Searcharxiv⌕ Search

arXiv subjects

Chirag Agarwal

Publications and source records attributed to Chirag Agarwal.

At least 19 recordsLinked to original sources

RadRef: A Reference-Free Metric for Radiology Report Generation

AI-generated radiology reports are commonly evaluated against radiologist-authored references, limiting the applicability of standard reference-based metrics when no reference is available. We introduce RadRef, a reference-free metric that scores report claims against image-derived beliefs over a clinical finding vocabulary, weighted by relevance to the clinical indication. By evaluating findings beyond the report's stated claims, RadRef assesses omissions as well as unsupported assertions. On ReXVal, RadRef achieves $73.1\%$ agreement with radiologist rankings among non-tied same-radiograph report pairs, with $61.8\%$ pair coverage and $64.3\%$ tie-inclusive agreement. Its non-tied agreement is close to GREEN ($73.9\%$) and RadFact ($72.6\%$), with its strongest relative results on omissions. We further evaluate $100$ model-generated and $100$ radiologist-authored reports from $100$ MIMIC-CXR studies, independently annotated by two board-certified radiologists. On same-image omission rankings, RadRef attains Kendall's $τ_b = 0.295$, compared with $0.386$ between the readers. In separate paired comparisons across generated reports, no evaluated comparator significantly outperforms RadRef on omissions. The same image-grounded representation provides a false-finding readout, RadRef-FF, with $τ_b = 0.32$ on ReXVal and $0.29$ in the reader study. Controlled perturbations confirm that RadRef depends on the radiograph, report, and clinical indication. RadRef enables image-grounded evaluation before a radiologist-authored reference exists, supporting quality monitoring of report-generation systems in clinical settings, and complements reference-based metrics when one is available.

cs.LG↗

MEA: A Reward-Driven Multi-Agent System for Faithful Model Explanations

Recent years have seen the employment of a plethora of machine learning (ML) models in high-stakes domains, but they remain largely opaque to the practitioners who act on their predictions. While post-hoc explanation methods offer a lens into this model behavior, wielding them effectively demands expertise most domain experts lack: navigating high-dimensional outputs, selecting the best explanations, and synthesizing evidence across disparate tools. To this end, we present MEA, a multi-agent framework that removes the explanation knowledge barrier entirely: a Proposer agent selects and configures explanation tools based on the question and modality, while an Actor agent is optimized end-to-end against faithfulness, transforming the outputs into natural language explanations grounded in model behavior across tabular, text, and vision modalities. Further, we introduce diverse question types spanning feature attribution, counterfactual reasoning, and spurious feature detection, each paired with a perturbation-based faithfulness metric. We find that frontier LLMs systematically produce unfaithful explanations. By optimizing against faithfulness rewards augmented with a modality-adaptive penalty, MEA consistently outperforms post hoc explainers, agentic, and closed-source baselines across six datasets, with reward-driven optimization yielding faithfulness gains of +28% (tabular), +21% (text), and +34% (vision) over the untrained backbone. More broadly, our findings suggest that AI agents themselves can serve as a scalable, adaptable interface to ML explainability, opening a path toward natural-language explainability that generalizes beyond the fixed, single-purpose tools that have long defined the field.

cs.AI↗

Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility

Sparse autoencoders (SAEs) are typically trained on hundreds of millions to billions of activation tokens, yet little work has tested whether this scale is necessary. We study how much activation data SAEs need by training four SAE architectures on three language models with budgets ranging from 5M tokens to a 500M-token reference. We measure two notions of convergence: functional behavior on standard evaluations and one-to-one reproducibility of learned features. Functional data requirements are strongly metric-dependent: reconstruction often approaches the full-data reference at substantially smaller budgets, whereas feature reproducibility continues to improve with data. For Gemma-2-2B BatchTopK, the fraction of strictly matched features rises from 2\% at 5M tokens to 29\% at 50M. Data selection also matters: random sampling better preserves the original activation distribution and generally favors reconstruction, whereas diversity-based sampling frequently improves AutoInterp. A theoretical analysis separates statistical consistency of subsampled training from the stronger identifiability and optimization conditions required for one-to-one feature recovery. Our results show that SAE training can often be made far more data-efficient when budgets are matched to the goal, and feature-level claims are backed by explicit reproducibility checks, paving the way toward cheaper and more reliable SAE research.

cs.LG↗

When Agents Collude: Weak-to-Strong On-Policy Distillation for Mitigating Strategic Deception in Multi-Agent Systems

While recent advances in autonomous agents enable them to coordinate on increasingly complex tasks, prior work shows that this growing capacity for coordination can create safety risks, including collusion, in which agents jointly pursue misaligned actions while concealing the true basis of those actions from oversight. We study this risk in a simulated insider-trading environment, comparing single agents with sequential and decentralized teams across seven language models from three families. By varying agents' system-level objectives and pressure conditions, we examine when coordinated agents act on material nonpublic information, conceal the basis of a misaligned trade in a report, and maintain that deception under follow-up questioning. Compared to single-agent decision-making, multi-agent interaction raises misaligned-action rates from 35.4% to 68.2% and persistent-deception rates from 31.0% to 62.2%, both measured over all runs. These effects are largest under existential pressure, where teams sustain deception through follow-up questioning in 87% of all runs. Qualitative analysis reveals shared cover stories and coordinated rationales, suggesting group-level strategic deception rather than isolated reporting errors. To mitigate this behavior, we introduce PACT, a weak-to-strong on-policy distillation framework for mitigating collusive deception. Using generic honesty prompts, PACT trains a stronger student on its own sampled continuations, matching its next-token distribution to that of a smaller, frozen honesty-aligned teacher via reverse-KL supervision. Under existential pressure with ethical instructions, PACT reduces multi-agent persistent deception from 80% to 3%. Our results show that multi-agent coordination amplifies strategic deception and that on-policy honesty alignment can substantially mitigate these failures.

cs.MA↗

Mechanistic Interpretability of Atmospheric Rivers in GraphCast

While AI weather models now rival operational forecasts, how they represent the atmosphere internally remains an open question: feature attribution reveals which input patterns matter, not what the model computes or how it combines information internally. We train sparse autoencoders (SAEs) on GraphCast to uncover its learned concepts, using atmospheric rivers as our phenomenon of focus. Both standard and Matryoshka SAEs show GraphCast computes atmospheric river intensity, measured by integrated vapor transport (IVT), as a stable internal variable, despite IVT being neither an input nor a target. In contrast to the unstructured concept retrieval of the standard SAE, the Matryoshka SAE orders concepts by importance and exposes their relations. Atmospheric river concepts persist across depth and direct interventions confirm causality. This method offers a way to find internal variables and determine which of them the model actually relies on, which is a prerequisite for asking whether those variables remain meaningful as the phenomenon changes under a warming climate.

cs.LG↗

The Fragility of Chain-of-Thought Monitoring Across Typologically Diverse Languages

Chain-of-thought (CoT) monitoring has been proposed as a promising safety mechanism for detecting misaligned behavior in large language models. However, its reliability remains largely unexplored beyond English and across diverse model families. We present the first large-scale evaluation of CoT monitorability across 13 diverse languages and seven frontier model families, comprising 16 models. Using adversarial-hint evaluations that require explicit intermediate computation, together with analysis of internal answer-token probabilities, we consistently find CoT unfaithfulness across languages and hint types, with an average rate of 95.9\% across 8B--120B parameter models. We find that frontier models systematically exhibit strategic manipulation, including answer-switching, post-hoc rationalization, and procedural exploitation of hints, making their reasoning difficult to reliably monitor. These deceptive patterns remain especially pronounced in low-resource languages, revealing fundamental limitations in current CoT-based oversight. Our results show that CoT monitoring is fragile under linguistic distribution shift, providing a substantially weaker safety signal than English-only studies suggest. These findings motivate the development of more robust CoT monitors and complementary white-box monitoring techniques, particularly for mid- and low-resource languages. Our code is available \href{https://multilingual-cot-monitoring.github.io/}{\textcolor{blue}{here}}.

cs.CL↗

STRUCTUREDAGENT: Planning with AND/OR Trees for Long-Horizon Web Tasks

Existing LLM-based web agents struggle on complex, long-horizon tasks due to limited in-context memory, weak planning abilities, and greedy behaviors that lead to premature termination. To address these challenges, we propose \SA{}, a hierarchical planning framework that interleaves planning and execution via dynamic $\ANDOR$ trees. The framework separates structural planning from LLM-based reasoning, enabling principled error recovery through node repair, systematic exploration of alternatives via OR nodes, and modular plans that can facilitate human intervention. On WebArena (630 tasks), \SA{} achieves a $\sim$53\% success rate vs.\ $\sim$46\% for AgentOccam, and on complex multi-constraint Amazon shopping tasks, gains reach 10\% over the strongest baseline.

cs.AI↗

PageGuide: Browser extension to assist users in navigating a webpage and locating information

Users browsing the web struggle to locate relevant information on cluttered pages and to complete web navigation tasks. Current web agents can answer questions and automate actions, but return answers without showing where the information comes from, forcing users to manually verify results and blindly trust every automated step. We present PageGuide, a browser extension that grounds LLM answers in page content (e.g. HTML DOM elements, text, maps, images, charts), via two features: FIND, which highlights supporting evidence in situ on a page, and GUIDE, which guides users across pages with evidence captured at each step. In a within-subject study (N=92), PageGuide improves overall verification accuracy from 65% to 77% and reduces verification time from 73s to 67s across FIND and GUIDE tasks. PageGuide also reduces manual search effort and improves users' perceived support for verification. Code is available at https://pageguide.github.io/.

cs.HC↗

The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing

Autonomous AI systems are transitioning from advisory roles to autonomous ones for medication prescriptions. Recent U.S. bill H.R. 238 and Utah's prescription-renewal pilot program both authorize AI to prescribe medications in an agentic capacity. While many regulatory guidelines suggest aggregate model performance metrics at the point of clearance, they do not require i) calibrated per-prediction confidence for action-gated thresholds, ii) differentiated communication between uncertainty arising from model ignorance (epistemic) from genuine clinical ambiguity (aleatoric), and iii) inferential transparency at the moment of decision enabling liability allocation. Here, we argue these three architectural features are minimum conditions for safe autonomous prescribing, and validate them with a survey of 136 U.S. prescribing clinicians. Our results suggest prescribing clinicians i) would not permit autonomous prescribing without a confidence-based escalation mechanism, ii) preferred a competing-options summary for aleatoric uncertainty but preferred abstention for epistemic uncertainty, and iii) were only willing to accept liability when inferential transparency enabled them to make a decision under acknowledged uncertainty. These findings indicate that our recommended architectural features would encourage higher rates of clinician adoption of autonomous AI prescribing, largely through collapsing much of what "autonomy" conventionally means.

cs.AI↗

A Mechanistic Perspective and Circuit-Guided Difficulty Metric for Unlearning

Machine unlearning is becoming essential for building trustworthy and compliant language models. Yet unlearning success varies considerably across individual samples: some are reliably erased, while others persist despite the same procedure. We argue that this disparity is not only a data-side phenomenon, but also reflects model-internal mechanisms that encode and protect memorized information. We study this problem from a mechanistic perspective based on model circuits--structured interaction pathways that govern how predictions are formed. We propose Circuit-guided Unlearning Difficulty (CUD), a {\em pre-unlearning} metric that assigns each sample a continuous difficulty score using circuit-level signals. Extensive experiments demonstrate that CUD reliably separates intrinsically easy and hard samples, and remains stable across unlearning methods. We identify key circuit-level patterns that reveal a mechanistic signature of difficulty: easy-to-unlearn samples are associated with shorter, shallower interactions concentrated in earlier-to-intermediate parts of the original model, whereas hard samples rely on longer and deeper pathways closer to late-stage computation. Compared to existing qualitative studies, CUD takes a first step toward a principled, fine-grained, and interpretable analysis of unlearning difficulty; and motivates the development of unlearning methods grounded in model mechanisms.

cs.LG↗

When Graph Tokens Sink: A Mechanistic Analysis of Graph Language Models

Graph Language Models (GLMs) have become a promising direction for adapting Large Language Models (LLMs) to graph learning tasks. By transforming graph topology and node information into graph tokens, GLMs allow LLMs to jointly process structured graph inputs and textual instructions. Yet, it remains unclear how LLMs internally interpret these graph tokens and whether graph tokens act as meaningful carriers of graph structure. In this work, we analyze how LLMs process graph information through graph-token behavior in representative GLM architectures. Findings. We find that the internal saliency of graph tokens in GLMs is not equivalent to graph information utilization. Graph sink tokens consistently emerge as activation-level outliers: they can be identified by massive activation values along a small set of hidden-state dimensions and are biased toward early graph-token positions. However, this activation-level saliency does not imply that these tokens are the main carriers of graph information. Unlike classical attention sinks in language and vision-language models, graph sink tokens do not necessarily attract the largest attention weights from query tokens. Through pruning, repositioning, and swapping interventions, we show that graph sink tokens are not the most important semantic or structural tokens for downstream prediction. Implications. Together, these results suggest that after current GLMs map graph structure into the LLM token space, the resulting graph-token representations do not naturally form a fully usable topology-aware internal representation; instead, they exhibit a decoupling between activation-level saliency and graph-semantic utility. This decoupling points to limitations in existing graph-token construction, placement, and alignment mechanisms.

cs.LG↗

GridVQA-X: A Framework for Evaluating Multimodal Explainability Methods

With the increasing development of Vision-Language Models, it becomes imperative that their predictions are readily explainable to relevant stakeholders. However, the field of explainability has not kept pace with the multimodal surge. While recent Multimodal Explainable AI (MxAI) methods generate explanations to attribute the interaction between different modalities, current evaluation protocols lack the ground truth required to distinguish between true cross-modal reasoning (e.g., spatial composition) and shallow cross-modal shortcuts (e.g., Bag-of-Words attribute matching). It remains unknown whether MxAI methods faithfully capture synergistic interactions or merely hallucinate reasoning on models acting as simple feature detectors. In this paper, we introduce GridVQA-X, the first diagnostic framework specifically designed to evaluate cross-modal explainability. Unlike natural datasets, GridVQA-X leverages a closed-world synthesis logic to generate unique, mathematically guaranteed explanations. We utilize this controlled environment to train paired ground-truth models on identical architectures: $M_{\text{pure}}$, which learns robust spatial-relational reasoning and $M_{\text{spur}}$, which is structurally forced to rely on cross-modal shortcuts. This behavioral divergence creates a rigorous testbed: a faithful explainer must report distinct reasoning pathways for each model. Our findings reveal that widely used methods fail to distinguish between models relying on genuine spatial-relational reasoning and those exploiting cross-modal shortcuts, highlighting a critical gap in capturing true cross-modal synergy and misrepresenting how multimodal models actually make decisions.

cs.CV↗

CURE-Med: Curriculum-Informed Reinforcement Learning for Multilingual Medical Reasoning

While large language models (LLMs) have shown to perform well on monolingual mathematical and commonsense reasoning, they remain unreliable for multilingual medical reasoning applications, hindering their deployment in multilingual healthcare settings. We address this by first introducing CUREMED-BENCH, a high-quality multilingual medical reasoning dataset with open-ended reasoning queries with a single verifiable answer, spanning thirteen languages, including underrepresented languages such as Amharic, Yoruba, and Swahili. Building on this dataset, we propose CURE-MED, a curriculum-informed reinforcement learning framework that integrates code-switching-aware supervised fine-tuning and Group Relative Policy Optimization to jointly improve logical correctness and language stability. Across thirteen languages, our approach consistently outperforms strong baselines and scales effectively, achieving 85.21% language consistency and 54.35% logical correctness at 7B parameters, and 94.96% language consistency and 70.04% logical correctness at 32B parameters. These results support reliable and equitable multilingual medical reasoning in LLMs. The code and dataset are available at https://cure-med.github.io/

cs.AI↗

Towards Understanding the Robustness of Sparse Autoencoders

Large Language Models (LLMs) remain vulnerable to optimization-based jailbreak attacks that exploit internal gradient structure. While Sparse Autoencoders (SAEs) are widely used for interpretability, their robustness implications remain underexplored. We present a study of integrating pretrained SAEs into transformer residual streams at inference time, without modifying model weights or blocking gradients. Across four model families (Gemma, LLaMA, Mistral, Qwen) and two strong white-box attacks (GCG, BEAST) plus three black-box benchmarks, SAE-augmented models achieve up to a 5x reduction in jailbreak success rate relative to the undefended baseline and reduce cross-model attack transferability. Parametric ablations reveal (i) a monotonic dose-response relationship between L0 sparsity and attack success rate, and (ii) a layer-dependent defense-utility tradeoff, where intermediate layers balance robustness and clean performance. These findings are consistent with a representational bottleneck hypothesis: sparse projection reshapes the optimization geometry exploited by jailbreak attacks.

cs.LG↗

Do Vision Language Models Need to Process Image Tokens?

Vision Language Models (VLMs) have achieved remarkable success by integrating visual encoders with large language models (LLMs). While VLMs process dense image tokens across deep transformer stacks (incurring substantial computational overhead), it remains fundamentally unclear whether sustained image-token processing is necessary for their performance or visual representations meaningfully evolve from early to later layers. In this work, we systematically investigate the functional role of image tokens in VLMs and show that visual representations rapidly converge to a bounded-complexity regime, \ie their entropy stabilizes, intrinsic dimensionality compresses, and trajectory curvature approaches a near-constant profile. In contrast, textual representations continue to undergo substantial restructuring across depth. Once stabilized, visual representations become largely interchangeable between layers, indicating limited additional transformation in deeper stages. Further, depth-wise visual truncation reveals that the necessity of visual processing is task-dependent, where single-token predictions remain comparatively robust to truncated visual depth, but multi-token generation require sustained access to visual representations. Under deterministic decoding, reducing visual depth perturbs intermediate reasoning trajectories more strongly than final outputs, suggesting that image tokens influence the structure of reasoning more than the ultimate conclusions. Collectively, these findings \textbf{question the assumption} that deeper visual processing is uniformly essential in VLMs, challenging the current paradigm of multimodal LLM architectures.

cs.CV↗

SynopticBench: Evaluating Vision-Language Models on Generating Weather Forecast Discussions of the Future

Recent advances in visual-language models (VLMs) have led to significant improvements in a plethora of complex multimodal tasks like image captioning, report generation, and visual perception. However, generating text from meteorological data is highly challenging because the atmosphere is a chaotic system that is rapidly changing at various spatial and temporal scales. Given the complexity of atmospheric phenomena, it is critical to verifiably quantify the effectiveness of existing VLMs on weather forecasting data. In this work, we present SynopticBench, a high-quality dataset consisting of 1,367,041 text samples of Area Forecast Discussions created by the National Weather Service over the continental United States paired to images of 500mb geopotential height, 2 meter temperature, and 850mb wind velocity in weather forecasts. We also present Synoptic Phenomena Alignment and Coverage Evaluation (SPACE), a novel evaluation framework that can be used to effectively estimate the quality of text descriptions of synoptic weather phenomena. Extensive experiments on generating forecast discussions using state-of-the-art VLMs show the sensitivity of existing evaluation metrics in this domain and enable further exploration into synoptic weather and climate text generation.

cs.CL↗

Quantifying Explanation Quality in Graph Neural Networks using Out-of-Distribution Generalization

Evaluating the quality of post-hoc explanations for Graph Neural Networks (GNNs) remains a significant challenge. While recent years have seen an increasing development of explainability methods, current evaluation metrics (e.g., fidelity, sparsity) often fail to assess whether an explanation identifies the true underlying causal variables. To address this, we propose the Explanation-Generalization Score (EGS), a metric that quantifies the causal relevance of GNN explanations. EGS is founded on the principle of feature invariance and posits that if an explanation captures true causal drivers, it should lead to stable predictions across distribution shifts. To quantify this, we introduce a framework that trains GNNs using explanatory subgraphs and evaluates their performance in Out-of-Distribution (OOD) settings (here, OOD generalization serves as a rigorous proxy for the explanation's causal validity). Through large-scale validation involving 11,200 model combinations across synthetic and real-world datasets, our results demonstrate that EGS provides a principled benchmark for ranking explainers based on their ability to capture causal substructures, offering a robust alternative to traditional fidelity-based metrics.

cs.LG↗

Improving Human Verification of LLM Reasoning through Interactive Explanation Interfaces

The reasoning capabilities of Large Language Models (LLMs) have led to their increasing employment in several critical applications, particularly education, where they support problem-solving, tutoring, and personalized study. Chain-of-thought (CoT) reasoning capabilities [1, 2] are well-known to help LLMs decompose a problem into steps and explore the solution spaces more effectively, leading to impressive performance on mathematical and reasoning benchmarks. As the length of CoT tokens per question increases substantially to even thousands of tokens per question [ 1], it is unknown how users could comprehend LLM reasoning and detect errors or hallucinations. To address this problem and understand how reasoning can improve human-AI interaction, we present three new interactive reasoning interfaces: interactive CoT (iCoT), interactive Program-of-Thought (iPoT), and interactive Graph (iGraph). That is, we ask LLMs themselves to generate an interactive web interface wrapped around the original CoT content, which may be presented in text (iCoT), graphs (iGraph) or code (iPoT). This interface allows users to interact with and provide a novel experience in reading and validating the reasoning chains of LLMs. Across a study of 125 participants, interactive interfaces significantly improve user performance. Specifically, iGraph users score the highest error detection rate (85.6%), followed by iPoT (82.5%), iCoT (80.6%), all outperforming standard CoT (73.5%). Interactive interfaces also lead to faster user validation time-iGraph users are faster (57.9 secs per question) than the users of iCoT and iPoT (60 secs) and the standard CoT (64.7 secs). A post-study questionnaire shows that users prefer iGraph, citing its superior ability to enable them to follow the LLM's reasoning. We discuss the implications of these results and provide recommendations for the future design of reasoning models.

cs.HC↗