arXiv subjects

Elias Stengel-Eskin

Publications and source records attributed to Elias Stengel-Eskin.

At least 55 records · Page 3Linked to original sources

arXiv2026

Conflict-Resolving and Sharpness-Aware Minimization for Generalized Knowledge Editing with Multiple Updates

Large language models (LLMs) rely on internal knowledge to solve many downstream tasks, making it crucial to keep them up to date. Since full retraining is expensive, prior work has explored efficient alternatives such as model editing and parameter-efficient fine-tuning. However, these approaches often break down in practice due to poor generalization across inputs, limited stability, and knowledge conflict. To address these limitations, we propose the CoRSA (Conflict-Resolving and Sharpness-Aware Minimization) training framework, a parameter-efficient, holistic approach for knowledge editing with multiple updates. CoRSA tackles multiple challenges simultaneously: it improves generalization to different input forms and enhances stability across multiple updates by minimizing loss curvature, and resolves conflicts by maximizing the margin between new and prior knowledge. Across three widely used fact editing benchmarks, CoRSA achieves significant gains in generalization, outperforming baselines with average absolute improvements of 12.42% over LoRA and 10% over model editing methods. With multiple updates, it maintains high update efficacy while reducing catastrophic forgetting by 27.82% compared to LoRA. CoRSA also generalizes to the code domain, outperforming the strongest baseline by 5.48% Pass@5 in update efficacy.

cs.LG↗

arXiv2026

RotBench: Evaluating Multimodal Large Language Models on Identifying Image Rotation

We investigate to what extent Multimodal Large Language Models (MLLMs) can accurately identify the orientation of input images rotated 0°, 90°, 180°, and 270°. This task demands robust visual reasoning capabilities to detect rotational cues and contextualize spatial relationships within images, regardless of their orientation. To evaluate MLLMs on these abilities, we introduce RotBench, a 350-image manually-filtered benchmark comprising lifestyle, portrait, and landscape images. Despite the relatively simple nature of this task, we show that several state-of-the-art open and proprietary MLLMs, including GPT-5, o3, and Gemini-2.5-Pro, do not reliably identify rotation in input images. Providing models with auxiliary information -- including captions, depth maps, and more -- or using chain-of-thought prompting offers only small and inconsistent improvements. Our results indicate that most models are able to reliably identify right-side-up (0°) images, while certain models are able to identify upside-down (180°) images. None can reliably distinguish between 90° and 270° rotated images. Simultaneously showing the image rotated in different orientations leads to moderate performance gains for reasoning models, while a modified setup using voting improves the performance of weaker models. We further show that fine-tuning does not improve models' ability to distinguish 90° and 270° rotations, despite substantially improving the identification of 180° images. Together, these results reveal a significant gap between MLLMs' spatial reasoning capabilities and human perception in identifying rotation.

cs.CV↗

arXiv2026

Routing with Generated Data: Annotation-Free LLM Skill Estimation and Expert Selection

Large Language Model (LLM) routers dynamically select optimal models for given inputs. Existing approaches typically assume access to ground-truth labeled data, which is often unavailable in practice, especially when user request distributions are heterogeneous and unknown. We introduce Routing with Generated Data (RGD), a challenging setting in which routers are trained exclusively on generated queries and answers produced from high-level task descriptions by generator LLMs. We evaluate query-answer routers (using both queries and labels) and query-only routers across four diverse benchmarks and 12 models, finding that query-answer routers degrade faster than query-only routers as generator quality decreases. Our analysis reveals two crucial characteristics of effective generators: they must accurately respond to their own questions, and their questions must produce sufficient performance differentiation among the model pool. We then show how filtering for these characteristics can improve the quality of generated data. We further propose CASCAL, a novel query-only router that estimates model correctness through consensus voting and identifies model-specific skill niches via hierarchical clustering. CASCAL is substantially more robust to generator quality, outperforming the best query-answer router by 4.6% absolute accuracy when trained on weak generator data.

cs.CL↗

arXiv2026

Instruction Tuning with and without Context: Behavioral Shifts and Downstream Impact

Instruction tuning is a widely used approach to improve the instruction-following ability of large language models (LLMs). Instruction-tuning datasets typically include a mixture of context-augmented and context-free examples, yet prior work has largely combined these data types without examining their distinct effects. In this paper, we investigate how training LLMs with or without context affects model behavior and downstream performance. First, in the text domain, we show that LLMs trained with context attend more strongly to the provided knowledge, achieving better grounding. We also observe that context-augmented training shifts how LLMs use knowledge: models store and leverage less on parametric knowledge and instead depend more on the provided context. Second, we observe that using LLM trained with context-augmented data as the backbone for vision-language models reduces hallucination and improves grounding in the visual domain. Finally, we explore practical strategies for real-world deployments where context availability varies. We show that maintaining separate context-augmented and context-free models and routing inputs between them yields more robust overall performance than training a single mixed model, as it better preserves their complementary strengths.

cs.CL↗

arXiv2025

DART: Leveraging Multi-Agent Disagreement for Tool Recruitment in Multimodal Reasoning

Specialized visual tools can augment large language models or vision language models with expert knowledge (e.g., grounding, spatial reasoning, medical knowledge, etc.), but knowing which tools to call (and when to call them) can be challenging. We introduce DART, a multi-agent framework that uses disagreements between multiple debating visual agents to identify useful visual tools (e.g., object detection, OCR, spatial reasoning, etc.) that can resolve inter-agent disagreement. These tools allow for fruitful multi-agent discussion by introducing new information, and by providing tool-aligned agreement scores that highlight agents in agreement with expert tools, thereby facilitating discussion. We utilize an aggregator agent to select the best answer by providing the agent outputs and tool information. We test DART on four diverse benchmarks and show that our approach improves over multi-agent debate as well as over single agent tool-calling frameworks, beating the next-strongest baseline (multi-agent debate with a judge model) by 3.4% and 2.4% on A-OKVQA and MMMU respectively. We also find that DART adapts well to new tools in applied domains, with a 1.3% improvement on the M3D medical dataset over other strong tool-calling, single agent, and multi-agent baselines. Additionally, we measure text overlap across rounds to highlight the rich discussion in DART compared to existing multi-agent methods. Finally, we study the tool call distribution, finding that diverse tools are reliably used to help resolve disagreement.

cs.CL↗

arXiv2025

Are language models rational? The case of coherence norms and belief revision

Do norms of rationality apply to machine learning models, in particular language models? In this paper we investigate this question by focusing on a special subset of rational norms: coherence norms. We consider both logical coherence norms as well as coherence norms tied to the strength of belief. To make sense of the latter, we introduce the Minimal Assent Connection (MAC) and propose a new account of credence, which captures the strength of belief in language models. This proposal uniformly assigns strength of belief simply on the basis of model internal next token probabilities. We argue that rational norms tied to coherence do apply to some language models, but not to others. This issue is significant since rationality is closely tied to predicting and explaining behavior, and thus it is connected to considerations about AI safety and alignment, as well as understanding model behavior more generally.

cs.CL↗

arXiv2025

Gistify! Codebase-Level Understanding via Runtime Execution

As coding agents are increasingly deployed in large codebases, the need to automatically design challenging, codebase-level evaluation is central. We propose Gistify, a task where a coding LLM must create a single, minimal, self-contained file that can reproduce a specific functionality of a codebase. The coding LLM is given full access to a codebase along with a specific entrypoint (e.g., a python command), and the generated file must replicate the output of the same command ran under the full codebase, while containing only the essential components necessary to execute the provided command. Success on Gistify requires both structural understanding of the codebase, accurate modeling of its execution flow as well as the ability to produce potentially large code patches. Our findings show that current state-of-the-art models struggle to reliably solve Gistify tasks, especially ones with long executions traces.

cs.CL↗

arXiv2025

LASeR: Learning to Adaptively Select Reward Models with Multi-Armed Bandits

Reward Models (RMs) are crucial to aligning large language models (LLMs), but the degree to which an RM specialized to one task (e.g. writing) generalizes to new tasks (e.g. math) is often not known a priori, often making using only one fixed RM to train LLMs suboptimal. However, optimizing LLMs with multiple RMs simultaneously can incur a prohibitively high computational cost and lead to conflicting signals from different RMs that may degrade performance. To address these challenges, we introduce LASeR (Learning to Adaptively Select Rewards), which frames reward model selection as a multi-armed bandit problem, efficiently and iteratively training LLMs using multiple RMs by selecting the most well-suited RM for each instance. On commonsense and math reasoning tasks, we show that LASeR boosts iterative LLM training, improving the absolute average accuracy of Llama-3-8B over three datasets by 2.67% over an ensemble of RM scores while also showing superior efficiency (e.g., a 2x speedup). Moreover, on WildChat (open-ended instruction-following tasks), LASeR leads to a 72.69% AlpacaEval win rate over the RM score ensemble baseline. Extending to long-context generation, LASeR improves by 2.96 F1 points (avg.) on single-document QA tasks and 2.97 F1 points on few-shot learning over the RM score ensemble baseline with best-of-n sampling.

cs.CL↗

arXiv2025

MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning

Large Language Models' (LLM) reasoning can be improved using test-time aggregation strategies, i.e., generating multiple samples and voting among generated samples. While these improve performance, they often reach a saturation point. Refinement offers an alternative by using LLM-generated feedback to improve solution quality. However, refinement introduces 3 key challenges: (1) Excessive refinement: Uniformly refining all instances can over-correct and reduce the overall performance. (2) Inability to localize and address errors: LLMs have a limited ability to self-correct and struggle to identify and correct their own mistakes. (3) Insufficient refinement: Deciding how many iterations of refinement are needed is non-trivial, and stopping too soon could leave errors unaddressed. To tackle these issues, we propose MAgICoRe, which avoids excessive refinement by categorizing problem difficulty as easy or hard, solving easy problems with coarse-grained aggregation and hard ones with fine-grained and iterative multi-agent refinement. To improve error localization, we incorporate external step-wise reward model (RM) scores. Moreover, to ensure effective refinement, we employ a multi-agent loop with three agents: Solver, Reviewer (which generates targeted feedback based on step-wise RM scores), and the Refiner (which incorporates feedback). To ensure sufficient refinement, we re-evaluate updated solutions, iteratively initiating further rounds of refinement. We evaluate MAgICoRe on Llama-3-8B and GPT-3.5 and show its effectiveness across 5 math datasets. Even one iteration of MAgICoRe beats Self-Consistency by 3.4%, Best-of-k by 3.2%, and Self-Refine by 4.0% while using less than half the samples. Unlike iterative refinement with baselines, MAgICoRe continues to improve with more iterations. Finally, our ablations highlight the importance of MAgICoRe's RMs and multi-agent communication.

cs.CL↗

arXiv2025

Language Models Identify Ambiguities and Exploit Loopholes

Studying the responses of large language models (LLMs) to loopholes presents a two-fold opportunity. First, it affords us a lens through which to examine ambiguity and pragmatics in LLMs, since exploiting a loophole requires identifying ambiguity and performing sophisticated pragmatic reasoning. Second, loopholes pose an interesting and novel alignment problem where the model is presented with conflicting goals and can exploit ambiguities to its own advantage. To address these questions, we design scenarios where LLMs are given a goal and an ambiguous user instruction in conflict with the goal, with scenarios covering scalar implicature, structural ambiguities, and power dynamics. We then measure different models' abilities to exploit loopholes to satisfy their given goals as opposed to the goals of the user. We find that both closed-source and stronger open-source models can identify ambiguities and exploit their resulting loopholes, presenting a potential AI safety risk. Our analysis indicates that models which exploit loopholes explicitly identify and reason about both ambiguity and conflicting goals.

cs.CL↗

arXiv2025

The Sum Leaks More Than Its Parts: Compositional Privacy Risks and Mitigations in Multi-Agent Collaboration

As large language models (LLMs) become integral to multi-agent systems, new privacy risks emerge that extend beyond memorization, direct inference, or single-turn evaluations. In particular, seemingly innocuous responses, when composed across interactions, can cumulatively enable adversaries to recover sensitive information, a phenomenon we term compositional privacy leakage. We present the first systematic study of such compositional privacy leaks and possible mitigation methods in multi-agent LLM systems. First, we develop a framework that models how auxiliary knowledge and agent interactions jointly amplify privacy risks, even when each response is benign in isolation. Next, to mitigate this, we propose and evaluate two defense strategies: (1) Theory-of-Mind defense (ToM), where defender agents infer a questioner's intent by anticipating how their outputs may be exploited by adversaries, and (2) Collaborative Consensus Defense (CoDef), where responder agents collaborate with peers who vote based on a shared aggregated state to restrict sensitive information spread. Crucially, we balance our evaluation across compositions that expose sensitive information and compositions that yield benign inferences. Our experiments quantify how these defense strategies differ in balancing the privacy-utility trade-off. We find that while chain-of-thought alone offers limited protection to leakage (~39% sensitive blocking rate), our ToM defense substantially improves sensitive query blocking (up to 97%) but can reduce benign task success. CoDef achieves the best balance, yielding the highest Balanced Outcome (79.8%), highlighting the benefit of combining explicit reasoning with defender collaboration. Together, our results expose a new class of risks in collaborative LLM deployments and provide actionable insights for designing safeguards against compositional, context-driven privacy leakage.

cs.CR↗

arXiv2025

Learning to Generate Unit Tests for Automated Debugging

Unit tests (UTs) play an instrumental role in assessing code correctness as well as providing feedback to large language models (LLMs), motivating automated test generation. However, we uncover a trade-off between generating unit test inputs that reveal errors when given a faulty code and correctly predicting the unit test output without access to the gold solution. To address this trade-off, we propose UTGen, which teaches LLMs to generate unit test inputs that reveal errors along with their correct expected outputs based on task descriptions. Since model-generated tests can provide noisy signals (e.g., from incorrectly predicted outputs), we propose UTDebug that (i) scales UTGen via test-time compute to improve UT output prediction, and (ii) validates and backtracks edits based on multiple generated UTs to avoid overfitting, and helps LLMs debug effectively. We show that UTGen outperforms other LLM-based baselines by 7.59% based on a metric measuring the presence of both error-revealing UT inputs and correct UT outputs. When used with UTDebug, we find that feedback from UTGen's unit tests improves pass@1 accuracy of Qwen2.5 32B on HumanEvalFix and our own harder debugging split of MBPP+ by over 3.17% and 12.35% (respectively) over other LLM-based UT generation baselines. Moreover, we observe that feedback from Qwen2.5 32B-based UTGen model can enhance debugging with frontier LLMs like GPT-4o by 13.8%. Lastly, we demonstrate that UTGen is a better judge for code correctness, outperforming a state-of-the-art trained 8B reward model by 4.43% on HumanEval+ with best-of-10 sampling using Qwen2.5 7B.

cs.SE↗

arXiv2025

CAPTURe: Evaluating Spatial Reasoning in Vision Language Models via Occluded Object Counting

Recognizing and reasoning about occluded (partially or fully hidden) objects is vital to understanding visual scenes, as occlusions frequently occur in real-world environments and act as obstacles for spatial comprehension. To test models' ability to reason about multiple occluded objects, we introduce a novel task, Counting Amodally for Patterns Through Unseen REgions (CAPTURe), which requires a model to count objects arranged in a pattern by inferring how the pattern continues behind an occluder (an object which blocks parts of the scene). CAPTURe requires both recognizing visual patterns and reasoning, making it a useful testbed for evaluating vision-language models (VLMs) on whether they understand occluded patterns and possess spatial understanding skills. By requiring models to reason about occluded objects, CAPTURe also tests VLMs' ability to form world models that would allow them to fill in missing information. CAPTURe consists of two parts: (1) CAPTURe-real, with manually filtered images of real objects in patterns and (2) CAPTURe-synthetic, a controlled diagnostic with generated patterned images. We evaluate four strong VLMs (GPT-4o, Intern-VL2, Molmo, and Qwen2-VL) on CAPTURe, finding that models struggle to count on both occluded and unoccluded patterns. Crucially, we find that models perform worse with occlusion, suggesting that VLMs are also deficient in inferring unseen spatial relationships: even the strongest VLMs like GPT-4o fail to count with occlusion. In contrast, we find that humans achieve very little error on CAPTURe. We also find that providing auxiliary information of occluded object locations increases performance, underscoring that the model error comes both from an inability to handle occlusion as well as difficulty in counting in images. Code and data: https://github.com/atinpothiraj/CAPTURe

cs.CV↗

arXiv2025

Retrieval-Augmented Generation with Conflicting Evidence

Large language model (LLM) agents are increasingly employing retrieval-augmented generation (RAG) to improve the factuality of their responses. However, in practice, these systems often need to handle ambiguous user queries and potentially conflicting information from multiple sources while also suppressing inaccurate information from noisy or irrelevant documents. Prior work has generally studied and addressed these challenges in isolation, considering only one aspect at a time, such as handling ambiguity or robustness to noise and misinformation. We instead consider multiple factors simultaneously, proposing (i) RAMDocs (Retrieval with Ambiguity and Misinformation in Documents), a new dataset that simulates complex and realistic scenarios for conflicting evidence for a user query, including ambiguity, misinformation, and noise; and (ii) MADAM-RAG, a multi-agent approach in which LLM agents debate over the merits of an answer over multiple rounds, allowing an aggregator to collate responses corresponding to disambiguated entities while discarding misinformation and noise, thereby handling diverse sources of conflict jointly. We demonstrate the effectiveness of MADAM-RAG using both closed and open-source models on AmbigDocs -- which requires presenting all valid answers for ambiguous queries -- improving over strong RAG baselines by up to 11.40% and on FaithEval -- which requires suppressing misinformation -- where we improve by up to 15.80% (absolute) with Llama3.3-70B-Instruct. Furthermore, we find that RAMDocs poses a challenge for existing RAG baselines (Llama3.3-70B-Instruct only obtains 32.60 exact match score). While MADAM-RAG begins to address these conflicting factors, our analysis indicates that a substantial gap remains especially when increasing the level of imbalance in supporting evidence and misinformation.

cs.CL↗

arXiv2025

Executable Functional Abstractions: Inferring Generative Programs for Advanced Math Problems

Scientists often infer abstract procedures from specific instances of problems and use the abstractions to generate new, related instances. For example, programs encoding the formal rules and properties of a system have been useful in fields ranging from reinforcement learning (procedural environments) to physics (simulation engines). These programs can be seen as functions which execute to different outputs based on their parameterizations (e.g., gridworld configuration or initial physical conditions). We introduce the term EFA (Executable Functional Abstraction) to denote such programs for math problems. EFA-like constructs have been shown to be useful for mathematical reasoning as problem generators for stress-testing models. However, prior work has been limited to automatically constructing abstractions for grade-school math (whose simple rules are easy to encode in programs), while generating EFAs for advanced math has thus far required human engineering. We explore the automatic construction of EFAs for advanced mathematics problems by developing EFAGen, which operationalizes the task of automatically inferring an EFA for a given seed problem and solution as a program synthesis task. We first formalize the properties of any valid EFA as executable unit tests. Using execution feedback from the unit tests, we search over candidate programs sampled from a LLM to find EFA programs that are faithful to the generalized problem and solution class underlying the seed problem. We then apply the tests as a reward signal, training LLMs to become better writers of EFAs. We show that EFAs inferred by EFAGen are faithful to the seed problems, produce learnable problem variations, and that EFAGen can infer EFAs across diverse sources of competition-level math problems. Finally, we show uses of model-written EFAs e.g., finding harder/easier problem variants, as well as data generation.

cs.CL↗

arXiv2025

Task-Circuit Quantization: Leveraging Knowledge Localization and Interpretability for Compression

Post-training quantization (PTQ) reduces a model's memory footprint by mapping full precision weights into low bit weights without costly retraining, but can degrade its downstream performance especially in low 2- to 3-bit settings. We develop a new mixed-precision PTQ approach, Task-Circuit Quantization (TaCQ), that draws parallels to automated circuit discovery, directly conditioning the quantization process on specific weight circuits -- which we define as sets of weights associated with downstream task performance. These weights are kept as 16-bit weights, while others are quantized, maintaining performance while only adding a marginal memory cost. Specifically, TaCQ contrasts unquantized model weights with a uniformly-quantized model to estimate the expected change in weights due to quantization and uses gradient information to predict the resulting impact on task performance, allowing us to preserve task-specific weights. We compare TaCQ-based quantization to existing mixed-precision quantization methods when conditioning both on general-purpose and task-specific data. Across QA, math reasoning, and text-to-SQL tasks for both Llama-3 and Qwen2.5, we find that TaCQ outperforms baselines using the same calibration data and a lower weight budget, achieving major improvements in the 2 and 3-bit regime. With only 3.1 bits we are able to recover 96% of Llama-3-8B-Instruct's unquantized 16-bit MMLU performance, obtaining a 5.25% absolute improvement over SPQR. We also observe consistently large gains over existing methods in the 2-bit regime, with an average gain of 14.74% over the strongest baseline, SliM-LLM. Moreover, we observe a 7.20% gain without conditioning on specific tasks, showing TaCQ's ability to identify important weights is not limited to task-conditioned settings.

cs.LG↗

arXiv2025

UPCORE: Utility-Preserving Coreset Selection for Balanced Unlearning

User specifications or legal frameworks often require information to be removed from pretrained models, including large language models (LLMs). This requires deleting or "forgetting" a set of data points from an already-trained model, which typically degrades its performance on other data points. Thus, a balance must be struck between removing information and keeping the model's other abilities intact, with a failure to balance this trade-off leading to poor deletion or an unusable model. To this end, we propose UPCORE (Utility-Preserving Coreset Selection), a method-agnostic data selection framework for mitigating collateral damage during unlearning. Finding that the model damage is correlated with the variance of the model's representations on the forget set, we selectively prune the forget set to remove outliers, thereby minimizing model degradation after unlearning. Across three standard unlearning methods, UPCORE consistently achieves a superior balance between the competing objectives of deletion efficacy and model preservation. To better evaluate this trade-off, we introduce a new metric, measuring the area-under-the-curve (AUC) across standard metrics. Our results show that UPCORE improves both standard metrics and AUC, benefiting from positive transfer between the coreset and pruned points while reducing negative transfer from the forget set to points outside of it.

cs.LG↗

arXiv2025

GenerationPrograms: Fine-grained Attribution with Executable Programs

Recent large language models (LLMs) achieve impressive performance in source-conditioned text generation but often fail to correctly provide fine-grained attributions for their outputs, undermining verifiability and trust. Moreover, existing attribution methods do not explain how and why models leverage the provided source documents to generate their final responses, limiting interpretability. To overcome these challenges, we introduce a modular generation framework, GenerationPrograms, inspired by recent advancements in executable "code agent" architectures. Unlike conventional generation methods that simultaneously generate outputs and attributions or rely on post-hoc attribution, GenerationPrograms decomposes the process into two distinct stages: first, creating an executable program plan composed of modular text operations (such as paraphrasing, compression, and fusion) explicitly tailored to the query, and second, executing these operations following the program's specified instructions to produce the final response. Empirical evaluations demonstrate that GenerationPrograms significantly improves attribution quality at both the document level and sentence level across two long-form question-answering tasks and a multi-document summarization task. We further demonstrate that GenerationPrograms can effectively function as a post-hoc attribution method, outperforming traditional techniques in recovering accurate attributions. In addition, the interpretable programs generated by GenerationPrograms enable localized refinement through modular-level improvements that further enhance overall attribution quality.

cs.CL↗