, they often edit and retry multiple times, creating rich "interaction traces" that reveal how they approach coding tasks and provide clues about their level of skill development. For novice programmers in particular, these traces reflect the diverse reasoning processes they employ to code, such as exploratory behavior to understand how a programming concept works, re-strategizing in response to bugs, and personalizing stylistic choices. In this work, we explore what can be learned from training language models on such reasoning traces: not just about code, but about coders, and particularly students learning to program. We introduce a dataset of over 3.8 million programming reasoning traces from users of Pencil Code, a free online educational platform used by students to learn simple programming concepts. Compared to models trained only on final programs or synthetically-generated traces, we find that models trained on real traces are stronger at modeling diverse student behavior. Through both behavioral and probing analyses, we also find that many properties of code traces, such as goal backtracking or number of comments, can be predicted from learned representations of the students who write them. Building on this result, we show that we can help students recover from mistakes by steering code generation models to identify a sequence of edits that will results in more correct code while remaining close to the original student's style. Together, our results suggest that many properties of code are properties of individual students and that training on edit traces can lead to models that are more steerable, more predictive of student behavior while programming, and better at generating programs in their final states. Code and data is available at https://github.com/meghabyte/pencilcode-public △ Less

Submitted 15 April, 2026; v1 submitted 6 October, 2025; originally announced October 2025.

Comments: Accepted to 27th International Conference on AI in Education (AIED 2026)

Language and Experience: A Computational Model of Social Learning in Complex Tasks

Authors: Cédric Colas, Tracey Mills, Ben Prystawski, Michael Henry Tessler, Noah Goodman, Jacob Andreas, Joshua Tenenbaum

Abstract: The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models social learning as joint probabilistic inference over structured, executable world models given sensorim… ▽ More The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models social learning as joint probabilistic inference over structured, executable world models given sensorimotor and linguistic data. We make this possible by turning a pretrained language model into a probabilistic model of how humans share advice conditioned on their beliefs, allowing our agents both to generate advice for others and to interpret linguistic input as evidence during Bayesian inference. Using behavioral experiments and simulations across 10 video games, we show how linguistic guidance can shape exploration and accelerate learning by reducing risky interactions and speeding up key discoveries in both humans and models. We further explore how knowledge can accumulate across generations through iterated learning experiments and demonstrate successful knowledge transfer between humans and models -- revealing how structured, language-compatible representations might enable human-machine collaborative learning. △ Less

Submitted 18 February, 2026; v1 submitted 26 August, 2025; originally announced September 2025.

Comments: Code: github.com/ccolas/language_and_experience Demo: cedriccolas.com/demos/language_and_experience

Journal ref: ICLR 2026; CogSci 2025

Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty

Authors: Mehul Damani, Isha Puri, Stewart Slocum, Idan Shenfeld, Leshem Choshen, Yoon Kim, Jacob Andreas

Abstract: When language models (LMs) are trained via reinforcement learning (RL) to generate natural language "reasoning chains", their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-conf… ▽ More When language models (LMs) are trained via reinforcement learning (RL) to generate natural language "reasoning chains", their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-confidence outputs, they often have the unintended side-effect of degrading calibration and increasing the rate at which LMs generate incorrect responses (or "hallucinate") in other problem domains. This paper describes RLCR (Reinforcement Learning with Calibration Rewards), an approach to training reasoning models that jointly improves accuracy and calibrated confidence estimation. During RLCR, LMs generate both predictions and numerical confidence estimates after reasoning. They are trained to optimize a reward function that augments a binary correctness score with a Brier score -- a scoring rule for confidence estimates that incentivizes calibrated prediction. We first prove that this reward function (or any reward function that uses a bounded, proper scoring rule) yields models whose predictions are both accurate and well-calibrated. We next show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy, on both in-domain and out-of-domain evaluations -- outperforming both ordinary RL training and classifiers trained to assign post-hoc confidence scores. While ordinary RL hurts calibration, RLCR improves it. Finally, we demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration via confidence-weighted scaling methods. Our results show that explicitly optimizing for calibration can produce more generally reliable reasoning models. Code, models, and further info is available at https://rl-calibration.github.io/. △ Less

Submitted 14 May, 2026; v1 submitted 22 July, 2025; originally announced July 2025.

Assessing Adaptive World Models in Machines with Novel Games

Authors: Lance Ying, Katherine M. Collins, Prafull Sharma, Cedric Colas, Kaiya Ivy Zhao, Adrian Weller, Zenna Tavares, Phillip Isola, Samuel J. Gershman, Jacob D. Andreas, Thomas L. Griffiths, Francois Chollet, Kelsey R. Allen, Joshua B. Tenenbaum

Abstract: Human intelligence exhibits a remarkable capacity for rapid adaptation and effective problem-solving in novel and unfamiliar contexts. We argue that this profound adaptability is fundamentally linked to the efficient construction and refinement of internal representations of the environment, commonly referred to as world models, and we refer to this adaptation mechanism as world model induction. H… ▽ More Human intelligence exhibits a remarkable capacity for rapid adaptation and effective problem-solving in novel and unfamiliar contexts. We argue that this profound adaptability is fundamentally linked to the efficient construction and refinement of internal representations of the environment, commonly referred to as world models, and we refer to this adaptation mechanism as world model induction. However, current understanding and evaluation of world models in artificial intelligence (AI) remains narrow, often focusing on static representations learned from training on massive corpora of data, instead of the efficiency and efficacy in learning these representations through interaction and exploration within a novel environment. In this Perspective, we provide a view of world model induction drawing on decades of research in cognitive science on how humans learn and adapt so efficiently; we then call for a new evaluation framework for assessing adaptive world models in AI. Concretely, we propose a new benchmarking paradigm based on suites of carefully designed games with genuine, deep and continually refreshing novelty in the underlying game structures -- we refer to this class of games as novel games. We detail key desiderata for constructing these games and propose appropriate metrics to explicitly challenge and evaluate the agent's ability for rapid world model induction. We hope that this new evaluation framework will inspire future evaluation efforts on world models in AI and provide a crucial step towards developing AI systems capable of human-like rapid adaptation and robust generalization -- a critical component of artificial general intelligence. △ Less

Submitted 22 July, 2025; v1 submitted 17 July, 2025; originally announced July 2025.

Comments: 17 pages, 4 figures

Modeling Open-World Cognition as On-Demand Synthesis of Probabilistic Models

Authors: Lionel Wong, Katherine M. Collins, Lance Ying, Cedegao E. Zhang, Adrian Weller, Tobias Gerstenberg, Timothy O'Donnell, Alexander K. Lew, Jacob D. Andreas, Joshua B. Tenenbaum, Tyler Brooke-Wilson

Abstract: When faced with novel situations, people are able to marshal relevant considerations from a wide range of background knowledge and put these to use in inferences and predictions. What permits us to draw in globally relevant information and reason over it coherently? Here, we explore the hypothesis that people use a combination of distributed and symbolic representations to construct bespoke mental… ▽ More When faced with novel situations, people are able to marshal relevant considerations from a wide range of background knowledge and put these to use in inferences and predictions. What permits us to draw in globally relevant information and reason over it coherently? Here, we explore the hypothesis that people use a combination of distributed and symbolic representations to construct bespoke mental models tailored to novel situations. We propose a computational implementation of this idea -- a ``Model Synthesis Architecture'' (MSA) -- using language models to implement global relevance-based retrieval and model synthesis and probabilistic programs to implement bespoke, coherent world models. We evaluate our MSA as a model of human judgments on a novel reasoning dataset. The dataset -- built around a `Model Olympics` domain of sports vignettes -- tests models' capacity for human-like, open-ended reasoning by requiring (i) judgments about novel causal structures described in language; (ii) drawing on large bodies of background knowledge; and (iii) doing both in light of observations that introduce arbitrary novel variables. Our MSA approach captures human judgments better than language model-only baselines, under both direct and chain-of-thought generations from the LM that supports model synthesis. These results suggest that MSAs can be implemented in a way that mirrors people's ability to deliver locally coherent reasoning over globally relevant variables, offering a path to understanding and replicating human reasoning in open-ended domains. △ Less

Submitted 18 July, 2025; v1 submitted 16 July, 2025; originally announced July 2025.

Comments: Presented at CogSci 2025

LLM Hypnosis: Exploiting User Feedback for Unauthorized Knowledge Injection to All Users

Authors: Almog Hilel, Riddhi Bhagwat, Idan Shenfeld, Jacob Andreas, Leshem Choshen

Abstract: We describe a vulnerability in language models (LMs) trained with user feedback, whereby a single user can persistently alter LM knowledge and behavior given only the ability to provide prompts and upvote / downvote feedback on LM outputs. To implement the attack, the attacker prompts the LM to stochastically output either a "poisoned" or benign response, then upvotes the poisoned response or down… ▽ More We describe a vulnerability in language models (LMs) trained with user feedback, whereby a single user can persistently alter LM knowledge and behavior given only the ability to provide prompts and upvote / downvote feedback on LM outputs. To implement the attack, the attacker prompts the LM to stochastically output either a "poisoned" or benign response, then upvotes the poisoned response or downvotes the benign one. When feedback signals are used in a subsequent preference tuning behavior, LMs exhibit increased probability of producing poisoned responses even in contexts without malicious prompts. We show that this attack can be used to (1) insert factual knowledge the model did not previously possess, (2) modify code generation patterns in ways that introduce exploitable security flaws, and (3) inject fake financial news. Our finding both identifies a new qualitative feature of language model preference tuning (showing that it even highly restricted forms of preference data can be used to exert fine-grained control over behavior), and a new attack mechanism for LMs trained with user feedback (extending work on pretraining-time data poisoning and deployment-time prompt injection). △ Less

Submitted 20 April, 2026; v1 submitted 3 July, 2025; originally announced July 2025.

Can Gradient Descent Simulate Prompting?

Authors: Eric Zhang, Leshem Choshen, Jacob Andreas

Abstract: There are two primary ways of incorporating new information into a language model (LM): changing its prompt or changing its parameters, e.g. via fine-tuning. Parameter updates incur no long-term storage cost for model changes. However, for many model updates, prompting is significantly more effective: prompted models can generalize robustly from single examples and draw logical inferences that do… ▽ More There are two primary ways of incorporating new information into a language model (LM): changing its prompt or changing its parameters, e.g. via fine-tuning. Parameter updates incur no long-term storage cost for model changes. However, for many model updates, prompting is significantly more effective: prompted models can generalize robustly from single examples and draw logical inferences that do not occur under standard fine-tuning. Can models be modified so that fine-tuning does emulate prompting? This paper describes a method for meta-training LMs such that gradient updates emulate the effects of conditioning on new information. Our approach uses tools from gradient-based meta-learning but uses an LM's own prompted predictions as targets, eliminating the need for ground-truth labels. Subsequent gradient descent training recovers some (and occasionally all) of prompted model performance -- showing improvement on the ``reversal curse'' tasks, and answering questions about text passages after a single gradient update. These results suggest that, with appropriate initialization, gradient descent can be surprisingly expressive. Our results suggest new avenues for long-context modeling and offer insight into the generalization capabilities of gradient-based learning. △ Less

Submitted 26 June, 2025; originally announced June 2025.

Comments: 14 pages, 2 figures

Sequential-Parallel Duality in Prefix Scannable Models

Authors: Morris Yau, Sharut Gupta, Valerie Engelmayer, Kazuki Irie, Stefanie Jegelka, Jacob Andreas

Abstract: Modern neural sequence models are designed to meet the dual mandate of parallelizable training and fast sequential inference. Recent developments have given rise to various models, such as Gated Linear Attention (GLA) and Mamba, that achieve such ``sequential-parallel duality.'' This raises a natural question: can we characterize the full class of neural sequence models that support near-constant-… ▽ More Modern neural sequence models are designed to meet the dual mandate of parallelizable training and fast sequential inference. Recent developments have given rise to various models, such as Gated Linear Attention (GLA) and Mamba, that achieve such ``sequential-parallel duality.'' This raises a natural question: can we characterize the full class of neural sequence models that support near-constant-time parallel evaluation and linear-time, constant-space sequential inference? We begin by describing a broad class of such models -- state space models -- as those whose state updates can be computed using the classic parallel prefix scan algorithm with a custom associative aggregation operator. We then define a more general class, Prefix-Scannable Models (PSMs), by relaxing the state aggregation operator to allow arbitrary (potentially non-associative) functions such as softmax attention. This generalization unifies many existing architectures, including element-wise RNNs (e.g., Mamba) and linear transformers (e.g., GLA, Mamba2, mLSTM), while also introducing new models with softmax-like operators that achieve O(1) amortized compute per token and log(N) memory for sequence length N. We empirically evaluate such models on illustrative small-scale language modeling and canonical synthetic tasks, including state tracking and associative recall. Empirically, we find that PSMs retain the expressivity of transformer-based architectures while matching the inference efficiency of state space models -- in some cases exhibiting better length generalization than either. △ Less

Submitted 11 March, 2026; v1 submitted 12 June, 2025; originally announced June 2025.

Line of Sight: On Linear Representations in VLLMs

Authors: Achyuta Rajaram, Sarah Schwettmann, Jacob Andreas, Arthur Conmy

Abstract: Language models can be equipped with multimodal capabilities by fine-tuning on embeddings of visual inputs. But how do such multimodal models represent images in their hidden activations? We explore representations of image concepts within LlaVA-Next, a popular open-source VLLM. We find a diverse set of ImageNet classes represented via linearly decodable features in the residual stream. We show th… ▽ More Language models can be equipped with multimodal capabilities by fine-tuning on embeddings of visual inputs. But how do such multimodal models represent images in their hidden activations? We explore representations of image concepts within LlaVA-Next, a popular open-source VLLM. We find a diverse set of ImageNet classes represented via linearly decodable features in the residual stream. We show that the features are causal by performing targeted edits on the model output. In order to increase the diversity of the studied linear features, we train multimodal Sparse Autoencoders (SAEs), creating a highly interpretable dictionary of text and image features. We find that although model representations across modalities are quite disjoint, they become increasingly shared in deeper layers. △ Less

Submitted 5 June, 2025; originally announced June 2025.

Comments: 8 pages, 9 figures

Breakpoint: Scalable evaluation of system-level reasoning in LLM code agents

Authors: Kaivalya Hariharan, Uzay Girit, Atticus Wang, Jacob Andreas

Abstract: Benchmarks for large language models (LLMs) have predominantly assessed short-horizon, localized reasoning. Existing long-horizon suites (e.g. SWE-bench) rely on manually curated issues, so expanding or tuning difficulty demands expensive human effort and evaluations quickly saturate. However, many real-world tasks, such as software engineering or scientific research, require agents to rapidly com… ▽ More Benchmarks for large language models (LLMs) have predominantly assessed short-horizon, localized reasoning. Existing long-horizon suites (e.g. SWE-bench) rely on manually curated issues, so expanding or tuning difficulty demands expensive human effort and evaluations quickly saturate. However, many real-world tasks, such as software engineering or scientific research, require agents to rapidly comprehend and manipulate novel, complex structures dynamically; evaluating these capabilities requires the ability to construct large and varied sets of problems for agents to solve. We introduce Breakpoint, a benchmarking methodology that automatically generates code-repair tasks by adversarially corrupting functions within real-world software repositories. Breakpoint systematically controls task difficulty along two clear dimensions: local reasoning (characterized by code complexity metrics such as cyclomatic complexity) and system-level reasoning (characterized by call-graph centrality and the number of simultaneously corrupted interdependent functions). In experiments across more than 900 generated tasks we demonstrate that our methodology can scale to arbitrary difficulty, with state-of-the-art models' success rates ranging from 55% on the easiest tasks down to 0% on the hardest. △ Less

Submitted 30 May, 2025; originally announced June 2025.

Comments: 21 pages, 14 figures

Log-Augmented Generation: Scaling Test-Time Reasoning with Reusable Computation

Authors: Peter Baile Chen, Yi Zhang, Dan Roth, Samuel Madden, Jacob Andreas, Michael Cafarella

Abstract: While humans naturally learn and adapt from past experiences, large language models (LLMs) and their agentic counterparts struggle to retain reasoning from previous tasks and apply them in future contexts. To address this limitation, we propose a novel framework, log-augmented generation (LAG) that directly reuses prior computation and reasoning from past logs at test time to enhance model's abili… ▽ More While humans naturally learn and adapt from past experiences, large language models (LLMs) and their agentic counterparts struggle to retain reasoning from previous tasks and apply them in future contexts. To address this limitation, we propose a novel framework, log-augmented generation (LAG) that directly reuses prior computation and reasoning from past logs at test time to enhance model's ability to learn from previous tasks and perform better on new, unseen challenges, all while keeping the system efficient and scalable. Specifically, our system represents task logs using key-value (KV) caches, encoding the full reasoning context of prior tasks while storing KV caches for only a selected subset of tokens. When a new task arises, LAG retrieves the KV values from relevant logs to augment generation. Our approach differs from reflection-based memory mechanisms by directly reusing prior reasoning and computations without requiring additional steps for knowledge extraction or distillation. Our method also goes beyond existing KV caching techniques, which primarily target efficiency gains rather than improving accuracy. Experiments on knowledge- and reasoning-intensive datasets demonstrate that our method significantly outperforms standard agentic systems that do not utilize logs, as well as existing solutions based on reflection and KV cache techniques. △ Less

Submitted 20 May, 2025; originally announced May 2025.

Comments: Data and code are available at https://peterbaile.github.io/lag/

Self-Steering Language Models

Authors: Gabriel Grand, Joshua B. Tenenbaum, Vikash K. Mansinghka, Alexander K. Lew, Jacob Andreas

Abstract: While test-time reasoning enables language models (LMs) to tackle complex tasks, searching or planning in natural language can be slow, costly, and error-prone. But even when LMs struggle to emulate the precise reasoning steps needed to solve a problem, they often excel at describing its abstract structure--both how to verify solutions and how to search for them. This paper introduces DisCIPL, a m… ▽ More While test-time reasoning enables language models (LMs) to tackle complex tasks, searching or planning in natural language can be slow, costly, and error-prone. But even when LMs struggle to emulate the precise reasoning steps needed to solve a problem, they often excel at describing its abstract structure--both how to verify solutions and how to search for them. This paper introduces DisCIPL, a method for "self-steering" LMs where a Planner model generates a task-specific inference program that is executed by a population of Follower models. Our approach equips LMs with the ability to write recursive search procedures that guide LM inference, enabling new forms of verifiable and efficient reasoning. When instantiated with a small Follower (e.g., Llama-3.2-1B or Qwen3-1.7B), DisCIPL matches (and sometimes outperforms) much larger models, including GPT-4o and o1, on challenging constrained generation tasks. Our work opens up a design space of highly-parallelized Monte Carlo inference strategies that outperform standard best-of-N sampling, require no finetuning, and can be implemented automatically by existing LMs. △ Less

Submitted 8 August, 2025; v1 submitted 9 April, 2025; originally announced April 2025.

Comments: Accepted to COLM 2025

ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning

Authors: Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, Shiyu Chang

Abstract: We present ThinkPrune, a simple yet effective method for pruning the thinking length for long-thinking LLMs, which has been found to often produce inefficient and redundant thinking processes. Existing preliminary explorations of reducing thinking length primarily focus on forcing the thinking process to early exit, rather than adapting the LLM to optimize and consolidate the thinking process, and… ▽ More We present ThinkPrune, a simple yet effective method for pruning the thinking length for long-thinking LLMs, which has been found to often produce inefficient and redundant thinking processes. Existing preliminary explorations of reducing thinking length primarily focus on forcing the thinking process to early exit, rather than adapting the LLM to optimize and consolidate the thinking process, and therefore the length-performance tradeoff observed so far is sub-optimal. To fill this gap, ThinkPrune offers a simple solution that continuously trains the long-thinking LLMs via reinforcement learning (RL) with an added token limit, beyond which any unfinished thoughts and answers will be discarded, resulting in a zero reward. To further preserve model performance, we introduce an iterative length pruning approach, where multiple rounds of RL are conducted, each with an increasingly more stringent token limit. We observed that ThinkPrune results in a remarkable performance-length tradeoff -- on the AIME24 dataset, the reasoning length of DeepSeek-R1-Distill-Qwen-1.5B can be reduced by half with only 2% drop in performance. We also observed that after pruning, the LLMs can bypass unnecessary steps while keeping the core reasoning process complete. Code is available at https://github.com/UCSB-NLP-Chang/ThinkPrune. △ Less

Submitted 1 April, 2025; originally announced April 2025.

Comments: 15 pages, 7 figures

(How) Do Language Models Track State?

Authors: Belinda Z. Li, Zifan Carl Guo, Jacob Andreas

Abstract: Transformer language models (LMs) exhibit behaviors -- from storytelling to code generation -- that seem to require tracking the unobserved state of an evolving world. How do they do this? We study state tracking in LMs trained or fine-tuned to compose permutations (i.e., to compute the order of a set of objects after a sequence of swaps). Despite the simple algebraic structure of this problem, ma… ▽ More Transformer language models (LMs) exhibit behaviors -- from storytelling to code generation -- that seem to require tracking the unobserved state of an evolving world. How do they do this? We study state tracking in LMs trained or fine-tuned to compose permutations (i.e., to compute the order of a set of objects after a sequence of swaps). Despite the simple algebraic structure of this problem, many other tasks (e.g., simulation of finite automata and evaluation of boolean expressions) can be reduced to permutation composition, making it a natural model for state tracking in general. We show that LMs consistently learn one of two state tracking mechanisms for this task. The first closely resembles the "associative scan" construction used in recent theoretical work by Liu et al. (2023) and Merrill et al. (2024). The second uses an easy-to-compute feature (permutation parity) to partially prune the space of outputs, and then refines this with an associative scan. LMs that learn the former algorithm tend to generalize better and converge faster, and we show how to steer LMs toward one or the other with intermediate training tasks that encourage or suppress the heuristics. Our results demonstrate that transformer LMs, whether pre-trained or fine-tuned, can learn to implement efficient and interpretable state-tracking mechanisms, and the emergence of these mechanisms can be predicted and controlled. △ Less

Submitted 30 October, 2025; v1 submitted 4 March, 2025; originally announced March 2025.

Comments: 21 pages, 17 figures, 1 table. Code: http://github.com/belindal/state-tracking

LM Agents for Coordinating Multi-User Information Gathering

Authors: Harsh Jhamtani, Jacob Andreas, Benjamin Van Durme

Abstract: This paper introduces PeopleJoin, a benchmark for evaluating LM-mediated collaborative problem solving. Given a user request, PeopleJoin agents must identify teammates who might be able to assist, converse with these teammates to gather information, and finally compile a useful answer or summary for the original user. PeopleJoin comprises two evaluation domains: PeopleJoin-QA, focused on questions… ▽ More This paper introduces PeopleJoin, a benchmark for evaluating LM-mediated collaborative problem solving. Given a user request, PeopleJoin agents must identify teammates who might be able to assist, converse with these teammates to gather information, and finally compile a useful answer or summary for the original user. PeopleJoin comprises two evaluation domains: PeopleJoin-QA, focused on questions about tabular data, and PeopleJoin-DocCreation, focused on document creation tasks. The two domains are adapted from existing NLP benchmarks for database question answering and multi-document summarization; here, however, the information needed to complete these tasks is distributed across synthetic ``organizations'' of 2--20 users, simulating natural multi-user collaboration scenarios. We implemented several popular LM agent architectures, evaluating their accuracy and efficiency at completing tasks, and highlight new research questions that can be studied using PeopleJoin. △ Less

Submitted 17 February, 2025; originally announced February 2025.

The Surprising Effectiveness of Test-Time Training for Few-Shot Learning

Authors: Ekin Akyürek, Mehul Damani, Adam Zweiger, Linlu Qiu, Han Guo, Jyothish Pari, Yoon Kim, Jacob Andreas

Abstract: Language models (LMs) have shown impressive performance on tasks within their training distribution, but often struggle with structurally novel tasks even when given a small number of in-context task examples. We investigate the effectiveness of test-time training (TTT) -- temporarily updating model parameters during inference using a loss derived from input data -- as a mechanism for improving LM… ▽ More Language models (LMs) have shown impressive performance on tasks within their training distribution, but often struggle with structurally novel tasks even when given a small number of in-context task examples. We investigate the effectiveness of test-time training (TTT) -- temporarily updating model parameters during inference using a loss derived from input data -- as a mechanism for improving LMs' reasoning and few-shot learning capabilities. On the Abstraction and Reasoning Corpus (ARC), performing TTT with in-context examples yields up to $6\times$ higher accuracy compared to fine-tuned baselines -- reaching $53.0\%$ on the public validation set with an 8B-parameter LM and $61.9\%$ when ensembled with program-synthesis methods, matching average human performance. On BIG-Bench Hard (BBH), TTT on in-context examples surpasses standard few-shot prompting in the $10$-shot setting by $7.3$ percentage points ($50.5\%$ to $57.8\%$). Our findings highlight the limitations of in-context learning for novel tasks and demonstrate the potential of test-time training to enhance language model adaptability. △ Less

Submitted 24 March, 2025; v1 submitted 11 November, 2024; originally announced November 2024.

Comments: Preprint

LoRA vs Full Fine-tuning: An Illusion of Equivalence

Authors: Reece Shuttleworth, Jacob Andreas, Antonio Torralba, Pratyusha Sharma

Abstract: Fine-tuning is a crucial paradigm for adapting pre-trained large language models to downstream tasks. Recently, methods like Low-Rank Adaptation (LoRA) have been shown to effectively fine-tune LLMs with an extreme reduction in trainable parameters. But, \emph{are their learned solutions really equivalent?} We study how LoRA and full-finetuning change pre-trained models by analyzing the model's wei… ▽ More Fine-tuning is a crucial paradigm for adapting pre-trained large language models to downstream tasks. Recently, methods like Low-Rank Adaptation (LoRA) have been shown to effectively fine-tune LLMs with an extreme reduction in trainable parameters. But, \emph{are their learned solutions really equivalent?} We study how LoRA and full-finetuning change pre-trained models by analyzing the model's weight matrices through the lens of their spectral properties. We find that LoRA and full fine-tuning yield weight matrices whose singular value decompositions exhibit very different structure: weight matrices trained with LoRA have new, high-ranking singular vectors, which we call \emph{intruder dimensions}, while those trained with full fine-tuning do not. Further, we extend the finding that LoRA forgets less than full fine-tuning and find its forgetting is vastly localized to the intruder dimension -- by causally intervening on the intruder dimensions by changing their associated singular values post-fine-tuning, we show that they cause forgetting. Moreover, scaling them down significantly improves modeling of the pre-training distribution with a minimal drop in downstream task performance. Given this, we should expect accumulating intruder dimensions to be harmful and lead to more forgetting. This will be amplified during continual learning because of sequentially fine-tuning, and we show that LoRA models do accumulate intruder dimensions here tend to perform worse in this setting, emphasizing the practicality of our findings. △ Less

Submitted 22 October, 2025; v1 submitted 28 October, 2024; originally announced October 2024.

A Hitchhiker's Guide to Scaling Law Estimation

Authors: Leshem Choshen, Yang Zhang, Jacob Andreas

Abstract: Scaling laws predict the loss of a target machine learning model by extrapolating from easier-to-train models with fewer parameters or smaller training sets. This provides an efficient way for practitioners and researchers alike to compare pretraining decisions involving optimizers, datasets, and model architectures. Despite the widespread use of scaling laws to model the dynamics of language mode… ▽ More Scaling laws predict the loss of a target machine learning model by extrapolating from easier-to-train models with fewer parameters or smaller training sets. This provides an efficient way for practitioners and researchers alike to compare pretraining decisions involving optimizers, datasets, and model architectures. Despite the widespread use of scaling laws to model the dynamics of language model training, there has been little work on understanding how to best estimate and interpret them. We collect (and release) a large-scale dataset containing losses and downstream evaluations for 485 previously published pretrained models. We use these to estimate more than 1000 scaling laws, then derive a set of best practices for estimating scaling laws in new model families. We find that fitting scaling laws to intermediate checkpoints of training runs (and not just their final losses) substantially improves accuracy, and that -- all else equal -- estimates of performance are generally most accurate when derived from other models of similar sizes. However, because there is a significant degree of variability across model seeds, training multiple small models is sometimes more useful than training a single large one. Moreover, while different model families differ scaling behavior, they are often similar enough that a target model's behavior can be predicted from a single model with the same architecture, along with scaling parameter estimates derived from other model families. △ Less

Submitted 2 June, 2025; v1 submitted 15 October, 2024; originally announced October 2024.

Comments: ICML

Learning Linear Attention in Polynomial Time

Authors: Morris Yau, Ekin Akyürek, Jiayuan Mao, Joshua B. Tenenbaum, Stefanie Jegelka, Jacob Andreas

Abstract: Previous research has explored the computational expressivity of Transformer models in simulating Boolean circuits or Turing machines. However, the learnability of these simulators from observational data has remained an open question. Our study addresses this gap by providing the first polynomial-time learnability results (specifically strong, agnostic PAC learning) for single-layer Transformers… ▽ More Previous research has explored the computational expressivity of Transformer models in simulating Boolean circuits or Turing machines. However, the learnability of these simulators from observational data has remained an open question. Our study addresses this gap by providing the first polynomial-time learnability results (specifically strong, agnostic PAC learning) for single-layer Transformers with linear attention. We show that linear attention may be viewed as a linear predictor in a suitably defined RKHS. As a consequence, the problem of learning any linear transformer may be converted into the problem of learning an ordinary linear predictor in an expanded feature space, and any such predictor may be converted back into a multiheaded linear transformer. Moving to generalization, we show how to efficiently identify training datasets for which every empirical risk minimizer is equivalent (up to trivial symmetries) to the linear Transformer that generated the data, thereby guaranteeing the learned model will correctly generalize across all inputs. Finally, we provide examples of computations expressible via linear attention and therefore polynomial-time learnable, including associative memories, finite automata, and a class of Universal Turing Machine (UTMs) with polynomially bounded computation histories. We empirically validate our theoretical findings on three tasks: learning random linear attention networks, key--value associations, and learning to execute finite automata. Our findings bridge a critical gap between theoretical expressivity and learnability of Transformers, and show that flexible and general models of computation are efficiently learnable. △ Less

Submitted 23 October, 2025; v1 submitted 13 October, 2024; originally announced October 2024.

MisinfoEval: Generative AI in the Era of "Alternative Facts"

Authors: Saadia Gabriel, Liang Lyu, James Siderius, Marzyeh Ghassemi, Jacob Andreas, Asu Ozdaglar

Abstract: The spread of misinformation on social media platforms threatens democratic processes, contributes to massive economic losses, and endangers public health. Many efforts to address misinformation focus on a knowledge deficit model and propose interventions for improving users' critical thinking through access to facts. Such efforts are often hampered by challenges with scalability, and by platform… ▽ More The spread of misinformation on social media platforms threatens democratic processes, contributes to massive economic losses, and endangers public health. Many efforts to address misinformation focus on a knowledge deficit model and propose interventions for improving users' critical thinking through access to facts. Such efforts are often hampered by challenges with scalability, and by platform users' personal biases. The emergence of generative AI presents promising opportunities for countering misinformation at scale across ideological barriers. In this paper, we introduce a framework (MisinfoEval) for generating and comprehensively evaluating large language model (LLM) based misinformation interventions. We present (1) an experiment with a simulated social media environment to measure effectiveness of misinformation interventions, and (2) a second experiment with personalized explanations tailored to the demographics and beliefs of users with the goal of countering misinformation by appealing to their pre-existing values. Our findings confirm that LLM-based interventions are highly effective at correcting user behavior (improving overall user accuracy at reliability labeling by up to 41.72%). Furthermore, we find that users favor more personalized interventions when making decisions about news reliability and users shown personalized interventions have significantly higher accuracy at identifying misinformation. △ Less

Submitted 14 October, 2024; v1 submitted 13 October, 2024; originally announced October 2024.

Comments: EMNLP 2024. Correspondence can be sent to skgabrie at cs dot ucla dot edu

Learning How Hard to Think: Input-Adaptive Allocation of LM Computation

Authors: Mehul Damani, Idan Shenfeld, Andi Peng, Andreea Bobu, Jacob Andreas

Abstract: Computationally intensive decoding procedures--including search, reranking, and self-critique--can improve the quality of language model (LM) outputs in problems spanning code generation, numerical reasoning, and dialog. Existing work typically applies the same decoding procedure for every input to an LM. But not all inputs require the same amount of computation to process. Can we allocate decodin… ▽ More Computationally intensive decoding procedures--including search, reranking, and self-critique--can improve the quality of language model (LM) outputs in problems spanning code generation, numerical reasoning, and dialog. Existing work typically applies the same decoding procedure for every input to an LM. But not all inputs require the same amount of computation to process. Can we allocate decoding computation adaptively, using more resources to answer questions whose answers will be harder to compute? We present an approach that predicts the distribution of rewards given an input and computation budget, then allocates additional computation to inputs for which it is predicted to be most useful. We apply this approach in two decoding procedures: first, an adaptive best-of-k procedure that dynamically selects the number of samples to generate as input to a reranker; second, a routing procedure that dynamically responds to a query using a decoding procedure that is expensive but accurate, or one that is cheaper but less capable. Across a suite of programming, mathematics, and dialog tasks, we show that accurate computation-allocation procedures can be learned, and reduce computation by up to 50% at no cost to response quality, or improve quality by up to 10% at a fixed computational budget. △ Less

Submitted 6 October, 2024; originally announced October 2024.

Algorithmic Capabilities of Random Transformers

Authors: Ziqian Zhong, Jacob Andreas

Abstract: Trained transformer models have been found to implement interpretable procedures for tasks like arithmetic and associative recall, but little is understood about how the circuits that implement these procedures originate during training. To what extent do they depend on the supervisory signal provided to models, and to what extent are they attributable to behavior already present in models at the… ▽ More Trained transformer models have been found to implement interpretable procedures for tasks like arithmetic and associative recall, but little is understood about how the circuits that implement these procedures originate during training. To what extent do they depend on the supervisory signal provided to models, and to what extent are they attributable to behavior already present in models at the beginning of training? To investigate these questions, we investigate what functions can be learned by randomly initialized transformers in which only the embedding layers are optimized, so that the only input--output mappings learnable from data are those already implemented (up to a choice of encoding scheme) by the randomly initialized model. We find that these random transformers can perform a wide range of meaningful algorithmic tasks, including modular arithmetic, in-weights and in-context associative recall, decimal addition, parenthesis balancing, and even some aspects of natural language text generation. Our results indicate that some algorithmic capabilities are present in transformers (and accessible via appropriately structured inputs) even before these models are trained. Code is available at https://github.com/fjzzq2002/random_transformers. △ Less

Submitted 6 October, 2024; originally announced October 2024.

Comments: Accepted by NeurIPS 2024

Adaptive Language-Guided Abstraction from Contrastive Explanations

Authors: Andi Peng, Belinda Z. Li, Ilia Sucholutsky, Nishanth Kumar, Julie A. Shah, Jacob Andreas, Andreea Bobu

Abstract: Many approaches to robot learning begin by inferring a reward function from a set of human demonstrations. To learn a good reward, it is necessary to determine which features of the environment are relevant before determining how these features should be used to compute reward. End-to-end methods for joint feature and reward learning (e.g., using deep networks or program synthesis techniques) ofte… ▽ More Many approaches to robot learning begin by inferring a reward function from a set of human demonstrations. To learn a good reward, it is necessary to determine which features of the environment are relevant before determining how these features should be used to compute reward. End-to-end methods for joint feature and reward learning (e.g., using deep networks or program synthesis techniques) often yield brittle reward functions that are sensitive to spurious state features. By contrast, humans can often generalizably learn from a small number of demonstrations by incorporating strong priors about what features of a demonstration are likely meaningful for a task of interest. How do we build robots that leverage this kind of background knowledge when learning from new demonstrations? This paper describes a method named ALGAE (Adaptive Language-Guided Abstraction from [Contrastive] Explanations) which alternates between using language models to iteratively identify human-meaningful features needed to explain demonstrated behavior, then standard inverse reinforcement learning techniques to assign weights to these features. Experiments across a variety of both simulated and real-world robot environments show that ALGAE learns generalizable reward functions defined on interpretable features using only small numbers of demonstrations. Importantly, ALGAE can recognize when features are missing, then extract and define those features without any human input -- making it possible to quickly and efficiently acquire rich representations of user behavior. △ Less

Submitted 13 September, 2024; v1 submitted 12 September, 2024; originally announced September 2024.

Comments: CoRL 2024

Unforgettable Generalization in Language Models

Authors: Eric Zhang, Leshem Choshen, Jacob Andreas

Abstract: When language models (LMs) are trained to forget (or "unlearn'') a skill, how precisely does their behavior change? We study the behavior of transformer LMs in which tasks have been forgotten via fine-tuning on randomized labels. Such LMs learn to generate near-random predictions for individual examples in the "training'' set used for forgetting. Across tasks, however, LMs exhibit extreme variabil… ▽ More When language models (LMs) are trained to forget (or "unlearn'') a skill, how precisely does their behavior change? We study the behavior of transformer LMs in which tasks have been forgotten via fine-tuning on randomized labels. Such LMs learn to generate near-random predictions for individual examples in the "training'' set used for forgetting. Across tasks, however, LMs exhibit extreme variability in whether LM predictions change on examples outside the training set. In some tasks (like entailment classification), forgetting generalizes robustly, and causes models to produce uninformative predictions on new task instances; in other tasks (like physical commonsense reasoning and scientific question answering) forgetting affects only the training examples, and models continue to perform the "forgotten'' task accurately even for examples very similar to those that appeared in the training set. Dataset difficulty is not predictive of whether a behavior can be forgotten; instead, generalization in forgetting is (weakly) predicted by the confidence of LMs' initial task predictions and the variability of LM representations of training data, with low confidence and low variability both associated with greater generalization. Perhaps most surprisingly, random-label forgetting appears to be somewhat insensitive to the contents of the training set: for example, models trained on science questions with random labels continue to answer other science questions accurately, but begin to produce random labels on entailment classification tasks. Finally, we show that even generalizable forgetting is shallow: linear probes trained on LMs' representations can still perform tasks reliably after forgetting. Our results highlight the difficulty and unpredictability of performing targeted skill removal from models via fine-tuning. △ Less

Submitted 5 August, 2026; v1 submitted 3 September, 2024; originally announced September 2024.

Comments: 18 pages, 9 figures, published in First Conference on Language Modeling 2024

Journal ref: First Conference on Language Modeling (2024)

Language Modeling with Editable External Knowledge

Authors: Belinda Z. Li, Emmy Liu, Alexis Ross, Abbas Zeitoun, Graham Neubig, Jacob Andreas

Abstract: When the world changes, so does the text that humans write about it. How do we build language models that can be easily updated to reflect these changes? One popular approach is retrieval-augmented generation, in which new documents are inserted into a knowledge base and retrieved during prediction for downstream tasks. Most prior work on these systems have focused on improving behavior during pre… ▽ More When the world changes, so does the text that humans write about it. How do we build language models that can be easily updated to reflect these changes? One popular approach is retrieval-augmented generation, in which new documents are inserted into a knowledge base and retrieved during prediction for downstream tasks. Most prior work on these systems have focused on improving behavior during prediction through better retrieval or reasoning. This paper introduces ERASE, which instead improves model behavior when new documents are acquired, by incrementally deleting or rewriting other entries in the knowledge base each time a document is added. In two new benchmark datasets evaluating models' ability to answer questions about a stream of news articles or conversations, ERASE improves accuracy relative to conventional retrieval-augmented generation by 7-13% (Mixtral-8x7B) and 6-10% (Llama-3-8B) absolute. Code and data are available at https://github.com/belindal/ERASE △ Less

Submitted 17 June, 2024; originally announced June 2024.

A Probabilistic Framework for LLM Hallucination Detection via Belief Tree Propagation

Authors: Bairu Hou, Yang Zhang, Jacob Andreas, Shiyu Chang

Abstract: This paper focuses on the task of hallucination detection, which aims to determine the truthfulness of LLM-generated statements. To address this problem, a popular class of methods utilize the LLM's self-consistencies in its beliefs in a set of logically related augmented statements generated by the LLM, which does not require external knowledge databases and can work with both white-box and black… ▽ More This paper focuses on the task of hallucination detection, which aims to determine the truthfulness of LLM-generated statements. To address this problem, a popular class of methods utilize the LLM's self-consistencies in its beliefs in a set of logically related augmented statements generated by the LLM, which does not require external knowledge databases and can work with both white-box and black-box LLMs. However, in many existing approaches, the augmented statements tend to be very monotone and unstructured, which makes it difficult to integrate meaningful information from the LLM beliefs in these statements. Also, many methods work with the binarized version of the LLM's belief, instead of the continuous version, which significantly loses information. To overcome these limitations, in this paper, we propose Belief Tree Propagation (BTProp), a probabilistic framework for LLM hallucination detection. BTProp introduces a belief tree of logically related statements by recursively decomposing a parent statement into child statements with three decomposition strategies, and builds a hidden Markov tree model to integrate the LLM's belief scores in these statements in a principled way. Experiment results show that our method improves baselines by 3%-9% (evaluated by AUROC and AUC-PR) on multiple hallucination detection benchmarks. Code is available at https://github.com/UCSB-NLP-Chang/BTProp. △ Less

Submitted 7 February, 2025; v1 submitted 11 June, 2024; originally announced June 2024.

Comments: NAACL 2025 (main conference)

Elements of World Knowledge (EWoK): A Cognition-Inspired Framework for Evaluating Basic World Knowledge in Language Models

Authors: Anna A. Ivanova, Aalok Sathe, Benjamin Lipkin, Unnathi Kumar, Setayesh Radkani, Thomas H. Clark, Carina Kauf, Jennifer Hu, R. T. Pramod, Gabriel Grand, Vivian Paulun, Maria Ryskina, Ekin Akyürek, Ethan Wilcox, Nafisa Rashid, Leshem Choshen, Roger Levy, Evelina Fedorenko, Joshua Tenenbaum, Jacob Andreas

Abstract: The ability to build and reason about models of the world is essential for situated language understanding. But evaluating world modeling capabilities in modern AI systems -- especially those based on language models -- has proven challenging, in large part because of the difficulty of disentangling conceptual knowledge about the world from knowledge of surface co-occurrence statistics. This paper… ▽ More The ability to build and reason about models of the world is essential for situated language understanding. But evaluating world modeling capabilities in modern AI systems -- especially those based on language models -- has proven challenging, in large part because of the difficulty of disentangling conceptual knowledge about the world from knowledge of surface co-occurrence statistics. This paper presents Elements of World Knowledge (EWoK), a framework for evaluating language models' understanding of the conceptual knowledge underlying world modeling. EWoK targets specific concepts from multiple knowledge domains known to be important for world modeling in humans, from social interactions (help, deceive) to spatial relations (left, right). Objects, agents, and locations in the items can be flexibly filled in, enabling easy generation of multiple controlled datasets. We then introduce EWoK-core-1.0, a dataset of 4,374 items covering 11 world knowledge domains. We evaluate 20 open-weights large language models (1.3B--70B parameters) and compare them with human performance. All tested models perform worse than humans, with results varying drastically across domains. Performance on social interactions and social properties was highest and performance on physical relations and spatial relations was lowest. Overall, this dataset highlights simple cases where even large models struggle and presents rich avenues for targeted research on LLM world modeling capabilities. △ Less

Submitted 3 July, 2025; v1 submitted 15 May, 2024; originally announced May 2024.

Comments: Accepted to Transactions of the ACL (TACL). Contains 25 pages (14 main), 6 figures. Visit http://ewok-core.github.io for data and code. Authors Anna Ivanova, Aalok Sathe, Benjamin Lipkin contributed equally

Learning Phonotactics from Linguistic Informants

Authors: Canaan Breiss, Alexis Ross, Amani Maina-Kilaas, Roger Levy, Jacob Andreas

Abstract: We propose an interactive approach to language learning that utilizes linguistic acceptability judgments from an informant (a competent language user) to learn a grammar. Given a grammar formalism and a framework for synthesizing data, our model iteratively selects or synthesizes a data-point according to one of a range of information-theoretic policies, asks the informant for a binary judgment, a… ▽ More We propose an interactive approach to language learning that utilizes linguistic acceptability judgments from an informant (a competent language user) to learn a grammar. Given a grammar formalism and a framework for synthesizing data, our model iteratively selects or synthesizes a data-point according to one of a range of information-theoretic policies, asks the informant for a binary judgment, and updates its own parameters in preparation for the next query. We demonstrate the effectiveness of our model in the domain of phonotactics, the rules governing what kinds of sound-sequences are acceptable in a language, and carry out two experiments, one with typologically-natural linguistic data and another with a range of procedurally-generated languages. We find that the information-theoretic policies that our model uses to select items to query the informant achieve sample efficiency comparable to, and sometimes greater than, fully supervised approaches. △ Less

Submitted 7 May, 2024; originally announced May 2024.