-
Off-Context GRPO: Learning to Reason on Hard Problems using Privileged Information
Authors:
Priyank Agrawal,
Ankur Samanta,
Shervin Ghasemlou,
Jalaj Bhandari,
Kavosh Asadi,
Daniel Jiang,
Aditya Modi
Abstract:
Reinforcement learning with verifiable rewards (RLVR) improves reasoning in large language models. Yet, typical RLVR approaches fail on difficult problems: when a model cannot generate any correct solutions, it receives \textit{zero} learning signal. Providing privileged guidance during training, such as solution prefixes, can help overcome this learning cliff by steering the model towards {correc…
▽ More
Reinforcement learning with verifiable rewards (RLVR) improves reasoning in large language models. Yet, typical RLVR approaches fail on difficult problems: when a model cannot generate any correct solutions, it receives \textit{zero} learning signal. Providing privileged guidance during training, such as solution prefixes, can help overcome this learning cliff by steering the model towards {correct solutions with non-zero reward}. {We call these rollouts \textit{off-context}: they are generated from a training prompt that contains privileged guidance, while the target objective is defined by the original prompt without that guidance.} {We introduce} Off-Context GRPO (OC-GRPO), a minimally modified variant of GRPO that uses guided rollouts but applies an importance-corrected objective to steer the update back toward the original unguided objective, avoiding the mismatch that destabilizes uncorrected guided training. Empirically, our algorithm achieves a 3.9\% absolute improvement (13.8\% relative gain) over vanilla GRPO on average across standard mathematical reasoning benchmarks with negligible additional cost.
△ Less
Submitted 21 July, 2026;
originally announced July 2026.
-
Dopamine: Brain Modes, Not Brains
Authors:
Shervin Ghasemlou
Abstract:
Parameter-efficient fine-tuning (PEFT) methods such as \lora{} adapt large pretrained models by adding small weight-space updates. While effective, weight deltas are hard to interpret mechanistically, and they do not directly expose \emph{which} internal computations are reused versus bypassed for a new task. We explore an alternative view inspired by neuromodulation: adaptation as a change in \em…
▽ More
Parameter-efficient fine-tuning (PEFT) methods such as \lora{} adapt large pretrained models by adding small weight-space updates. While effective, weight deltas are hard to interpret mechanistically, and they do not directly expose \emph{which} internal computations are reused versus bypassed for a new task. We explore an alternative view inspired by neuromodulation: adaptation as a change in \emph{mode} -- selecting and rescaling existing computations -- rather than rewriting the underlying weights. We propose \methodname{}, a simple activation-space PEFT technique that freezes base weights and learns per-neuron \emph{thresholds} and \emph{gains}. During training, a smooth gate decides whether a neuron's activation participates; at inference the gate can be hardened to yield explicit conditional computation and neuron-level attributions.
As a proof of concept, we study ``mode specialization'' on MNIST (0$^\circ$) versus rotated MNIST (45$^\circ$). We pretrain a small MLP on a 50/50 mixture (foundation), freeze its weights, and then specialize to the rotated mode using \methodname{}. Across seeds, \methodname{} improves rotated accuracy over the frozen baseline while using only a few hundred trainable parameters per layer, and exhibits partial activation sparsity (a minority of units strongly active). Compared to \lora{}, \methodname{} trades some accuracy for substantially fewer trainable parameters and a more interpretable ``which-neurons-fire'' mechanism. We discuss limitations, including reduced expressivity when the frozen base lacks features needed for the target mode.
△ Less
Submitted 12 February, 2026;
originally announced February 2026.
-
Multimodal Generative Recommendation for Fusing Semantic and Collaborative Signals
Authors:
Moritz Vandenhirtz,
Kaveh Hassani,
Shervin Ghasemlou,
Shuai Shao,
Hamid Eghbalzadeh,
Fuchun Peng,
Jun Liu,
Michael Louis Iuzzolino
Abstract:
Sequential recommender systems rank relevant items by modeling a user's interaction history and computing the inner product between the resulting user representation and stored item embeddings. To avoid the significant memory overhead of storing large item sets, the generative recommendation paradigm instead models each item as a series of discrete semantic codes. Here, the next item is predicted…
▽ More
Sequential recommender systems rank relevant items by modeling a user's interaction history and computing the inner product between the resulting user representation and stored item embeddings. To avoid the significant memory overhead of storing large item sets, the generative recommendation paradigm instead models each item as a series of discrete semantic codes. Here, the next item is predicted by an autoregressive model that generates the code sequence corresponding to the predicted item. However, despite promising ranking capabilities on small datasets, these methods have yet to surpass traditional sequential recommenders on large item sets, limiting their adoption in the very scenarios they were designed to address. To resolve this, we propose MSCGRec, a Multimodal Semantic and Collaborative Generative Recommender. MSCGRec incorporates multiple semantic modalities and introduces a novel self-supervised quantization learning approach for images based on the DINO framework. Additionally, MSCGRec fuses collaborative and semantic signals by extracting collaborative features from sequential recommenders and treating them as a separate modality. Finally, we propose constrained sequence learning that restricts the large output space during training to the set of permissible tokens. We empirically demonstrate on three large real-world datasets that MSCGRec outperforms both sequential and generative recommendation baselines and provide an extensive ablation study to validate the impact of each component.
△ Less
Submitted 3 February, 2026;
originally announced February 2026.
-
CRAG-MM: Multi-modal Multi-turn Comprehensive RAG Benchmark
Authors:
Jiaqi Wang,
Xiao Yang,
Kai Sun,
Parth Suresh,
Sanat Sharma,
Adam Czyzewski,
Derek Andersen,
Surya Appini,
Arkav Banerjee,
Sajal Choudhary,
Shervin Ghasemlou,
Ziqiang Guan,
Akil Iyer,
Haidar Khan,
Lingkun Kong,
Roy Luo,
Tiffany Ma,
Zhen Qiao,
David Tran,
Wenfang Xu,
Skyler Yeatman,
Chen Zhou,
Gunveer Gujral,
Yinglong Xia,
Shane Moon
, et al. (16 additional authors not shown)
Abstract:
Wearable devices such as smart glasses are transforming the way people interact with their surroundings, enabling users to seek information regarding entities in their view. Multi-Modal Retrieval-Augmented Generation (MM-RAG) plays a key role in supporting such questions, yet there is still no comprehensive benchmark for this task, especially regarding wearables scenarios. To fill this gap, we pre…
▽ More
Wearable devices such as smart glasses are transforming the way people interact with their surroundings, enabling users to seek information regarding entities in their view. Multi-Modal Retrieval-Augmented Generation (MM-RAG) plays a key role in supporting such questions, yet there is still no comprehensive benchmark for this task, especially regarding wearables scenarios. To fill this gap, we present CRAG-MM -- a Comprehensive RAG benchmark for Multi-modal Multi-turn conversations. CRAG-MM contains a diverse set of 6.5K (image, question, answer) triplets and 2K visual-based multi-turn conversations across 13 domains, including 6.2K egocentric images designed to mimic captures from wearable devices. We carefully constructed the questions to reflect real-world scenarios and challenges, including five types of image-quality issues, six question types, varying entity popularity, differing information dynamism, and different conversation turns. We design three tasks: single-source augmentation, multi-source augmentation, and multi-turn conversations -- each paired with an associated retrieval corpus and APIs for both image-KG retrieval and webpage retrieval. Our evaluation shows that straightforward RAG approaches achieve only 32% and 43% truthfulness on CRAG-MM single- and multi-turn QA, respectively, whereas state-of-the-art industry solutions have similar quality (32%/45%), underscoring ample room for improvement. The benchmark has hosted KDD Cup 2025, attracting about 1K participants and 5K submissions, with winning solutions improving baseline performance by 28%, highlighting its early impact on advancing the field.
△ Less
Submitted 30 October, 2025;
originally announced October 2025.
-
Doppelgänger's Watch: A Split Objective Approach to Large Language Models
Authors:
Shervin Ghasemlou,
Ashish Katiyar,
Aparajita Saraf,
Seungwhan Moon,
Mangesh Pujari,
Pinar Donmez,
Babak Damavandi,
Anuj Kumar
Abstract:
In this paper, we investigate the problem of "generation supervision" in large language models, and present a novel bicameral architecture to separate supervision signals from their core capability, helpfulness. Doppelgänger, a new module parallel to the underlying language model, supervises the generation of each token, and learns to concurrently predict the supervision score(s) of the sequences…
▽ More
In this paper, we investigate the problem of "generation supervision" in large language models, and present a novel bicameral architecture to separate supervision signals from their core capability, helpfulness. Doppelgänger, a new module parallel to the underlying language model, supervises the generation of each token, and learns to concurrently predict the supervision score(s) of the sequences up to and including each token. In this work, we present the theoretical findings, and leave the report on experimental results to a forthcoming publication.
△ Less
Submitted 9 September, 2024;
originally announced September 2024.
-
Toward a language-theoretic foundation for planning and filtering
Authors:
Fatemeh Zahra Saberifar,
Shervin Ghasemlou,
Dylan A. Shell,
Jason M. O'Kane
Abstract:
We address problems underlying the algorithmic question of automating the co-design of robot hardware in tandem with its apposite software. Specifically, we consider the impact that degradations of a robot's sensor and actuation suites may have on the ability of that robot to complete its tasks. We introduce a new formal structure that generalizes and consolidates a variety of well-known structure…
▽ More
We address problems underlying the algorithmic question of automating the co-design of robot hardware in tandem with its apposite software. Specifically, we consider the impact that degradations of a robot's sensor and actuation suites may have on the ability of that robot to complete its tasks. We introduce a new formal structure that generalizes and consolidates a variety of well-known structures including many forms of plans, planning problems, and filters, into a single data structure called a procrustean graph, and give these graph structures semantics in terms of ideas based in formal language theory. We describe a collection of operations on procrustean graphs (both semantics-preserving and semantics-mutating), and show how a family of questions about the destructiveness of a change to the robot hardware can be answered by applying these operations. We also highlight the connections between this new approach and existing threads of research, including combinatorial filtering, Erdmann's strategy complexes, and hybrid automata.
△ Less
Submitted 23 July, 2018;
originally announced July 2018.