Showing 1–50 of 117 results for author: Hazra, R

AI Research Preference Models

Authors: Thomas Simon Foster, Bassel Al Omari, Tingchen Fu, Thomas Mann, Carl Domond, Lucia Cipolina-Kun, Bhavul Gauri, Muna Aghamelu, Alexander D. Goldie, Eryk Helenowski, Jean-Christophe Gagnon-Audet, Alberto Pepe, Saba Nazir, Daniel Izcovich, Noam Levi, Rishi Hazra, Karen Hambardzumyan, Nicolas Baldwin, Xian Li, Martin Josifoski, Paris Giampouras, Masoud Jalili Sabet, Anya Sims, Hela Momand, Tatiana Shavrina , et al. (8 additional authors not shown)

Abstract: AI research agents (AIRA) can now propose, implement, and evaluate their own machine learning experiments, but progress on frontier tasks is throttled by cost: a candidate solution can be written in minutes, whereas evaluating it can take hours to days of GPU time. An agent can therefore propose far more candidates than it can afford to run, and its progress depends on its research preference: how… ▽ More AI research agents (AIRA) can now propose, implement, and evaluate their own machine learning experiments, but progress on frontier tasks is throttled by cost: a candidate solution can be written in minutes, whereas evaluating it can take hours to days of GPU time. An agent can therefore propose far more candidates than it can afford to run, and its progress depends on its research preference: how it allocates a fixed execution budget across many candidates. We introduce AI Research Preference Models (RPMs) that predict which of multiple candidate solutions are most worth executing, without paying the cost of executing them all. We build RPMs from frozen pretrained language models (with no task-specific training), in two forms: an inference-only model that reasons over candidate plans, code, and prior executed solutions, and an agentic model that additionally runs small-scale pilot experiments before deciding. We integrate both into the AIRA-dojo search agent and evaluate on AIRS-Bench, a recent benchmark of machine learning research tasks for AI research agents. The two variants raise the average normalized score from 0.684 to 0.711 and 0.729 respectively, and reach the unguided agent's 24-hour performance in roughly 15 hours, using less than two-thirds of its execution budget. Our best RPMs also yield new state-of-the-art results on two AIRS-Bench tasks. △ Less

Submitted 14 August, 2026; originally announced August 2026.

Comments: 34 pages, 17 figures, 6 tables

The Generalized Friendship Paradox for Eigenvectors

Authors: Bishakh Bhattacharya, Arijit Chakrabarty, Rajat Subhra Hazra

Abstract: In this paper, we investigate the generalized friendship paradox for eigenvectors (alternatively called the eigen friendship paradox and abbreviated hereafter as EFP) in the setting of inhomogeneous Erdős--Rényi random graphs whose edge probabilities are generated by a continuous graphon. We consider the adjacency matrix of the graph and take the entries of the eigenvector corresponding to its lar… ▽ More In this paper, we investigate the generalized friendship paradox for eigenvectors (alternatively called the eigen friendship paradox and abbreviated hereafter as EFP) in the setting of inhomogeneous Erdős--Rényi random graphs whose edge probabilities are generated by a continuous graphon. We consider the adjacency matrix of the graph and take the entries of the eigenvector corresponding to its largest eigenvalue as the vertex attributes. It was shown in \cite{hazra2026generalized} that the generalized friendship paradox holds in this setting. We study the empirical distribution of the resulting bias values across the vertices and derive its limiting distribution explicitly in terms of the principal eigenvalue and the corresponding eigenfunction of the integral operator whose kernel is the underlying graphon. △ Less

Submitted 21 July, 2026; originally announced July 2026.

MSC Class: 05C80; 60B20; 60B10; 46L54

Renormalisation of Inhomogeneous Random Graphs

Authors: Luca Avena, Diego Garlaschelli, Rajat Subhra Hazra, Frank den Hollander

Abstract: We consider inhomogeneous random graphs in which vertices are assigned i.i.d.\ random weights, pairs of distinct vertices are connected by an edge independently with a probability that is a bi-variate function of the weights of the vertices, and single vertices are connected to themselves by a self-loop independently with a probability that is a uni-variate function of the weight of the vertex. We… ▽ More We consider inhomogeneous random graphs in which vertices are assigned i.i.d.\ random weights, pairs of distinct vertices are connected by an edge independently with a probability that is a bi-variate function of the weights of the vertices, and single vertices are connected to themselves by a self-loop independently with a probability that is a uni-variate function of the weight of the vertex. We apply a renormalisation transformation in which vertices are aggregated into groups of equal size according to a greedy algorithm, namely, distinct groups of aggregated vertices are connected by an aggregated edge if and only if there is at least one edge connecting two constituent vertices across the groups, while a group of aggregated vertices is connected to itself by an aggregated self-loop if and only if there is at least one self-loop at an internal vertex or one edge connecting a pair of distinct internal vertices. We analyse what happens when the renormalisation transformation is iterated. In particular, we show that, starting from appropriately scaled connection functions, the iterated renormalised graphs converge to a two-parameter family of random graphs, acting as an attractor in a universality class. We consider a light-tailed regime, for which the scaling limit is a homogeneous Erdős--Rényi random graph, and a heavy-tailed regime, for which the scaling limit is an inhomogeneous random graph with stable infinite-mean random weights and an exponential disconnection function. Different scalings are needed for the two regimes. Which of the two regimes prevails depends on the choice of the connection functions and the choice of the law of the random weights. △ Less

Submitted 14 July, 2026; originally announced July 2026.

Comments: 31 pages

Spectrum of Directed Inhomogeneous Random Graphs

Authors: Rajat Subhra Hazra, Giacomo Passuello

Abstract: We study the spectrum of the adjacency matrix $A_n$ of directed inhomogeneous random graphs on $n$ vertices. We assume that $A_n$ has independent entries and diverging average degree scale $s_n$. This framework includes, as special cases, the directed Chung--Lu random graph and directed stochastic block models. Assuming boundedness of the variance profile and that $s_n$ diverges faster than a suit… ▽ More We study the spectrum of the adjacency matrix $A_n$ of directed inhomogeneous random graphs on $n$ vertices. We assume that $A_n$ has independent entries and diverging average degree scale $s_n$. This framework includes, as special cases, the directed Chung--Lu random graph and directed stochastic block models. Assuming boundedness of the variance profile and that $s_n$ diverges faster than a suitable logarithmic function of $n$, we show that the rank-one Chung--Lu model satisfies a non-homogeneous version of the circular law, which in some situations allows for an explicit expression. Moreover, under mild conditions, we identify the asymptotic singular value distribution using tools from free probability. Finally, for finite-rank directed models, we prove the existence of eigenvalues outside the bulk and establish their joint Gaussian fluctuations at the scale $\sqrt{s_n/n}$, with an explicit covariance matrix. These results extend the theory of spectral outliers and their fluctuations to directed inhomogeneous random graphs. △ Less

Submitted 9 July, 2026; originally announced July 2026.

Comments: 35 pages, 2 figures

MSC Class: 60B20; 05C80

Lost in Interpretation: The Plausibility-Faithfulness Trade-off in Cross-Lingual Explanations

Authors: Somnath Banerjee, Pranav Jha, Rima Hazra, Animesh Mukherjee

Abstract: LLMs deployed multilingually are often audited via English explanations for non-English inputs. We evaluate extractive explanations ''where the model identifies input token spans as evidence alongside a generated rationale'' and uncover a systematic trade-off: English-pivot explanations can achieve higher span agreement with human rationales while their evidence becomes less causally grounded in t… ▽ More LLMs deployed multilingually are often audited via English explanations for non-English inputs. We evaluate extractive explanations ''where the model identifies input token spans as evidence alongside a generated rationale'' and uncover a systematic trade-off: English-pivot explanations can achieve higher span agreement with human rationales while their evidence becomes less causally grounded in the model's prediction, as measured by both comprehensiveness and sufficiency. Across 3 tasks, 5~languages, and 2~multilingual LLM families, we find that English explanations frequently produce fluent but loosely anchored rationales, with comprehensiveness degrading by up to 5.7x relative to native-language conditions - even as task accuracy remains stable across settings. For socially nuanced classification, English pivots also fail to preserve pragmatic cues, reducing both faithfulness and span agreement. We recommend auditing explanations in the input language, reporting multi-faceted faithfulness metrics beyond lexical overlap, and treating English rationales as communication summaries rather than faithful decision traces. △ Less

Submitted 18 May, 2026; originally announced May 2026.

COvolve: Adversarial Co-Evolution of Large-Language-Model-Generated Policies and Environments via Two-Player Zero-Sum Game

Authors: Alkis Sygkounas, Rishi Hazra, Andreas Persson, Pedro Zuidberg Dos Martires, Amy Loutfi

Abstract: A central challenge in building continually improving agents is that training environments are typically static or manually constructed. This restricts continual learning and generalization beyond the training distribution. We address this with COvolve, a co-evolutionary framework that leverages large language models (LLMs) to generate both environments and agent policies, expressed as executable… ▽ More A central challenge in building continually improving agents is that training environments are typically static or manually constructed. This restricts continual learning and generalization beyond the training distribution. We address this with COvolve, a co-evolutionary framework that leverages large language models (LLMs) to generate both environments and agent policies, expressed as executable Python code. We model the interaction between environment and policy designers as a two-player zero-sum game, ensuring adversarial co-evolution in which environments expose policy weaknesses and policies adapt in response. This process induces an automated curriculum in which environments and policies co-evolve toward increasing complexity. To guarantee robustness and prevent forgetting as the curriculum progresses, we compute the mixed-strategy Nash equilibrium (MSNE) of the zero-sum game, thereby yielding a meta-policy. This MSNE meta-policy ensures that the agent does not forget to solve previously seen environments while learning to solve previously unseen ones. Experiments in urban driving, symbolic maze-solving, and geometric navigation showcase that COvolve produces progressively more complex environments. Our results demonstrate the potential of LLM-driven co-evolution to achieve open-ended learning without predefined task distributions or manual intervention. △ Less

Submitted 30 March, 2026; originally announced March 2026.

Comments: Accepted at GECCO 2026

AIRA_2: Overcoming Bottlenecks in AI Research Agents

Authors: Karen Hambardzumyan, Nicolas Baldwin, Edan Toledo, Rishi Hazra, Michael Kuchnik, Bassel Al Omari, Thomas Simon Foster, Anton Protopopov, Jean-Christophe Gagnon-Audet, Ishita Mediratta, Kelvin Niu, Michael Shvartsman, Alisia Lupidi, Alexis Audran-Reiss, Parth Pathak, Tatiana Shavrina, Despoina Magka, Hela Momand, Derek Dunfield, Nicola Cancedda, Pontus Stenetorp, Carole-Jean Wu, Jakob Nicolaus Foerster, Yoram Bachrach, Martin Josifoski

Abstract: Existing research has identified three structural performance bottlenecks in AI research agents: (1) synchronous single-GPU execution constrains sample throughput, limiting the benefit of search; (2) a generalization gap where validation-based selection causes overfitting and performance to degrade over extended search horizons; and (3) the limited capability of fixed, single-turn LLM operators im… ▽ More Existing research has identified three structural performance bottlenecks in AI research agents: (1) synchronous single-GPU execution constrains sample throughput, limiting the benefit of search; (2) a generalization gap where validation-based selection causes overfitting and performance to degrade over extended search horizons; and (3) the limited capability of fixed, single-turn LLM operators imposes a ceiling on search performance. We introduce AIRA$_2$, which addresses these bottlenecks through three architectural choices: an asynchronous multi-GPU worker pool that increases experiment throughput linearly; a Hidden Consistent Evaluation protocol that delivers a reliable evaluation signal; and ReAct agents that dynamically scope their actions and debug interactively. On MLE-bench-30, AIRA$^{\dagger}_{2}$ achieves a mean Percentile Rank of 81.5% at 24 hours and 83.1% at 72 hours, outperforming the strongest baseline, which achieves 72.7%. On AIRS-Bench, AIRA$_2$ exceeds human state-of-the-art on 6 out of 20 diverse research tasks. Ablations confirm that each architectural component is necessary, that performance follows a predictable scaling law that transfers across LLM backbones, and that the "overfitting" reported in prior work was driven by evaluation noise rather than true data memorization. △ Less

Submitted 13 April, 2026; v1 submitted 27 March, 2026; originally announced March 2026.

SafeMath: Inference-time Safety improves Math Accuracy

Authors: Sagnik Basu, Subhrajit Mitra, Aman Juneja, Somnath Banerjee, Rima Hazra, Animesh Mukherjee

Abstract: Recent research points toward LLMs being manipulated through adversarial and seemingly benign inputs, resulting in harmful, biased, or policy-violating outputs. In this paper, we study an underexplored issue concerning harmful and toxic mathematical word problems. We show that math questions, particularly those framed as natural language narratives, can serve as a subtle medium for propagating bia… ▽ More Recent research points toward LLMs being manipulated through adversarial and seemingly benign inputs, resulting in harmful, biased, or policy-violating outputs. In this paper, we study an underexplored issue concerning harmful and toxic mathematical word problems. We show that math questions, particularly those framed as natural language narratives, can serve as a subtle medium for propagating biased, unethical, or psychologically harmful content, with heightened risks in educational settings involving children. To support a systematic study of this phenomenon, we introduce ToxicGSM, a dataset of 1.9k arithmetic problems in which harmful or sensitive context is embedded while preserving mathematically well-defined reasoning tasks. Using this dataset, we audit the behaviour of existing LLMs and analyse the trade-offs between safety enforcement and mathematical correctness. We further propose SafeMath -- a safety alignment technique that reduces harmful outputs while maintaining, and in some cases improving, mathematical reasoning performance. Our results highlight the importance of disentangling linguistic harm from math reasoning and demonstrate that effective safety alignment need not come at the cost of accuracy. We release the source code and dataset at https://github.com/Swagnick99/SafeMath/tree/main. △ Less

Submitted 26 March, 2026; originally announced March 2026.

Comments: Submitted in ARR March 2026

SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems

Authors: Rima Hazra, Bikram Ghuku, Ilona Marchenko, Yaroslava Tokarieva, Sayan Layek, Somnath Banerjee, Julia Stoyanovich, Mykola Pechenizkiy

Abstract: Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic co… ▽ More Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction. △ Less

Submitted 18 March, 2026; originally announced March 2026.

Bridging the Multilingual Safety Divide: Efficient, Culturally-Aware Alignment for Global South Languages

Authors: Somnath Banerjee, Rima Hazra, Animesh Mukherjee

Abstract: Large language models (LLMs) are being deployed across the Global South, where everyday use involves low-resource languages, code-mixing, and culturally specific norms. Yet safety pipelines, benchmarks, and alignment still largely target English and a handful of high-resource languages, implicitly assuming safety and factuality ''transfer'' across languages. Evidence increasingly shows they do not… ▽ More Large language models (LLMs) are being deployed across the Global South, where everyday use involves low-resource languages, code-mixing, and culturally specific norms. Yet safety pipelines, benchmarks, and alignment still largely target English and a handful of high-resource languages, implicitly assuming safety and factuality ''transfer'' across languages. Evidence increasingly shows they do not. We synthesize recent findings indicating that (i) safety guardrails weaken sharply on low-resource and code-mixed inputs, (ii) culturally harmful behavior can persist even when standard toxicity scores look acceptable, and (iii) English-only knowledge edits and safety patches often fail to carry over to low-resource languages. In response, we outline a practical agenda for researchers and students in the Global South: parameter-efficient safety steering, culturally grounded evaluation and preference data, and participatory workflows that empower local communities to define and mitigate harm. Our aim is to make multilingual safety a core requirement-not an add-on-for equitable AI in underrepresented regions. △ Less

Submitted 14 February, 2026; originally announced February 2026.

Comments: Accepted to the EGSAI Workshop at AAAI 2026

From Fluent to Verifiable: Claim-Level Auditability for Deep Research Agents

Authors: Razeen A Rasheed, Somnath Banerjee, Animesh Mukherjee, Rima Hazra

Abstract: A deep research agent produces a fluent scientific report in minutes; a careful reader then tries to verify the main claims and discovers the real cost is not reading, but tracing: which sentence is supported by which passage, what was ignored, and where evidence conflicts. We argue that as research generation becomes cheap, auditability becomes the bottleneck, and the dominant risk shifts from is… ▽ More A deep research agent produces a fluent scientific report in minutes; a careful reader then tries to verify the main claims and discovers the real cost is not reading, but tracing: which sentence is supported by which passage, what was ignored, and where evidence conflicts. We argue that as research generation becomes cheap, auditability becomes the bottleneck, and the dominant risk shifts from isolated factual errors to scientifically styled outputs whose claim-evidence links are weak, missing, or misleading. This perspective proposes claim-level auditability as a first-class design and evaluation target for deep research agents, distills recurring long-horizon failure modes (objective drift, transient constraints, and unverifiable inference), and introduces the Auditable Autonomous Research (AAR) standard, a compact measurement framework that makes auditability testable via provenance coverage, provenance soundness, contradiction transparency, and audit effort. We then argue for semantic provenance with protocolized validation: persistent, queryable provenance graphs that encode claim--evidence relations (including conflicts) and integrate continuous validation during synthesis rather than after publication, with practical instrumentation patterns to support deployment at scale. △ Less

Submitted 14 February, 2026; originally announced February 2026.

Outlier eigenvalues and eigenvectors of generalized Wigner matrices with finite-rank perturbations

Authors: Bishakh Bhattacharya, Arijit Chakrabarty, Rajat Subhra Hazra

Abstract: A generalized Wigner matrix perturbed by a finite-rank deterministic matrix is considered. The fluctuations of the largest eigenvalues, which emerge outside the bulk of the spectrum, and the corresponding eigenvectors, are studied. Under certain assumptions on the perturbation and the matrix structure, we derive the first-order behavior of these eigenvalues and show that they are well separated fr… ▽ More A generalized Wigner matrix perturbed by a finite-rank deterministic matrix is considered. The fluctuations of the largest eigenvalues, which emerge outside the bulk of the spectrum, and the corresponding eigenvectors, are studied. Under certain assumptions on the perturbation and the matrix structure, we derive the first-order behavior of these eigenvalues and show that they are well separated from the bulk. The fluctuations of these eigenvalues are shown to follow a multivariate Gaussian distribution, and the asymptotic behavior of the associated eigenvectors is also studied. We prove central limit theorems that describe the asymptotic alignment of these eigenvectors with the perturbation's eigenvectors, as well as their Gaussian fluctuations around the origin for non-aligned components. Furthermore, we discuss the convergence of the eigenvector process in a Sobolev space framework. △ Less

Submitted 15 January, 2026; originally announced January 2026.

Comments: 61 pages. Preliminary version, comments are welcomed

Training AI Co-Scientists Using Rubric Rewards

Authors: Shashwat Goel, Rishi Hazra, Dulhan Jayalath, Timon Willi, Parag Jain, William F. Shen, Ilias Leontiadis, Francesco Barbieri, Yoram Bachrach, Jonas Geiping, Chenxi Whitehouse

Abstract: AI co-scientists are emerging as a tool to assist human researchers in achieving their research goals. A crucial feature of these AI co-scientists is the ability to generate a research plan given a set of aims and constraints. The plan may be used by researchers for brainstorming, or may even be implemented after further refinement. However, language models currently struggle to generate research… ▽ More AI co-scientists are emerging as a tool to assist human researchers in achieving their research goals. A crucial feature of these AI co-scientists is the ability to generate a research plan given a set of aims and constraints. The plan may be used by researchers for brainstorming, or may even be implemented after further refinement. However, language models currently struggle to generate research plans that follow all constraints and implicit requirements. In this work, we study how to leverage the vast corpus of existing research papers to train language models that generate better research plans. We build a scalable, diverse training corpus by automatically extracting research goals and goal-specific grading rubrics from papers across several domains. We then train models for research plan generation via reinforcement learning with self-grading. A frozen copy of the initial policy acts as the grader during training, with the rubrics creating a generator-verifier gap that enables improvements without external human supervision. To validate this approach, we conduct a study with human experts for machine learning research goals, spanning 225 hours. The experts prefer plans generated by our finetuned Qwen3-30B-A3B model over the initial model for 70% of research goals, and approve 84% of the automatically extracted goal-specific grading rubrics. To assess generality, we also extend our approach to research goals from medical papers, and new arXiv preprints, evaluating with a jury of frontier models. Our finetuning yields 12-22% relative improvements and significant cross-domain generalization, proving effective even in problem settings like medical research where execution feedback is infeasible. Together, these findings demonstrate the potential of a scalable, automated training recipe as a step towards improving general AI co-scientists. △ Less

Submitted 29 December, 2025; originally announced December 2025.

Comments: 11 pages in the main paper, total 119 including sample outputs in the Appendix

ProSocialAlign: Preference Conditioned Test Time Alignment in Language Models

Authors: Somnath Banerjee, Sayan Layek, Sayantan Adak, Mykola Pechenizkiy, Animesh Mukherjee, Rima Hazra

Abstract: Current language model safety paradigms often fall short in emotionally charged or high-stakes settings, where refusal-only approaches may alienate users and naive compliance can amplify risk. We propose ProSocialAlign, a test-time, parameter-efficient framework that steers generation toward safe, empathetic, and value-aligned responses without retraining the base model. We formalize five human-ce… ▽ More Current language model safety paradigms often fall short in emotionally charged or high-stakes settings, where refusal-only approaches may alienate users and naive compliance can amplify risk. We propose ProSocialAlign, a test-time, parameter-efficient framework that steers generation toward safe, empathetic, and value-aligned responses without retraining the base model. We formalize five human-centered objectives and cast safety as lexicographic constrained generation: first, applying hard constraints to eliminate harmful continuations; then optimizing for prosocial quality within the safe set. Our method combines (i) directional regulation, a harm-mitigation mechanism that subtracts a learned "harm vector" in parameter space, and (ii) preference-aware autoregressive reward modeling trained jointly across attributes with gradient conflict resolution, enabling fine-grained, user-controllable decoding. Empirical evaluations across five safety benchmarks demonstrate state-of-the-art performance, reducing unsafe leakage and boosting alignment to human values, with strong gains across multiple evaluation metrics. ProSocialAlign offers a robust and modular foundation for generating context-sensitive, safe, and human-aligned responses at inference time. △ Less

Submitted 6 December, 2025; originally announced December 2025.

Functional central limit theorem for subgraph counts in a dynamic random connection model

Authors: Rajat Subhra Hazra, Nikolai Kriukov, Michel Mandjes, Moritz Otto

Abstract: We prove a functional central limit theorem for subgraph counts in a dynamic version of the random connection model. To establish tightness, we develop a dynamic extension of the cumulant method. We prove a functional central limit theorem for subgraph counts in a dynamic version of the random connection model. To establish tightness, we develop a dynamic extension of the cumulant method. △ Less

Submitted 22 November, 2025; originally announced November 2025.

MSC Class: 05C80; 60D05; 60G55

SAM 3: Segment Anything with Concepts

Authors: Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman Rädle, Triantafyllos Afouras, Effrosyni Mavroudi, Katherine Xu, Tsung-Han Wu, Yu Zhou, Liliane Momeni , et al. (13 additional authors not shown)

Abstract: We present Segment Anything Model (SAM) 3, a unified model that detects, segments, and tracks objects in images and videos based on concept prompts, which we define as either short noun phrases (e.g., "yellow school bus"), image exemplars, or a combination of both. Promptable Concept Segmentation (PCS) takes such prompts and returns segmentation masks and unique identities for all matching object… ▽ More We present Segment Anything Model (SAM) 3, a unified model that detects, segments, and tracks objects in images and videos based on concept prompts, which we define as either short noun phrases (e.g., "yellow school bus"), image exemplars, or a combination of both. Promptable Concept Segmentation (PCS) takes such prompts and returns segmentation masks and unique identities for all matching object instances. To advance PCS, we build a scalable data engine that produces a high-quality dataset with 4M unique concept labels, including hard negatives, across images and videos. Our model consists of an image-level detector and a memory-based video tracker that share a single backbone. Recognition and localization are decoupled with a presence head, which boosts detection accuracy. SAM 3 doubles the accuracy of existing systems in both image and video PCS, and improves previous SAM capabilities on visual segmentation tasks. We open source SAM 3 along with our new Segment Anything with Concepts (SA-Co) benchmark for promptable concept segmentation. △ Less

Submitted 28 March, 2026; v1 submitted 20 November, 2025; originally announced November 2025.

What Does It Take to Be a Good AI Research Agent? Studying the Role of Ideation Diversity

Authors: Alexis Audran-Reiss, Jordi Armengol-Estapé, Karen Hambardzumyan, Amar Budhiraja, Martin Josifoski, Edan Toledo, Rishi Hazra, Despoina Magka, Michael Shvartsman, Parth Pathak, Justine T Kao, Lucia Cipolina-Kun, Bhavul Gauri, Jean-Christophe Gagnon-Audet, Emanuel Tewolde, Jenny Zhang, Taco Cohen, Yossi Adi, Tatiana Shavrina, Yoram Bachrach

Abstract: AI research agents offer the promise to accelerate scientific progress by automating the design, implementation, and training of machine learning models. However, the field is still in its infancy, and the key factors driving the success or failure of agent trajectories are not fully understood. We examine the role that ideation diversity plays in agent performance. First, we analyse agent traject… ▽ More AI research agents offer the promise to accelerate scientific progress by automating the design, implementation, and training of machine learning models. However, the field is still in its infancy, and the key factors driving the success or failure of agent trajectories are not fully understood. We examine the role that ideation diversity plays in agent performance. First, we analyse agent trajectories on MLE-bench, a well-known benchmark to evaluate AI research agents, across different models and agent scaffolds. Our analysis reveals that different models and agent scaffolds yield varying degrees of ideation diversity, and that higher-performing agents tend to have increased ideation diversity. Further, we run a controlled experiment where we modify the degree of ideation diversity, demonstrating that higher ideation diversity results in stronger performance. Finally, we strengthen our results by examining additional evaluation metrics beyond the standard medal-based scoring of MLE-bench, showing that our findings still hold across other agent performance metrics. △ Less

Submitted 9 December, 2025; v1 submitted 19 November, 2025; originally announced November 2025.

LexiCon: a Benchmark for Planning under Temporal Constraints in Natural Language

Authors: Periklis Mantenoglou, Rishi Hazra, Pedro Zuidberg Dos Martires, Luc De Raedt

Abstract: Owing to their reasoning capabilities, large language models (LLMs) have been evaluated on planning tasks described in natural language. However, LLMs have largely been tested on planning domains without constraints. In order to deploy them in real-world settings where adherence to constraints, in particular safety constraints, is critical, we need to evaluate their performance on constrained plan… ▽ More Owing to their reasoning capabilities, large language models (LLMs) have been evaluated on planning tasks described in natural language. However, LLMs have largely been tested on planning domains without constraints. In order to deploy them in real-world settings where adherence to constraints, in particular safety constraints, is critical, we need to evaluate their performance on constrained planning tasks. We introduce LexiCon -- a natural language-based (Lexi) constrained (Con) planning benchmark, consisting of a suite of environments, that can be used to evaluate the planning capabilities of LLMs in a principled fashion. The core idea behind LexiCon is to take existing planning environments and impose temporal constraints on the states. These constrained problems are then translated into natural language and given to an LLM to solve. A key feature of LexiCon is its extensibility. That is, the set of supported environments can be extended with new (unconstrained) environment generators, for which temporal constraints are constructed automatically. This renders LexiCon future-proof: the hardness of the generated planning problems can be increased as the planning capabilities of LLMs improve. Our experiments reveal that the performance of state-of-the-art LLMs, including reasoning models like GPT-5, o3, and R1, deteriorates as the degree of constrainedness of the planning tasks increases. △ Less

Submitted 7 October, 2025; originally announced October 2025.

Spectra of random graphs with discrete scale invariance

Authors: Alessio Catanzaro, Rajat Subhra Hazra, Diego Garlaschelli

Abstract: Random graphs defined by an occurrence probability that is invariant under node aggregation have been identified recently in the context of network renormalization. The invariance property requires that edges are drawn with a specific probability that, in the annealed case, depends on a necessarily infinite-mean node fitness. The diverging mean determines many properties that are uncommon in model… ▽ More Random graphs defined by an occurrence probability that is invariant under node aggregation have been identified recently in the context of network renormalization. The invariance property requires that edges are drawn with a specific probability that, in the annealed case, depends on a necessarily infinite-mean node fitness. The diverging mean determines many properties that are uncommon in models with independent edges, but at the same time widespread in real-world networks. Here we focus on the leading eigenvalues and eigenvectors of the adjacency matrix of the model, where the n nodes are assigned a Pareto($α$)-distributed fitness with 0 < $α$ < 1. We find that the leading eigenvalues are all of order square root of n, alternate in sign and are located at the intersection between the real axis and a logarithmic spiral in the complex plane, which we characterize analytically in terms of the Gamma function. We also calculate the associated eigenvectors, finding that they display complexvalued scaling exponents and log-periodicity, which are signatures of discrete scale invariance. In contrast with the typical finite-rank behaviour of random graphs with finite-mean variables, we find that a growing number of the leading eigenvalues emerges from the bulk, whose edge extends up to order square root of n and therefore reaches the same scale as that of the structural eigenvalues. △ Less

Submitted 20 August, 2026; v1 submitted 15 September, 2025; originally announced September 2025.

AURA: Affordance-Understanding and Risk-aware Alignment Technique for Large Language Models

Authors: Sayantan Adak, Pratyush Chatterjee, Somnath Banerjee, Rima Hazra, Somak Aditya, Animesh Mukherjee

Abstract: Present day LLMs face the challenge of managing affordance-based safety risks-situations where outputs inadvertently facilitate harmful actions due to overlooked logical implications. Traditional safety solutions, such as scalar outcome-based reward models, parameter tuning, or heuristic decoding strategies, lack the granularity and proactive nature needed to reliably detect and intervene during s… ▽ More Present day LLMs face the challenge of managing affordance-based safety risks-situations where outputs inadvertently facilitate harmful actions due to overlooked logical implications. Traditional safety solutions, such as scalar outcome-based reward models, parameter tuning, or heuristic decoding strategies, lack the granularity and proactive nature needed to reliably detect and intervene during subtle yet crucial reasoning steps. Addressing this fundamental gap, we introduce AURA, an innovative, multi-layered framework centered around Process Reward Models (PRMs), providing comprehensive, step level evaluations across logical coherence and safety-awareness. Our framework seamlessly combines introspective self-critique, fine-grained PRM assessments, and adaptive safety-aware decoding to dynamically and proactively guide models toward safer reasoning trajectories. Empirical evidence clearly demonstrates that this approach significantly surpasses existing methods, significantly improving the logical integrity and affordance-sensitive safety of model outputs. This research represents a pivotal step toward safer, more responsible, and contextually aware AI, setting a new benchmark for alignment-sensitive applications. △ Less

Submitted 8 August, 2025; originally announced August 2025.

The Generalized Friendship Paradox for Spectral Centralities

Authors: Rajat Subhra Hazra, Evgeny Verbitskiy

Abstract: We revisit the classical friendship paradox which states that on an average ones friends have at least as many friends as oneself and generalize it to a variety of network centrality indices. For a broad class of spectral centralities on connected undirected graphs degree, eigenvector centrality, walk counts, Katz centrality and PageRank, we show that the average centrality of a nodes neighbours a… ▽ More We revisit the classical friendship paradox which states that on an average ones friends have at least as many friends as oneself and generalize it to a variety of network centrality indices. For a broad class of spectral centralities on connected undirected graphs degree, eigenvector centrality, walk counts, Katz centrality and PageRank, we show that the average centrality of a nodes neighbours always exceeds the global average centrality. △ Less

Submitted 8 January, 2026; v1 submitted 17 July, 2025; originally announced July 2025.

Comments: 12 pages. Title changed. Typos corrected. To appear in Journal of Complex Networks

AI Research Agents for Machine Learning: Search, Exploration, and Generalization in MLE-bench

Authors: Edan Toledo, Karen Hambardzumyan, Martin Josifoski, Rishi Hazra, Nicolas Baldwin, Alexis Audran-Reiss, Michael Kuchnik, Despoina Magka, Minqi Jiang, Alisia Maria Lupidi, Andrei Lupu, Roberta Raileanu, Kelvin Niu, Tatiana Shavrina, Jean-Christophe Gagnon-Audet, Michael Shvartsman, Shagun Sodhani, Alexander H. Miller, Abhishek Charnalia, Derek Dunfield, Carole-Jean Wu, Pontus Stenetorp, Nicola Cancedda, Jakob Nicolaus Foerster, Yoram Bachrach

Abstract: AI research agents are demonstrating great potential to accelerate scientific progress by automating the design, implementation, and training of machine learning models. We focus on methods for improving agents' performance on MLE-bench, a challenging benchmark where agents compete in Kaggle competitions to solve real-world machine learning problems. We formalize AI research agents as search polic… ▽ More AI research agents are demonstrating great potential to accelerate scientific progress by automating the design, implementation, and training of machine learning models. We focus on methods for improving agents' performance on MLE-bench, a challenging benchmark where agents compete in Kaggle competitions to solve real-world machine learning problems. We formalize AI research agents as search policies that navigate a space of candidate solutions, iteratively modifying them using operators. By designing and systematically varying different operator sets and search policies (Greedy, MCTS, Evolutionary), we show that their interplay is critical for achieving high performance. Our best pairing of search strategy and operator set achieves a state-of-the-art result on MLE-bench lite, increasing the success rate of achieving a Kaggle medal from 39.6% to 47.7%. Our investigation underscores the importance of jointly considering the search strategy, operator design, and evaluation methodology in advancing automated machine learning. △ Less

Submitted 4 November, 2025; v1 submitted 3 July, 2025; originally announced July 2025.

Comments: Code: https://github.com/facebookresearch/aira-dojo

Voter model on heterogeneous directed networks

Authors: Luca Avena, Federico Capannoli, Diego Garlaschelli, Rajat Subhra Hazra

Abstract: We investigate the consensus dynamics of the voter model on large random graphs with heterogeneous and directed features, focusing in particular on networks with power-law degree distributions. By extending recent results on sparse directed graphs, we derive exact first-order asymptotics for the expected consensus time in directed configuration models with i.i.d. Pareto-distributed in- and out-deg… ▽ More We investigate the consensus dynamics of the voter model on large random graphs with heterogeneous and directed features, focusing in particular on networks with power-law degree distributions. By extending recent results on sparse directed graphs, we derive exact first-order asymptotics for the expected consensus time in directed configuration models with i.i.d. Pareto-distributed in- and out-degrees. For any tail exponent α>0, we derive the mean consensus time scaling depending on the network size and a pre-factor that encodes detailed structural properties of the degree sequences. We give an explicit description of the pre factor in the directed setting. This extends and sharpens previous mean-field predictions from statistical physics, providing the first explicit consensus-time formula in the directed heavy-tailed setting. Through extensive simulations, we confirm the validity of our predictions across a wide range of heterogeneity regimes, including networks with infinite variance and infinite mean degree distribution. We further explore the interplay between network topology and voter dynamics, highlighting how degree fluctuations and maximal degrees shape the consensus landscape. Complementing the asymptotic analysis, we provide numerical evidence for the emergence of Wright-Fisher diffusive behavior in both directed and undirected ensembles under suitable mixing conditions, and demonstrate the breakdown of this approximation in the in the infinite mean regime. △ Less

Submitted 18 June, 2025; v1 submitted 13 June, 2025; originally announced June 2025.

The friendship paradox for trees

Authors: Rajat Subhra Hazra, Frank den Hollander, Nelly Litvak, Azadeh Parvaneh

Abstract: We analyse the friendship paradox on finite and infinite trees. In particular, we monitor the vertices for which the friendship-bias is positive, neutral and negative, respectively. For an arbitrary finite tree, we show that the number of positive vertices is at least as large as the number of negative vertices, a property we refer to as significance, and derive a lower bound in terms of the branc… ▽ More We analyse the friendship paradox on finite and infinite trees. In particular, we monitor the vertices for which the friendship-bias is positive, neutral and negative, respectively. For an arbitrary finite tree, we show that the number of positive vertices is at least as large as the number of negative vertices, a property we refer to as significance, and derive a lower bound in terms of the branching points in the tree. For an infinite Galton-Watson tree, we compute the densities of the positive and the negative vertices and show that either may dominate the other, depending on the offspring distribution. We also compute the densities of the edges having two given types of vertices at their ends, and give conditions in terms of the offspring distribution under which these types are positively or negatively correlated. △ Less

Submitted 24 July, 2026; v1 submitted 27 May, 2025; originally announced May 2025.

Comments: To appear in Electronic Journal of Probability

Attributional Safety Failures in Large Language Models under Code-Mixed Perturbations

Authors: Somnath Banerjee, Pratyush Chatterjee, Shanu Kumar, Sayan Layek, Parag Agrawal, Rima Hazra, Animesh Mukherjee

Abstract: While LLMs appear robustly safety-aligned in English, we uncover a catastrophic, overlooked weakness: attributional collapse under code-mixed perturbations. Our systematic evaluation of open models shows that the linguistic camouflage of code-mixing -- ``blending languages within a single conversation'' -- can cause safety guardrails to fail dramatically. Attack success rates (ASR) spike from a be… ▽ More While LLMs appear robustly safety-aligned in English, we uncover a catastrophic, overlooked weakness: attributional collapse under code-mixed perturbations. Our systematic evaluation of open models shows that the linguistic camouflage of code-mixing -- ``blending languages within a single conversation'' -- can cause safety guardrails to fail dramatically. Attack success rates (ASR) spike from a benign 9\% in monolingual English to 69\% under code-mixed inputs, with rates exceeding 90\% in non-Western contexts such as Arabic and Hindi. These effects hold not only on controlled synthetic datasets but also on real-world social media traces, revealing a serious risk for billions of users. To explain why this happens, we introduce saliency drift attribution (SDA), an interpretability framework that shows how, under code-mixing, the model's internal attention drifts away from safety-critical tokens (e.g., ``violence'' or ``corruption''), effectively blinding it to harmful intent. Finally, we propose a lightweight translation-based restoration strategy that recovers roughly 80\% of the safety lost to code-mixing, offering a practical path toward more equitable and robust LLM safety. △ Less

Submitted 30 November, 2025; v1 submitted 20 May, 2025; originally announced May 2025.

MiniMax Learning of Interpretable Factored Stochastic Policies from Conjoint Data, with Uncertainty Quantification

Authors: Connor T. Jerzak, Priyanshi Chandra, Rishi Hazra

Abstract: We study offline policy optimization over exponentially large factorial action spaces from randomized preference data, showing how conjoint experiments can estimate interpretable stochastic policies with asymptotically valid uncertainty under regularity conditions. Conjoint analyses typically report Average Marginal Component Effects (AMCEs) by averaging over opponent attributes and thus ignore st… ▽ More We study offline policy optimization over exponentially large factorial action spaces from randomized preference data, showing how conjoint experiments can estimate interpretable stochastic policies with asymptotically valid uncertainty under regularity conditions. Conjoint analyses typically report Average Marginal Component Effects (AMCEs) by averaging over opponent attributes and thus ignore strategic interdependence. We instead learn stochastic interventions -- product-of-Categorical policies over factor levels -- that (i) optimize expected outcomes in an average-case setting and (ii) extend to a two-player minimax (adversarial) setting that realistically captures simultaneous strategic candidate selection. Methodologically, we derive a closed-form optimizer for a tractable two-way interaction regime with L2 variance regularization, and provide a general gradient-based procedure for richer model classes. Uncertainty from the outcome model propagates asymptotically to both the optimal policy and its value via a Delta method approximation. We further model institutional details (e.g., primaries) inside the minimax objective and introduce a data-driven measure of strategic divergence between parties. On synthetic data, we empirically characterize finite-sample error and coverage as dimensionality and $n$ vary. On a U.S. presidential conjoint, adversarially learned policies produce restricted-equilibrium vote shares that align with historical election ranges in our data, in stark contrast to non-adversarial (averaging) optimizers. △ Less

Submitted 29 May, 2026; v1 submitted 26 April, 2025; originally announced April 2025.

Comments: ICML 2026

MSC Class: 62P25 ACM Class: G.3; G.1.6; J.4

Spectral properties of the Laplacian of Scale-Free Percolation models

Authors: Rajat Subhra Hazra, Nandan Malhotra

Abstract: We consider scale-free percolation on a discrete torus $\mathbf{V}_N$ of size $N$. Conditionally on an i.i.d. sequence of Pareto weights $(W_i)_{i\in \mathbf{V}_N}$ with tail exponent $τ-1>0$, we connect any two points $i$ and $j$ on the torus with probability $$p_{ij}= \frac{W_iW_j}{\|i-j\|^α} \wedge 1$$ for some parameter $α>0$. We focus on the (centred) Laplacian operator of this random gra… ▽ More We consider scale-free percolation on a discrete torus $\mathbf{V}_N$ of size $N$. Conditionally on an i.i.d. sequence of Pareto weights $(W_i)_{i\in \mathbf{V}_N}$ with tail exponent $τ-1>0$, we connect any two points $i$ and $j$ on the torus with probability $$p_{ij}= \frac{W_iW_j}{\|i-j\|^α} \wedge 1$$ for some parameter $α>0$. We focus on the (centred) Laplacian operator of this random graph and study its empirical spectral distribution. We explicitly identify the limiting distribution when $α<1$ and $τ>3$, in terms of the spectral distribution of some non-commutative unbounded operators. △ Less

Submitted 24 November, 2025; v1 submitted 24 April, 2025; originally announced April 2025.

Comments: Revised version, Lemma 3.2, 3.3 corrected typos and the argument in section 5.3 expanded

Have Large Language Models Learned to Reason? A Characterization via 3-SAT Phase Transition

Authors: Rishi Hazra, Gabriele Venturato, Pedro Zuidberg Dos Martires, Luc De Raedt

Abstract: Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. In theory, autoregressive LLMs with Chain-of-Thought (CoT) can perform more serial computations to solve complex reasoning tasks. However, recent studies suggest that, despite this capacity, LLMs do not truly learn to reason but instead fit on statistical features. To study the reasoning capabilitie… ▽ More Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. In theory, autoregressive LLMs with Chain-of-Thought (CoT) can perform more serial computations to solve complex reasoning tasks. However, recent studies suggest that, despite this capacity, LLMs do not truly learn to reason but instead fit on statistical features. To study the reasoning capabilities in a principled fashion, we adopt a computational theory perspective and propose an experimental protocol centered on 3-SAT -- the prototypical NP-complete problem lying at the core of logical reasoning and constraint satisfaction tasks. Specifically, we examine the phase transitions in random 3-SAT and characterize the reasoning abilities of state-of-the-art LLMs by varying the inherent hardness of the problem instances. By comparing DeepSeek R1 with other LLMs, our findings reveal two key insights (1) LLM accuracy drops significantly on harder instances, suggesting all current models struggle when statistical shortcuts are unavailable (2) Unlike other LLMs, R1 shows signs of having learned the underlying reasoning. Following a principled experimental protocol, our study moves beyond the benchmark-driven evidence often found in LLM reasoning research. Our findings highlight important gaps and suggest clear directions for future research. △ Less

Submitted 4 April, 2025; originally announced April 2025.

Comments: An updated version of arXiv:2408.07215v2, featuring: (1) inclusion of recent LRMs and recent LLMs, (2) revised conclusions reflecting recent developments, and (3) updated analysis

The multi-level friendship paradox for sparse random graphs

Authors: Rajat Subhra Hazra, Frank den Hollander, Azadeh Parvaneh

Abstract: In Hazra, den Hollander and Parvaneh (2025) we analysed the friendship paradox for sparse random graphs. For four classes of random graphs we characterised the empirical distribution of the friendship biases between vertices and their neighbours at distance $1$, proving convergence as $n\to\infty$ to a limiting distribution, with $n$ the number of vertices, and identifying moments and tail exponen… ▽ More In Hazra, den Hollander and Parvaneh (2025) we analysed the friendship paradox for sparse random graphs. For four classes of random graphs we characterised the empirical distribution of the friendship biases between vertices and their neighbours at distance $1$, proving convergence as $n\to\infty$ to a limiting distribution, with $n$ the number of vertices, and identifying moments and tail exponents of the limiting distribution. In the present paper we look at the multi-level friendship bias between vertices and their neighbours at distance $k \in \mathbb{N}$ obtained via a $k$-step exploration according to a backtracking or a non-backtracking random walk. We identify the limit of empirical distribution of the multi-level friendship biases as $n\to\infty$ and/or $k\to\infty$. We show that for non-backtracking exploration the two limits commute for a large class of sparse random graphs, including those that locally converge to a rooted Galton-Watson tree. In particular, we show that the same limit arises when $k$ depends on $n$, i.e., $k=k_n$, provided $\lim_{n\to\infty} k_n = \infty$ under some mild conditions. We exhibit cases where the two limits do not commute and show the relevance of the mixing time of the exploration. △ Less

Submitted 10 January, 2026; v1 submitted 24 February, 2025; originally announced February 2025.

Comments: Accepted for publication in Stochastic Processes and their Applications

MSC Class: 05C80; 60C05; 60F15; 60J80; 60G50

Parameter estimation in a dynamic Chung-Lu random graph

Authors: Rajat Subhra Hazra, Michel Mandjes, Jiesen Wang

Abstract: In this paper we consider a dynamic version of the Chung-Lu random graph in which the edges alternate between being present and absent. The main contribution concerns a technique by which one can estimate the underlying dynamics from partial information, in particular from snapshots of the total number of edges present. The efficacy of our inference method is demonstrated through a series of numer… ▽ More In this paper we consider a dynamic version of the Chung-Lu random graph in which the edges alternate between being present and absent. The main contribution concerns a technique by which one can estimate the underlying dynamics from partial information, in particular from snapshots of the total number of edges present. The efficacy of our inference method is demonstrated through a series of numerical experiments. △ Less

Submitted 25 July, 2025; v1 submitted 17 February, 2025; originally announced February 2025.

MemeSense: An Adaptive In-Context Framework for Social Commonsense Driven Meme Moderation

Authors: Sayantan Adak, Somnath Banerjee, Rajarshi Mandal, Avik Halder, Sayan Layek, Rima Hazra, Animesh Mukherjee

Abstract: Online memes are a powerful yet challenging medium for content moderation, often masking harmful intent behind humor, irony, or cultural symbolism. Conventional moderation systems "especially those relying on explicit text" frequently fail to recognize such subtle or implicit harm. We introduce MemeSense, an adaptive framework designed to generate socially grounded interventions for harmful memes… ▽ More Online memes are a powerful yet challenging medium for content moderation, often masking harmful intent behind humor, irony, or cultural symbolism. Conventional moderation systems "especially those relying on explicit text" frequently fail to recognize such subtle or implicit harm. We introduce MemeSense, an adaptive framework designed to generate socially grounded interventions for harmful memes by combining visual and textual understanding with curated, semantically aligned examples enriched with commonsense cues. This enables the model to detect nuanced complexed threats like misogyny, stereotyping, or vulgarity "even in memes lacking overt language". Across multiple benchmark datasets, MemeSense outperforms state-of-the-art methods, achieving up to 35% higher semantic similarity and 9% improvement in BERTScore for non-textual memes, and notable gains for text-rich memes as well. These results highlight MemeSense as a promising step toward safer, more context-aware AI systems for real-world content moderation. Code and data available at: https://github.com/sayantan11995/MemeSense △ Less

Submitted 17 October, 2025; v1 submitted 16 February, 2025; originally announced February 2025.

Comments: Accepted at Transactions on Machine Learning Research (TMLR)

Soteria: Language-Specific Functional Parameter Steering for Multilingual Safety Alignment

Authors: Somnath Banerjee, Sayan Layek, Pratyush Chatterjee, Animesh Mukherjee, Rima Hazra

Abstract: Ensuring consistent safety across multiple languages remains a significant challenge for large language models (LLMs). We introduce Soteria, a lightweight yet powerful strategy that locates and minimally adjusts the "functional heads" most responsible for harmful content generation in each language. By altering only a fraction of parameters, Soteria drastically reduces policy violations without sa… ▽ More Ensuring consistent safety across multiple languages remains a significant challenge for large language models (LLMs). We introduce Soteria, a lightweight yet powerful strategy that locates and minimally adjusts the "functional heads" most responsible for harmful content generation in each language. By altering only a fraction of parameters, Soteria drastically reduces policy violations without sacrificing overall model performance, even in low-resource settings. To rigorously evaluate our approach, we also present XThreatBench, a specialized multilingual dataset capturing fine-grained harmful behaviors drawn from real policy guidelines. Experiments with leading open-source LLMs (e.g., Llama, Qwen, Mistral) show that Soteria consistently improves safety metrics across high-, mid-, and low-resource languages. These findings highlight a promising path toward scalable, linguistically attuned, and ethically aligned LLMs worldwide. △ Less

Submitted 22 August, 2025; v1 submitted 16 February, 2025; originally announced February 2025.

Comments: Accepted at EMNLP 2025

The spectrum of dense kernel-based random graphs

Authors: Alessandra Cipriani, Rajat Subhra Hazra, Nandan Malhotra, Michele Salvi

Abstract: Kernel-based random graphs (KBRGs) are a broad class of random graph models that account for inhomogeneity among vertices. We consider KBRGs on a discrete $d-$dimensional torus $\mathbf{V}_N$ of size $N^d$. Conditionally on an i.i.d.~sequence of {Pareto} weights $(W_i)_{i\in \mathbf{V}_N}$ with tail exponent $τ-1>0$, we connect any two points $i$ and $j$ on the torus with probability… ▽ More Kernel-based random graphs (KBRGs) are a broad class of random graph models that account for inhomogeneity among vertices. We consider KBRGs on a discrete $d-$dimensional torus $\mathbf{V}_N$ of size $N^d$. Conditionally on an i.i.d.~sequence of {Pareto} weights $(W_i)_{i\in \mathbf{V}_N}$ with tail exponent $τ-1>0$, we connect any two points $i$ and $j$ on the torus with probability $$p_{ij}= \frac{κ_σ(W_i,W_j)}{\|i-j\|^α} \wedge 1$$ for some parameter $α>0$ and $κ_σ(u,v)= (u\vee v)(u \wedge v)^σ$ for some $σ\in(0,τ-1)$. We focus on the adjacency operator of this random graph and study its empirical spectral distribution. For $α<d$ and $τ>2$, we show that a non-trivial limiting distribution exists as $N\to\infty$ and that the corresponding measure $μ_{σ,τ}$ is absolutely continuous with respect to the Lebesgue measure. $μ_{σ,τ}$ is given by an operator-valued semicircle law, whose Stieltjes transform is characterised by a fixed point equation in an appropriate Banach space. We analyse the moments of $μ_{σ,τ}$ and prove that the second moment is finite even when the weights have infinite variance. In the case $σ=1$, corresponding to the so-called scale-free percolation random graph, we can explicitly describe the limiting measure and study its tail. △ Less

Submitted 14 March, 2025; v1 submitted 13 February, 2025; originally announced February 2025.

Comments: 48 pages, 6 figures

Functional Central Limit Theorem for the simultaneous subgraph count of dynamic Erdős-Rényi random graphs

Authors: Rajat Subhra Hazra, Nikolai Kriukov, Michel Mandjes

Abstract: In this paper we consider a dynamic Erdős-Rényi random graph with independent identically distributed edge processes. Our aim is to describe the joint evolution of the entries of a subgraph count vector. The main result of this paper is a functional central limit theorem: we establish, under an appropriate centering and scaling, the joint functional convergence of the vector of subgraph counts to… ▽ More In this paper we consider a dynamic Erdős-Rényi random graph with independent identically distributed edge processes. Our aim is to describe the joint evolution of the entries of a subgraph count vector. The main result of this paper is a functional central limit theorem: we establish, under an appropriate centering and scaling, the joint functional convergence of the vector of subgraph counts to a specific multidimensional Gaussian process. The result holds under mild assumptions on the edge processes, most notably a Lipschitz-type condition. △ Less

Submitted 26 November, 2025; v1 submitted 3 February, 2025; originally announced February 2025.

Comments: 38 pages. To appear in Electronic Journal of Probability

Evaluating Efficiency and Engagement in Scripted and LLM-Enhanced Human-Robot Interactions

Authors: Tim Schreiter, Jens V. Rüppel, Rishi Hazra, Andrey Rudenko, Martin Magnusson, Achim J. Lilienthal

Abstract: To achieve natural and intuitive interaction with people, HRI frameworks combine a wide array of methods for human perception, intention communication, human-aware navigation and collaborative action. In practice, when encountering unpredictable behavior of people or unexpected states of the environment, these frameworks may lack the ability to dynamically recognize such states, adapt and recover… ▽ More To achieve natural and intuitive interaction with people, HRI frameworks combine a wide array of methods for human perception, intention communication, human-aware navigation and collaborative action. In practice, when encountering unpredictable behavior of people or unexpected states of the environment, these frameworks may lack the ability to dynamically recognize such states, adapt and recover to resume the interaction. Large Language Models (LLMs), owing to their advanced reasoning capabilities and context retention, present a promising solution for enhancing robot adaptability. This potential, however, may not directly translate to improved interaction metrics. This paper considers a representative interaction with an industrial robot involving approach, instruction, and object manipulation, implemented in two conditions: (1) fully scripted and (2) including LLM-enhanced responses. We use gaze tracking and questionnaires to measure the participants' task efficiency, engagement, and robot perception. The results indicate higher subjective ratings for the LLM condition, but objective metrics show that the scripted condition performs comparably, particularly in efficiency and focus during simple tasks. We also note that the scripted condition may have an edge over LLM-enhanced responses in terms of response latency and energy consumption, especially for trivial and repetitive interactions. △ Less

Submitted 21 January, 2025; originally announced January 2025.

Comments: Accepted as a Late-Breaking Report to the 2025, 20th ACM/IEEE International Conference on Human-Robot Interaction (HRI)

The voter model on random regular graphs with random rewiring

Authors: Luca Avena, Rangel Baldasso, Rajat Subhra Hazra, Frank den Hollander, Matteo Quattropani

Abstract: We consider the voter model with binary opinions on a random regular graph with $n$ vertices of degree $d \geq 3$, subject to a rewiring dynamics in which pairs of edges are rewired, i.e., broken into four half-edges and subsequently reconnected at random. A parameter $ν\in (0,\infty)$ regulates the frequency at which the rewirings take place, in such a way that any given edge is rewired exponenti… ▽ More We consider the voter model with binary opinions on a random regular graph with $n$ vertices of degree $d \geq 3$, subject to a rewiring dynamics in which pairs of edges are rewired, i.e., broken into four half-edges and subsequently reconnected at random. A parameter $ν\in (0,\infty)$ regulates the frequency at which the rewirings take place, in such a way that any given edge is rewired exponentially at a rate $ν$ in the limit as $n\to\infty$. We show that, under the joint law of the random rewiring dynamics and the random opinion dynamics, the fraction of vertices with either one of the two opinions converges on time scale $n$ to the Fisher-Wright diffusion with an explicit diffusion constant $\vartheta_{d,ν}$ in the limit as $n\to\infty$. In particular, we identify $\vartheta_{d,ν}$ in terms of a continued-fraction expansion and analyse its dependence on $d$ and $ν$. A key role in our analysis is played by the set of discordant edges, which constitutes the boundary between the sets of vertices carrying the two opinions. △ Less

Submitted 15 January, 2025; originally announced January 2025.

Comments: 53 pages

Turning Logic Against Itself : Probing Model Defenses Through Contrastive Questions

Authors: Rachneet Sachdeva, Rima Hazra, Iryna Gurevych

Abstract: Large language models, despite extensive alignment with human values and ethical principles, remain vulnerable to sophisticated jailbreak attacks that exploit their reasoning abilities. Existing safety measures often detect overt malicious intent but fail to address subtle, reasoning-driven vulnerabilities. In this work, we introduce POATE (Polar Opposite query generation, Adversarial Template con… ▽ More Large language models, despite extensive alignment with human values and ethical principles, remain vulnerable to sophisticated jailbreak attacks that exploit their reasoning abilities. Existing safety measures often detect overt malicious intent but fail to address subtle, reasoning-driven vulnerabilities. In this work, we introduce POATE (Polar Opposite query generation, Adversarial Template construction, and Elaboration), a novel jailbreak technique that harnesses contrastive reasoning to provoke unethical responses. POATE crafts semantically opposing intents and integrates them with adversarial templates, steering models toward harmful outputs with remarkable subtlety. We conduct extensive evaluation across six diverse language model families of varying parameter sizes to demonstrate the robustness of the attack, achieving significantly higher attack success rates (~44%) compared to existing methods. To counter this, we propose Intent-Aware CoT and Reverse Thinking CoT, which decompose queries to detect malicious intent and reason in reverse to evaluate and reject harmful responses. These methods enhance reasoning robustness and strengthen the model's defense against adversarial exploits. △ Less

Submitted 30 September, 2025; v1 submitted 3 January, 2025; originally announced January 2025.

Comments: Accepted at EMNLP 2025 (Main)

Navigating the Cultural Kaleidoscope: A Hitchhiker's Guide to Sensitivity in Large Language Models

Authors: Somnath Banerjee, Sayan Layek, Hari Shrawgi, Rajarshi Mandal, Avik Halder, Shanu Kumar, Sagnik Basu, Parag Agrawal, Rima Hazra, Animesh Mukherjee

Abstract: As LLMs are increasingly deployed in global applications, the importance of cultural sensitivity becomes paramount, ensuring that users from diverse backgrounds feel respected and understood. Cultural harm can arise when these models fail to align with specific cultural norms, resulting in misrepresentations or violations of cultural values. This work addresses the challenges of ensuring cultural… ▽ More As LLMs are increasingly deployed in global applications, the importance of cultural sensitivity becomes paramount, ensuring that users from diverse backgrounds feel respected and understood. Cultural harm can arise when these models fail to align with specific cultural norms, resulting in misrepresentations or violations of cultural values. This work addresses the challenges of ensuring cultural sensitivity in LLMs, especially in small-parameter models that often lack the extensive training data needed to capture global cultural nuances. We present two key contributions: (1) A cultural harm test dataset, created to assess model outputs across different cultural contexts through scenarios that expose potential cultural insensitivities, and (2) A culturally aligned preference dataset, aimed at restoring cultural sensitivity through fine-tuning based on feedback from diverse annotators. These datasets facilitate the evaluation and enhancement of LLMs, ensuring their ethical and safe deployment across different cultural landscapes. Our results show that integrating culturally aligned feedback leads to a marked improvement in model behavior, significantly reducing the likelihood of generating culturally insensitive or harmful content. Ultimately, this work paves the way for more inclusive and respectful AI systems, fostering a future where LLMs can safely and ethically navigate the complexities of diverse cultural landscapes. △ Less

Submitted 24 January, 2025; v1 submitted 15 October, 2024; originally announced October 2024.

Comments: Accepted at NAACL 2025 (Main track). [Project Page](https://neuralsentinel.github.io/KaleidoCulture/)

Largest eigenvalue of positive mean Gaussian matrices

Authors: Arijit Chakrabarty, Rajat Subhra Hazra, Moumanti Podder

Abstract: This short note studies the fluctuations of the largest eigenvalue of symmetric random matrices with correlated Gaussian entries having positive mean. Under the assumption that the covariance kernel is absolutely summable, it is proved that the largest eigenvalue, after centering, converges in distribution to normal with an explicitly defined mean and variance. This result generalizes known findin… ▽ More This short note studies the fluctuations of the largest eigenvalue of symmetric random matrices with correlated Gaussian entries having positive mean. Under the assumption that the covariance kernel is absolutely summable, it is proved that the largest eigenvalue, after centering, converges in distribution to normal with an explicitly defined mean and variance. This result generalizes known findings for Wigner matrices with independent entries. △ Less

Submitted 17 October, 2024; v1 submitted 9 September, 2024; originally announced September 2024.

Comments: 11 pages

Bidirectional Intent Communication: A Role for Large Foundation Models

Authors: Tim Schreiter, Rishi Hazra, Jens Rüppel, Andrey Rudenko

Abstract: Integrating multimodal foundation models has significantly enhanced autonomous agents' language comprehension, perception, and planning capabilities. However, while existing works adopt a \emph{task-centric} approach with minimal human interaction, applying these models to developing assistive \emph{user-centric} robots that can interact and cooperate with humans remains underexplored. This paper… ▽ More Integrating multimodal foundation models has significantly enhanced autonomous agents' language comprehension, perception, and planning capabilities. However, while existing works adopt a \emph{task-centric} approach with minimal human interaction, applying these models to developing assistive \emph{user-centric} robots that can interact and cooperate with humans remains underexplored. This paper introduces ``Bident'', a framework designed to integrate robots seamlessly into shared spaces with humans. Bident enhances the interactive experience by incorporating multimodal inputs like speech and user gaze dynamics. Furthermore, Bident supports verbal utterances and physical actions like gestures, making it versatile for bidirectional human-robot interactions. Potential applications include personalized education, where robots can adapt to individual learning styles and paces, and healthcare, where robots can offer personalized support, companionship, and everyday assistance in the home and workplace environments. △ Less

Submitted 20 August, 2024; originally announced August 2024.

Comments: 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (RO-MAN), Workshop: Large Language Models in the RoMan Age

Topology of total cut complexes and cut complexes of grid graphs

Authors: Himanshu Chandrakar, Nisith Ranjan Hazra, Debotosh Rout, Anurag Singh

Abstract: Inspired by the work of Fr{ö}berg (1990) and Eagon and Reiner (1998), Bayer et al. recently introduced two new graph complexes: total cut complexes and cut complexes. In this article, we investigate these complexes specifically for (rectangular) grid graphs, focusing on $2 \times n$ and $3 \times n$ cases. We extend and refine the work of Bayer et al., proving and strengthening several of their co… ▽ More Inspired by the work of Fr{ö}berg (1990) and Eagon and Reiner (1998), Bayer et al. recently introduced two new graph complexes: total cut complexes and cut complexes. In this article, we investigate these complexes specifically for (rectangular) grid graphs, focusing on $2 \times n$ and $3 \times n$ cases. We extend and refine the work of Bayer et al., proving and strengthening several of their conjectures, thereby enhancing the understanding of these graph complexes' topological and combinatorial properties. △ Less

Submitted 12 January, 2026; v1 submitted 14 August, 2024; originally announced August 2024.

Comments: To appear in the SIAM Journal on Discrete Mathematics

Can Large Language Models Reason? A Characterization via 3-SAT

Authors: Rishi Hazra, Gabriele Venturato, Pedro Zuidberg Dos Martires, Luc De Raedt

Abstract: Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. However, recent works have shown that LLMs often bypass true reasoning using shortcuts, sparking skepticism. To study the reasoning capabilities in a principled fashion, we adopt a computational theory perspective and propose an experimental protocol centered on 3-SAT -- the prototypical NP-complete… ▽ More Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. However, recent works have shown that LLMs often bypass true reasoning using shortcuts, sparking skepticism. To study the reasoning capabilities in a principled fashion, we adopt a computational theory perspective and propose an experimental protocol centered on 3-SAT -- the prototypical NP-complete problem lying at the core of logical reasoning and constraint satisfaction tasks. Specifically, we examine the phase transitions in random 3-SAT and characterize the reasoning abilities of LLMs by varying the inherent hardness of the problem instances. Our experimental evidence shows that LLMs are incapable of performing true reasoning, as required for solving 3-SAT problems. Moreover, we observe significant performance variation based on the inherent hardness of the problems -- performing poorly on harder instances and vice versa. Importantly, we show that integrating external reasoners can considerably enhance LLM performance. By following a principled experimental protocol, our study draws concrete conclusions and moves beyond the anecdotal evidence often found in LLM reasoning research. △ Less

Submitted 22 October, 2024; v1 submitted 13 August, 2024; originally announced August 2024.

Functional Central Limit Theorem for the principal eigenvalue of dynamic Erdős-Rényi random graphs

Authors: Rajat Subhra Hazra, Nikolai Kriukov, Michel Mandjes

Abstract: In this paper we consider a dynamic version of the Erdős-Rényi random graph, in which edges independently appear and disappear in time, with the on- and off times being exponentially distributed. The focus lies on the evolution of the principle eigenvalue of the adjacency matrix in the regime that the number of vertices grows large. The main result is a functional central limit theorem, which disp… ▽ More In this paper we consider a dynamic version of the Erdős-Rényi random graph, in which edges independently appear and disappear in time, with the on- and off times being exponentially distributed. The focus lies on the evolution of the principle eigenvalue of the adjacency matrix in the regime that the number of vertices grows large. The main result is a functional central limit theorem, which displays that the principal eigenvalue essentially inherits the characteristics of the dynamics of the individual edges. △ Less

Submitted 2 July, 2024; originally announced July 2024.

MSC Class: 05C80; 15B52; 60B20

SafeInfer: Context Adaptive Decoding Time Safety Alignment for Large Language Models

Authors: Somnath Banerjee, Sayan Layek, Soham Tripathy, Shanu Kumar, Animesh Mukherjee, Rima Hazra

Abstract: Safety-aligned language models often exhibit fragile and imbalanced safety mechanisms, increasing the likelihood of generating unsafe content. In addition, incorporating new knowledge through editing techniques to language models can further compromise safety. To address these issues, we propose SafeInfer, a context-adaptive, decoding-time safety alignment strategy for generating safe responses to… ▽ More Safety-aligned language models often exhibit fragile and imbalanced safety mechanisms, increasing the likelihood of generating unsafe content. In addition, incorporating new knowledge through editing techniques to language models can further compromise safety. To address these issues, we propose SafeInfer, a context-adaptive, decoding-time safety alignment strategy for generating safe responses to user queries. SafeInfer comprises two phases: the safety amplification phase, which employs safe demonstration examples to adjust the model's hidden states and increase the likelihood of safer outputs, and the safety-guided decoding phase, which influences token selection based on safety-optimized distributions, ensuring the generated content complies with ethical guidelines. Further, we present HarmEval, a novel benchmark for extensive safety evaluations, designed to address potential misuse scenarios in accordance with the policies of leading AI tech giants. △ Less

Submitted 14 December, 2024; v1 submitted 18 June, 2024; originally announced June 2024.

Comments: Accepted at AAAI 2025 (AI Alignment Track). Also selected for Microsoft Academic Partnership Grant (MAPG) 2024

Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations

Authors: Rima Hazra, Sayan Layek, Somnath Banerjee, Soujanya Poria

Abstract: Ensuring the safe alignment of large language models (LLMs) with human values is critical as they become integral to applications like translation and question answering. Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to generating harmful content. We propose Safety Arithmetic, a training-free framework enhancing LLM safety across d… ▽ More Ensuring the safe alignment of large language models (LLMs) with human values is critical as they become integral to applications like translation and question answering. Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to generating harmful content. We propose Safety Arithmetic, a training-free framework enhancing LLM safety across different scenarios: Base models, Supervised fine-tuned models (SFT), and Edited models. Safety Arithmetic involves Harm Direction Removal to avoid harmful content and Safety Alignment to promote safe responses. Additionally, we present NoIntentEdit, a dataset highlighting edit instances that could compromise model safety if used unintentionally. Our experiments show that Safety Arithmetic significantly improves safety measures, reduces over-safety, and maintains model utility, outperforming existing methods in ensuring safe content generation. △ Less

Submitted 28 October, 2024; v1 submitted 17 June, 2024; originally announced June 2024.

Comments: EMNLP 2024 Main. Codes are available at: https://github.com/declare-lab/safety-arithmetic

Breaking Boundaries: Investigating the Effects of Model Editing on Cross-linguistic Performance

Authors: Somnath Banerjee, Avik Halder, Rajarshi Mandal, Sayan Layek, Ian Soboroff, Rima Hazra, Animesh Mukherjee

Abstract: The integration of pretrained language models (PLMs) like BERT and GPT has revolutionized NLP, particularly for English, but it has also created linguistic imbalances. This paper strategically identifies the need for linguistic equity by examining several knowledge editing techniques in multilingual contexts. We evaluate the performance of models such as Mistral, TowerInstruct, OpenHathi, Tamil-Ll… ▽ More The integration of pretrained language models (PLMs) like BERT and GPT has revolutionized NLP, particularly for English, but it has also created linguistic imbalances. This paper strategically identifies the need for linguistic equity by examining several knowledge editing techniques in multilingual contexts. We evaluate the performance of models such as Mistral, TowerInstruct, OpenHathi, Tamil-Llama, and Kan-Llama across languages including English, German, French, Italian, Spanish, Hindi, Tamil, and Kannada. Our research identifies significant discrepancies in normal and merged models concerning cross-lingual consistency. We employ strategies like 'each language for itself' (ELFI) and 'each language for others' (ELFO) to stress-test these models. Our findings demonstrate the potential for LLMs to overcome linguistic barriers, laying the groundwork for future research in achieving linguistic inclusivity in AI technologies. △ Less

Submitted 18 March, 2025; v1 submitted 16 June, 2024; originally announced June 2024.

Comments: Accepted at NAACL 2025 (Industry track)

REvolve: Reward Evolution with Large Language Models using Human Feedback

Authors: Rishi Hazra, Alkis Sygkounas, Andreas Persson, Amy Loutfi, Pedro Zuidberg Dos Martires

Abstract: Designing effective reward functions is crucial to training reinforcement learning (RL) algorithms. However, this design is non-trivial, even for domain experts, due to the subjective nature of certain tasks that are hard to quantify explicitly. In recent works, large language models (LLMs) have been used for reward generation from natural language task descriptions, leveraging their extensive ins… ▽ More Designing effective reward functions is crucial to training reinforcement learning (RL) algorithms. However, this design is non-trivial, even for domain experts, due to the subjective nature of certain tasks that are hard to quantify explicitly. In recent works, large language models (LLMs) have been used for reward generation from natural language task descriptions, leveraging their extensive instruction tuning and commonsense understanding of human behavior. In this work, we hypothesize that LLMs, guided by human feedback, can be used to formulate reward functions that reflect human implicit knowledge. We study this in three challenging settings -- autonomous driving, humanoid locomotion, and dexterous manipulation -- wherein notions of ``good" behavior are tacit and hard to quantify. To this end, we introduce REvolve, a truly evolutionary framework that uses LLMs for reward design in RL. REvolve generates and refines reward functions by utilizing human feedback to guide the evolution process, effectively translating implicit human knowledge into explicit reward functions for training (deep) RL agents. Experimentally, we demonstrate that agents trained on REvolve-designed rewards outperform other state-of-the-art baselines. △ Less

Submitted 23 May, 2025; v1 submitted 3 June, 2024; originally announced June 2024.

Comments: Published in ICLR 2025. Project page: https://rishihazra.github.io/REvolve

DistALANER: Distantly Supervised Active Learning Augmented Named Entity Recognition in the Open Source Software Ecosystem

Authors: Somnath Banerjee, Avik Dutta, Aaditya Agrawal, Rima Hazra, Animesh Mukherjee

Abstract: With the AI revolution in place, the trend for building automated systems to support professionals in different domains such as the open source software systems, healthcare systems, banking systems, transportation systems and many others have become increasingly prominent. A crucial requirement in the automation of support tools for such systems is the early identification of named entities, which… ▽ More With the AI revolution in place, the trend for building automated systems to support professionals in different domains such as the open source software systems, healthcare systems, banking systems, transportation systems and many others have become increasingly prominent. A crucial requirement in the automation of support tools for such systems is the early identification of named entities, which serves as a foundation for developing specialized functionalities. However, due to the specific nature of each domain, different technical terminologies and specialized languages, expert annotation of available data becomes expensive and challenging. In light of these challenges, this paper proposes a novel named entity recognition (NER) technique specifically tailored for the open-source software systems. Our approach aims to address the scarcity of annotated software data by employing a comprehensive two-step distantly supervised annotation process. This process strategically leverages language heuristics, unique lookup tables, external knowledge sources, and an active learning approach. By harnessing these powerful techniques, we not only enhance model performance but also effectively mitigate the limitations associated with cost and the scarcity of expert annotators. It is noteworthy that our model significantly outperforms the state-of-the-art LLMs by a substantial margin. We also show the effectiveness of NER in the downstream task of relation extraction. △ Less

Submitted 20 June, 2024; v1 submitted 25 February, 2024; originally announced February 2024.

Comments: Accepted at ECML-PKDD 2024 (Long Paper)

How (un)ethical are instruction-centric responses of LLMs? Unveiling the vulnerabilities of safety guardrails to harmful queries

Authors: Somnath Banerjee, Sayan Layek, Rima Hazra, Animesh Mukherjee

Abstract: In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including 'jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking the… ▽ More In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including 'jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking them to generate responses that are instruction-centric such as a pseudocode, a program or a software snippet as opposed to vanilla text. To investigate this question, we introduce TechHazardQA, a dataset containing complex queries which should be answered in both text and instruction-centric formats (e.g., pseudocodes), aimed at identifying triggers for unethical responses. We query a series of LLMs -- Llama-2-13b, Llama-2-7b, Mistral-V2 and Mistral 8X7B -- and ask them to generate both text and instruction-centric responses. For evaluation we report the harmfulness score metric as well as judgements from GPT-4 and humans. Overall, we observe that asking LLMs to produce instruction-centric responses enhances the unethical response generation by ~2-38% across the models. As an additional objective, we investigate the impact of model editing using the ROME technique, which further increases the propensity for generating undesirable content. In particular, asking edited LLMs to generate instruction-centric responses further increases the unethical response generation by ~3-16% across the different models. △ Less

Submitted 16 November, 2024; v1 submitted 23 February, 2024; originally announced February 2024.

Comments: Accepted at AAAI Conference on Web and Social Media (ICWSM) 2025. [Dataset](https://huggingface.co/datasets/SoftMINER-Group/TechHazardQA)

Context Matters: Pushing the Boundaries of Open-Ended Answer Generation with Graph-Structured Knowledge Context

Authors: Somnath Banerjee, Amruit Sahoo, Sayan Layek, Avik Dutta, Rima Hazra, Animesh Mukherjee

Abstract: In the continuously advancing AI landscape, crafting context-rich and meaningful responses via Large Language Models (LLMs) is essential. Researchers are becoming more aware of the challenges that LLMs with fewer parameters encounter when trying to provide suitable answers to open-ended questions. To address these hurdles, the integration of cutting-edge strategies, augmentation of rich external d… ▽ More In the continuously advancing AI landscape, crafting context-rich and meaningful responses via Large Language Models (LLMs) is essential. Researchers are becoming more aware of the challenges that LLMs with fewer parameters encounter when trying to provide suitable answers to open-ended questions. To address these hurdles, the integration of cutting-edge strategies, augmentation of rich external domain knowledge to LLMs, offers significant improvements. This paper introduces a novel framework that combines graph-driven context retrieval in conjunction to knowledge graphs based enhancement, honing the proficiency of LLMs, especially in domain specific community question answering platforms like AskUbuntu, Unix, and ServerFault. We conduct experiments on various LLMs with different parameter sizes to evaluate their ability to ground knowledge and determine factual accuracy in answers to open-ended questions. Our methodology GraphContextGen consistently outperforms dominant text-based retrieval systems, demonstrating its robustness and adaptability to a larger number of use cases. This advancement highlights the importance of pairing context rich data retrieval with LLMs, offering a renewed approach to knowledge sourcing and generation in AI systems. We also show that, due to rich contextual data retrieval, the crucial entities, along with the generated answer, remain factually coherent with the gold answer. △ Less

Submitted 15 October, 2024; v1 submitted 23 January, 2024; originally announced January 2024.

Comments: Accepted at EMNLP 2024