arXiv subjects

Dhruv Kumar

Publications and source records attributed to Dhruv Kumar.

At least 55 records · Page 3Linked to original sources

arXiv2026

Trust Regions Sell, But Who's Buying? Overlap Geometry as an Alternative Trust Region for Policy Optimization

Standard trust-region methods constrain policy updates via Kullback-Leibler (KL) divergence. However, KL controls only an average divergence and does not directly prevent rare, large likelihood-ratio excursions that destabilize training--precisely the failure mode that motivates heuristics such as PPO's clipping. We propose overlap geometry as an alternative trust region, constraining distributional overlap via the Bhattacharyya coefficient (closely related to the Hellinger/Renyi-1/2 geometry). This objective penalizes separation in the ratio tails, yielding tighter control over likelihood-ratio excursions without relying on total variation bounds that can be loose in tail regimes. We derive Bhattacharyya-TRPO (BTRPO) and Bhattacharyya-PPO (BPPO), enforcing overlap constraints via square-root ratio updates: BPPO clips the square-root ratio q = sqrt(r), and BTRPO applies a quadratic Hellinger/Bhattacharyya penalty. Empirically, overlap-based updates improve robustness and aggregate performance as measured by RLiable under matched training budgets, suggesting overlap constraints as a practical, principled alternative to KL for stable policy optimization.

cs.LG↗

arXiv2026

How to Trick Your AI TA: A Systematic Study of Academic Jailbreaking in LLM Code Evaluation

The use of Large Language Models (LLMs) as automatic judges for code evaluation is becoming increasingly prevalent in academic environments. But their reliability can be compromised by students who may employ adversarial prompting strategies in order to induce misgrading and secure undeserved academic advantages. In this paper, we present the first large-scale study of jailbreaking LLM-based automated code evaluators in academic context. Our contributions are: (i) We systematically adapt 20+ jailbreaking strategies for jailbreaking AI code evaluators in the academic context, defining a new class of attacks termed academic jailbreaking. (ii) We release a poisoned dataset of 25K adversarial student submissions, specifically designed for the academic code-evaluation setting, sourced from diverse real-world coursework and paired with rubrics and human-graded references, and (iii) In order to capture the multidimensional impact of academic jailbreaking, we systematically adapt and define three jailbreaking metrics (Jailbreak Success Rate, Score Inflation, and Harmfulness). (iv) We comprehensively evalulate the academic jailbreaking attacks using six LLMs. We find that these models exhibit significant vulnerability, particularly to persuasive and role-play-based attacks (up to 97% JSR). Our adversarial dataset and benchmark suite lay the groundwork for next-generation robust LLM-based evaluators in academic code assessment.

cs.SE↗

arXiv2026

The Compliance Paradox: Semantic-Instruction Decoupling in Automated Academic Code Evaluation

The rapid integration of Large Language Models (LLMs) into educational assessment rests on the unverified assumption that instruction following capability translates directly to objective adjudication. We demonstrate that this assumption is fundamentally flawed. Instead of evaluating code quality, models frequently decouple from the submission's logic to satisfy hidden directives, a systemic vulnerability we term the Compliance Paradox, where models fine-tuned for extreme helpfulness are vulnerable to adversarial manipulation. To expose this, we introduce the Semantic-Preserving Adversarial Code Injection (SPACI) Framework and the Abstract Syntax Tree-Aware Semantic Injection Protocol (AST-ASIP). These methods exploit the Syntax-Semantics Gap by embedding adversarial directives into syntactically inert regions (trivia nodes) of the Abstract Syntax Tree. Through a large-scale evaluation of 9 SOTA models across 25,000 submissions in Python, C, C++, and Java, we reveal catastrophic failure rates (>95%) in high-capacity open-weights models like DeepSeek-V3, which systematically prioritize hidden formatting constraints over code correctness. We quantify this failure using our novel tripartite framework measuring Decoupling Probability, Score Divergence, and Pedagogical Severity to demonstrate the widespread "False Certification" of functionally broken code. Our findings suggest that current alignment paradigms create a "Trojan" vulnerability in automated grading, necessitating a shift from standard RLHF toward domain-specific Adjudicative Robustness, where models are conditioned to prioritize evidence over instruction compliance. We release our complete dataset and injection framework to facilitate further research on the topic.

cs.CL↗

arXiv2026

A Hybrid Supervised-LLM Pipeline for Actionable Suggestion Mining in Unstructured Customer Reviews

Extracting actionable suggestions from customer reviews is essential for operational decision-making, yet these directives are often embedded within mixed-intent, unstructured text. Existing approaches either classify suggestion-bearing sentences or generate high-level summaries, but rarely isolate the precise improvement instructions businesses need. We evaluate a hybrid pipeline combining a high-recall RoBERTa classifier trained with a precision-recall surrogate to reduce unrecoverable false negatives with a controlled, instruction-tuned LLM for suggestion extraction, categorization, clustering, and summarization. Across real-world hospitality and food datasets, the hybrid system outperforms prompt-only, rule-based, and classifier-only baselines in extraction accuracy and cluster coherence. Human evaluations further confirm that the resulting suggestions and summaries are clear, faithful, and interpretable. Overall, our results show that hybrid reasoning architectures achieve meaningful improvements fine-grained actionable suggestion mining while highlighting challenges in domain adaptation and efficient local deployment.

cs.CL↗

arXiv2026

neuralFOMO: Can LLMs Handle Being Second Best? Measuring Envy-Like Preferences in Multi-Agent Settings

Envy shapes competitiveness and cooperation in human groups, yet its role in large language model interactions remains largely unexplored. As LLMs increasingly operate in multi-agent settings, it is important to examine whether they exhibit envy-like preferences under social comparison. We evaluate LLM behavior across two scenarios: (1) a point-allocation game testing sensitivity to relative versus absolute payoff, and (2) comparative evaluations across general and contextual settings. To ground our analysis in psychological theory, we adapt four established psychometric questionnaires spanning general, domain-specific, workplace, and sibling-based envy. Our results reveal heterogeneous envy-like patterns across models and contexts, with some models sacrificing personal gain to reduce a peer's advantage, while others prioritize individual maximization. These findings highlight competitive dispositions as a design and safety consideration for multi-agent LLM systems.

cs.AI↗

arXiv2026

PhysicsSolutionAgent: Towards Multimodal Explanations for Numerical Physics Problem Solving

Explaining numerical physics problems often requires more than text-based solutions; clear visual reasoning can substantially improve conceptual understanding. While large language models (LLMs) demonstrate strong performance on many physics questions in textual form, their ability to generate long, high-quality visual explanations remains insufficiently explored. In this work, we introduce PhysicsSolutionAgent (PSA), an autonomous agent that generates physics-problem explanation videos of up to six minutes using Manim animations. To evaluate the generated videos, we design an assessment pipeline that performs automated checks across 15 quantitative parameters and incorporates feedback from a vision-language model (VLM) to iteratively improve video quality. We evaluate PSA on 32 videos spanning numerical and theoretical physics problems. Our results reveal systematic differences in video quality depending on problem difficulty and whether the task is numerical or theoretical. Using GPT-5-mini, PSA achieves a 100% video-completion rate with an average automated score of 3.8/5. However, qualitative analysis and human inspection uncover both minor and major issues, including visual layout inconsistencies and errors in how visual content is interpreted during feedback. These findings expose key limitations in reliable Manim code generation and highlight broader challenges in multimodal reasoning and evaluation for visual explanations of numerical physics problems. Our work underscores the need for improved visual understanding, verification, and evaluation frameworks in future multimodal educational systems

cs.CL↗

arXiv2026

Actionable Advice from Reviews via Mixture of LoRA Experts: A Two-LLM Pipeline for Issue Extraction and Business Recommendations

Customer reviews contain detailed, domain specific signals about service failures and user expectations, but converting this unstructured feedback into actionable business decisions remains difficult. We study review-to-action generation: producing concrete, implementable recommendations grounded in review text. We propose a modular two-LLM framework in which an Issue model extracts salient issues and assigns coarse themes, and an Advice model generates targeted operational fixes conditioned on the extracted issue representation. To enable specialization without expensive full fine-tuning, we adapt the Advice model using a mixture of LoRA experts strategy: multiple low-rank adapters are trained and a lightweight gating mechanism performs token-level expert mixing at inference, combining complementary expertise across issue types. We construct synthetic review-issue-advice triples from Yelp reviews (airlines and restaurants) to supervise training, and evaluate recommendations using an eight dimension operational rubric spanning actionability, specificity, feasibility, expected impact, novelty, non-redundancy, bias, and clarity. Across both domains, our approach consistently outperforms prompting-only and single-adapter baselines, yielding higher actionability and specificity while retaining favorable efficiency-quality trade-offs.

cs.AI↗

arXiv2026

TabPFN Through The Looking Glass: An interpretability study of TabPFN and its internal representations

Tabular foundational models are pre-trained models designed for a wide range of tabular data tasks. They have shown strong performance across domains, yet their internal representations and learned concepts remain poorly understood. This lack of interpretability makes it important to study how these models process and transform input features. In this work, we analyze the information encoded inside the model's hidden representations and examine how these representations evolve across layers. We run a set of probing experiments that test for the presence of linear regression coefficients, intermediate values from complex expressions, and the final answer in early layers. These experiments allow us to reason about the computations the model performs internally. Our results provide evidence that meaningful and structured information is stored inside the representations of tabular foundational models. We observe clear signals that correspond to both intermediate and final quantities involved in the model's prediction process. This gives insight into how the model refines its inputs and how the final output emerges. Our findings contribute to a deeper understanding of the internal mechanics of tabular foundational models. They show that these models encode concrete and interpretable information, which moves us closer to making their decision processes more transparent and trustworthy.

cs.LG↗

arXiv2026

SAC: A Framework for Measuring and Inducing Personality Traits in LLMs with Dynamic Intensity Control

Large language models (LLMs) have gained significant traction across a wide range of fields in recent years. There is also a growing expectation for them to display human-like personalities during interactions. To meet this expectation, numerous studies have proposed methods for modelling LLM personalities through psychometric evaluations. However, most existing models face two major limitations: they rely on the Big Five (OCEAN) framework, which only provides coarse personality dimensions, and they lack mechanisms for controlling trait intensity. In this paper, we address this gap by extending the Machine Personality Inventory (MPI), which originally used the Big Five model, to incorporate the 16 Personality Factor (16PF) model, allowing expressive control over sixteen distinct traits. We also developed a structured framework known as Specific Attribute Control (SAC) for evaluating and dynamically inducing trait intensity in LLMs. Our method introduces adjective-based semantic anchoring to guide trait intensity expression and leverages behavioural questions across five intensity factors: \textit{Frequency}, \textit{Depth}, \textit{Threshold}, \textit{Effort}, and \textit{Willingness}. Through experimentation, we find that modelling intensity as a continuous spectrum yields substantially more consistent and controllable personality expression compared to binary trait toggling. Moreover, we observe that changes in target trait intensity systematically influence closely related traits in psychologically coherent directions, suggesting that LLMs internalize multi-dimensional personality structures rather than treating traits in isolation. Our work opens new pathways for controlled and nuanced human-machine interactions in domains such as healthcare, education, and interviewing processes, bringing us one step closer to truly human-like social machines.

cs.CL↗

arXiv2026

When Reject Turns into Accept: Quantifying the Vulnerability of LLM-Based Scientific Reviewers to Indirect Prompt Injection

Driven by surging submission volumes, scientific peer review has catalyzed two parallel trends: individual over-reliance on LLMs and institutional AI-powered assessment systems. This study investigates the robustness of "LLM-as-a-Judge" systems to adversarial PDF manipulation via invisible text injections and layout aware encoding attacks. We specifically target the distinct incentive of flipping "Reject" decisions to "Accept," a vulnerability that fundamentally compromises scientific integrity. To measure this, we introduce the Weighted Adversarial Vulnerability Score (WAVS), a novel metric that quantifies susceptibility by weighting score inflation against the severity of decision shifts relative to ground truth. We adapt 15 domain-specific attack strategies, ranging from semantic persuasion to cognitive obfuscation, and evaluate them across 13 diverse language models (including GPT-5 and DeepSeek) using a curated dataset of 200 official and real-world accepted and rejected submissions (e.g., ICLR OpenReview). Our results demonstrate that obfuscation techniques like "Maximum Mark Magyk" and "Symbolic Masking & Context Redirection" successfully manipulate scores, achieving decision flip rates of up to 86.26% in open-source models, while exposing distinct "reasoning traps" in proprietary systems. We release our complete dataset and injection framework to facilitate further research on the topic (https://anonymous.4open.sciencer/llm-jailbreak-FC9E/).

cs.AI↗

arXiv2025

Beyond Single Bugs: Benchmarking Large Language Models for Multi-Vulnerability Detection

Large Language Models (LLMs) have demonstrated significant potential in automated software security, particularly in vulnerability detection. However, existing benchmarks primarily focus on isolated, single-vulnerability samples or function-level classification, failing to reflect the complexity of real-world software where multiple interacting vulnerabilities often coexist within large files. Recent studies indicate that LLMs suffer from "count bias" and "selection bias" in multi-label tasks, yet this has not been rigorously quantified in the domain of code security. In this work, we introduce a comprehensive benchmark for Multi-Vulnerability Detection across four major languages: C, C++, Python, and JavaScript. We construct a dataset of 40,000 files by systematically injecting controlled counts of vulnerabilities (1, 3, 5, and 9) into long-context code samples (7.5k-10k tokens) sourced from CodeParrot. We evaluate five state-of-the-art LLMs, including GPT-4o-mini, Llama-3.3-70B, and the Qwen-2.5 series. Our results reveal a sharp degradation in performance as vulnerability density increases. While Llama-3.3-70B achieves near-perfect F1 scores (approximately 0.97) on single-vulnerability C tasks, performance drops by up to 40% in high-density settings. Notably, Python and JavaScript show distinct failure modes compared to C/C++, with models exhibiting severe "under-counting" (Recall dropping to less than 0.30) in complex Python files.

cs.CR↗

arXiv2025

Toward Intelligent Scene Augmentation for Context-Aware Object Placement and Sponsor-Logo Integration

Intelligent image editing increasingly relies on advances in computer vision, multimodal reasoning, and generative modeling. While vision-language models (VLMs) and diffusion models enable guided visual manipulation, existing work rarely ensures that inserted objects are \emph{contextually appropriate}. We introduce two new tasks for advertising and digital media: (1) \emph{context-aware object insertion}, which requires predicting suitable object categories, generating them, and placing them plausibly within the scene; and (2) \emph{sponsor-product logo augmentation}, which involves detecting products and inserting correct brand logos, even when items are unbranded or incorrectly branded. To support these tasks, we build two new datasets with category annotations, placement regions, and sponsor-product labels.

cs.CV↗

arXiv2025

Understanding Virality: A Rubric based Vision-Language Model Framework for Short-Form Edutainment Evaluation

Evaluating short-form video content requires moving beyond surface-level quality metrics toward human-aligned, multimodal reasoning. While existing frameworks like VideoScore-2 assess visual and semantic fidelity, they do not capture how specific audiovisual attributes drive real audience engagement. In this work, we propose a data-driven evaluation framework that uses Vision-Language Models (VLMs) to extract unsupervised audiovisual features, clusters them into interpretable factors, and trains a regression-based evaluator to predict engagement on short-form edutainment videos. Our curated YouTube Shorts dataset enables systematic analysis of how VLM-derived features relate to human engagement behavior. Experiments show strong correlations between predicted and actual engagement, demonstrating that our lightweight, feature-based evaluator provides interpretable and scalable assessments compared to traditional metrics (e.g., SSIM, FID). By grounding evaluation in both multimodal feature importance and human-centered engagement signals, our approach advances toward robust and explainable video understanding.

cs.CV↗

arXiv2025

Evaluating Novelty in AI-Generated Research Plans Using Multi-Workflow LLM Pipelines

The integration of Large Language Models (LLMs) into the scientific ecosystem raises fundamental questions about the creativity and originality of AI-generated research. Recent work has identified ``smart plagiarism'' as a concern in single-step prompting approaches, where models reproduce existing ideas with terminological shifts. This paper investigates whether agentic workflows -- multi-step systems employing iterative reasoning, evolutionary search, and recursive decomposition -- can generate more novel and feasible research plans. We benchmark five reasoning architectures: Reflection-based iterative refinement, Sakana AI v2 evolutionary algorithms, Google Co-Scientist multi-agent framework, GPT Deep Research (GPT-5.1) recursive decomposition, and Gemini~3 Pro multimodal long-context pipeline. Using evaluations from thirty proposals each on novelty, feasibility, and impact, we find that decomposition-based and long-context workflows achieve mean novelty of 4.17/5, while reflection-based approaches score significantly lower (2.33/5). Results reveal varied performance across research domains, with high-performing workflows maintaining feasibility without sacrificing creativity. These findings support the view that carefully designed multi-stage agentic workflows can advance AI-assisted research ideation.

cs.CL↗

arXiv2025

Domain-Partitioned Hybrid RAG for Legal Reasoning: Toward Modular and Explainable Legal AI for India

Legal research in India involves navigating long and heterogeneous documents spanning statutes, constitutional provisions, penal codes, and judicial precedents, where purely keyword-based or embedding-only retrieval systems often fail to support structured legal reasoning. Recent retrieval augmented generation (RAG) approaches improve grounding but struggle with multi-hop reasoning, citation chaining, and cross-domain dependencies inherent to legal texts. We propose a domain partitioned hybrid RAG and Knowledge Graph architecture designed specifically for Indian legal research. The system integrates three specialized RAG pipelines covering Supreme Court case law, statutory and constitutional texts, and the Indian Penal Code, each optimized for domain specific retrieval. To enable relational reasoning beyond semantic similarity, we construct a Neo4j based Legal Knowledge Graph capturing structured relationships among cases, statutes, IPC sections, judges, and citations. An LLM driven agentic orchestrator dynamically routes queries across retrieval modules and the knowledge graph, fusing evidence into grounded and citation aware responses. We evaluate the system using a 40 question synthetic legal question answer benchmark curated from authoritative Indian legal sources and assessed via an LLM as a Judge framework. Results show that the hybrid architecture achieves a 70 percent pass rate, substantially outperforming a RAG only baseline at 37.5 percent, with marked improvements in completeness and legal reasoning quality. These findings demonstrate that combining domain partitioned retrieval with structured relational knowledge provides a scalable and interpretable foundation for advanced legal AI systems in the Indian judicial context.

cs.IR↗

arXiv2025

Investigating Spatial Attention Bias in Vision-Language Models

Vision-Language Models have demonstrated remarkable capabilities in understanding visual content, yet systematic biases in their spatial processing remain largely unexplored. This work identifies and characterizes a systematic spatial attention bias where VLMs consistently prioritize describing left-positioned content before right-positioned content in horizontally concatenated images. Through controlled experiments on image pairs using both open-source and closed-source models, we demonstrate that this bias persists across different architectures, with models describing left-positioned content first in approximately 97% of cases under neutral prompting conditions. Testing on an Arabic-finetuned model reveals that the bias persists despite right-to-left language training, ruling out language reading direction as the primary cause. Investigation of training dataset annotation guidelines from PixMo and Visual Genome reveals no explicit left-first ordering instructions, suggesting the bias is consistent with architectural factors rather than explicit training data instructions. These findings reveal fundamental limitations in how current VLMs process spatial information.

cs.CV↗

arXiv2025

Large Language Models as Pokémon Battle Agents: Strategic Play and Content Generation

Strategic decision-making in Pokémon battles presents a unique testbed for evaluating large language models. Pokémon battles demand reasoning about type matchups, statistical trade-offs, and risk assessment, skills that mirror human strategic thinking. This work examines whether Large Language Models (LLMs) can serve as competent battle agents, capable of both making tactically sound decisions and generating novel, balanced game content. We developed a turn-based Pokémon battle system where LLMs select moves based on battle state rather than pre-programmed logic. The framework captures essential Pokémon mechanics: type effectiveness multipliers, stat-based damage calculations, and multi-Pokémon team management. Through systematic evaluation across multiple model architectures we measured win rates, decision latency, type-alignment accuracy, and token efficiency. These results suggest LLMs can function as dynamic game opponents without domain-specific training, offering a practical alternative to reinforcement learning for turn-based strategic games. The dual capability of tactical reasoning and content creation, positions LLMs as both players and designers, with implications for procedural generation and adaptive difficulty systems in interactive entertainment.

cs.AI↗

arXiv2025

From Facts to Conclusions : Integrating Deductive Reasoning in Retrieval-Augmented LLMs

Retrieval-Augmented Generation (RAG) grounds large language models (LLMs) in external evidence, but fails when retrieved sources conflict or contain outdated or subjective information. Prior work address these issues independently but lack unified reasoning supervision. We propose a reasoning-trace-augmented RAG framework that adds structured, interpretable reasoning across three stages : (1) document-level adjudication, (2) conflict analysis, and (3) grounded synthesis, producing citation-linked answers or justified refusals. A Conflict-Aware Trust-Score (CATS) pipeline is introduced which evaluates groundedness, factual correctness, refusal accuracy, and conflict-behavior alignment using an LLM-as-a-Judge. Our 539-query reasoning dataset and evaluation pipeline establish a foundation for conflict-aware, interpretable RAG systems. Experimental results demonstrate substantial gains over baselines, most notably with Qwen, where Supervised Fine-Tuning improved End-to-End answer correctness from 0.069 to 0.883 and behavioral adherence from 0.074 to 0.722.

cs.CL↗