arXiv subjects

Dhruv Kumar

Publications and source records attributed to Dhruv Kumar.

At least 73 records · Page 4Linked to original sources

arXiv2025

Evaluating LLMs for Zeolite Synthesis Event Extraction (ZSEE): A Systematic Analysis of Prompting Strategies

Extracting structured information from zeolite synthesis experimental procedures is critical for materials discovery, yet existing methods have not systematically evaluated Large Language Models (LLMs) for this domain-specific task. This work addresses a fundamental question: what is the efficacy of different prompting strategies when applying LLMs to scientific information extraction? We focus on four key subtasks: event type classification (identifying synthesis steps), trigger text identification (locating event mentions), argument role extraction (recognizing parameter types), and argument text extraction (extracting parameter values). We evaluate four prompting strategies - zero-shot, few-shot, event-specific, and reflection-based - across six state-of-the-art LLMs (Gemma-3-12b-it, GPT-5-mini, O4-mini, Claude-Haiku-3.5, DeepSeek reasoning and non-reasoning) using the ZSEE dataset of 1,530 annotated sentences. Results demonstrate strong performance on event type classification (80-90\% F1) but modest performance on fine-grained extraction tasks, particularly argument role and argument text extraction (50-65\% F1). GPT-5-mini exhibits extreme prompt sensitivity with 11-79\% F1 variation. Notably, advanced prompting strategies provide minimal improvements over zero-shot approaches, revealing fundamental architectural limitations. Error analysis identifies systematic hallucination, over-generalization, and inability to capture synthesis-specific nuances. Our findings demonstrate that while LLMs achieve high-level understanding, precise extraction of experimental parameters requires domain-adapted models, providing quantitative benchmarks for scientific information extraction.

cs.CL↗

arXiv2025

Lights, Camera, Consistency: A Multistage Pipeline for Character-Stable AI Video Stories

Generating long, cohesive video stories with consistent characters is a significant challenge for current text-to-video AI. We introduce a method that approaches video generation in a filmmaker-like manner. Instead of creating a video in one step, our proposed pipeline first uses a large language model to generate a detailed production script. This script guides a text-to-image model in creating consistent visuals for each character, which then serve as anchors for a video generation model to synthesize each scene individually. Our baseline comparisons validate the necessity of this multi-stage decomposition; specifically, we observe that removing the visual anchoring mechanism results in a catastrophic drop in character consistency scores (from 7.99 to 0.55), confirming that visual priors are essential for identity preservation. Furthermore, we analyze cultural disparities in current models, revealing distinct biases in subject consistency and dynamic degree between Indian vs Western-themed generations.

cs.CV↗

arXiv2025

Confidence is Not Competence

Large language models (LLMs) often exhibit a puzzling disconnect between their asserted confidence and actual problem-solving competence. We offer a mechanistic account of this decoupling by analyzing the geometry of internal states across two phases - pre-generative assessment and solution execution. A simple linear probe decodes the internal "solvability belief" of a model, revealing a well-ordered belief axis that generalizes across model families and across math, code, planning, and logic tasks. Yet, the geometries diverge - although belief is linearly decodable, the assessment manifold has high linear effective dimensionality as measured from the principal components, while the subsequent reasoning trace evolves on a much lower-dimensional manifold. This sharp reduction in geometric complexity from thought to action mechanistically explains the confidence-competence gap. Causal interventions that steer representations along the belief axis leave final solutions unchanged, indicating that linear nudges in the complex assessment space do not control the constrained dynamics of execution. We thus uncover a two-system architecture - a geometrically complex assessor feeding a geometrically simple executor. These results challenge the assumption that decodable beliefs are actionable levers, instead arguing for interventions that target the procedural dynamics of execution rather than the high-level geometry of assessment.

cs.CL↗

arXiv2025

Policy Optimization Prefers The Path of Least Resistance

Policy optimization (PO) algorithms are used to refine Large Language Models for complex, multi-step reasoning. Current state-of-the-art pipelines enforce a strict think-then-answer format to elicit chain-of-thought (CoT); however, the behavior of PO when these rigid constraints are relaxed into an open-ended CoT structure remains an under-studied question. We investigate this gap with an extensive suite of controlled experiments and identify a consistent principle: \textit{policy optimization consistently follows the path of least resistance}. When afforded the flexibility to interleave reasoning and response, policy optimization consistently learns to discard explicit reasoning, causing the policy to degenerate to a direct \texttt{ }-only format. This outcome holds true across various models and algorithms. We find that this collapse in format is persistent even when the complex \texttt{ } format is assigned up to 4x larger reward weights. We formalize this principle through a series of controlled reward decomposition experiments, demonstrating a clear hierarchy: PO systematically optimizes for the simplest reward component first, a preference that holds even when faced with mutually exclusive choices or strong incentives for more complex behaviors. Finally, we show that successful convergence on the high-reward shortcut is not a low-effort drift but is driven by the optimization process that requires the KL-regularized policy to have sufficient freedom to make a significant shift from its initial prior. Our findings reveal that granting policies the freedom to diverge is a double-edged sword: while necessary for discovering high-reward shortcuts, it also creates a powerful incentive to game the simplest aspects of the reward function, posing a critical challenge for reward hacking under alignment.

cs.CL↗

arXiv2025

time2time: Causal Intervention in Hidden States to Simulate Rare Events in Time Series Foundation Models

While transformer-based foundation models excel at forecasting routine patterns, two questions remain: do they internalize semantic concepts such as market regimes, or merely fit curves? And can their internal representations be leveraged to simulate rare, high-stakes events such as market crashes? To investigate this, we introduce activation transplantation, a causal intervention that manipulates hidden states by imposing the statistical moments of one event (e.g., a historical crash) onto another (e.g., a calm period) during the forward pass. This procedure deterministically steers forecasts: injecting crash semantics induces downturn predictions, while injecting calm semantics suppresses crashes and restores stability. Beyond binary control, we find that models encode a graded notion of event severity, with the latent vector norm directly correlating with the magnitude of systemic shocks. Validated across two architecturally distinct TSFMs, Toto (decoder only) and Chronos (encoder-decoder), our results demonstrate that steerable, semantically grounded representations are a robust property of large time series transformers. Our findings provide evidence for a latent concept space that governs model predictions, shifting interpretability from post-hoc attribution to direct causal intervention, and enabling semantic "what-if" analysis for strategic stress-testing.

cs.LG↗

arXiv2025

Talking with Oompa Loompas: A novel framework for evaluating linguistic acquisition of LLM agents

Existing evaluation studies on linguistic competence of large language models (LLM agents) have focused primarily on vocabulary learning, morphological rule induction, syntactic generalization, pragmatic inference, and cross-linguistic transfer. However, none assess whether LLM agents can acquire a language through pattern recognition and interactive feedback, a central feature of human language acquisition. We propose a novel experimental framework in which an LLM agent is evaluated on its ability to acquire and use a newly constructed language (Tinkatongue) in conversation with a bot that understands only Tinkatongue. Our findings show that LLM agents fail to establish a conversation within 100 responses, yet they adopt distinct strategies that mirror human approaches to language learning. The results suggest a new direction for evaluation benchmarks and open pathways to model designs that learn more effectively from interactive feedback.

cs.CL↗

arXiv2025

HAEPO: History-Aggregated Exploratory Policy Optimization

Exploration is essential in modern learning, from reinforcement learning environments with small neural policies to large language models (LLMs). Existing work, such as DPO, leverages full sequence log-likelihoods to capture an entire trajectory of the model's decisions, while methods like GRPO aggregate per-token ratios into a trajectory-level update. However, both often limit exploration on long-horizon tasks. We introduce History-Aggregated Exploratory Policy Optimization (HAEPO), a history-aware exploratory loss to combat these shortcomings. HAEPO compresses each trajectory into the sum of its logarithmic probabilities (a cumulative logarithmic likelihood), and applies a Plackett-Luce softmax across trajectories to obtain normalized weights proportional to their returns, thus encouraging broader exploration. We add entropy regularization to stabilize the aggressive updates to prevent premature collapse and a soft KL penalty relative to a frozen copy of the previous (reference) policy. Empirically, HAEPO converges fast, explores thoroughly, aligns closely with true rewards, and demonstrates robust learning behavior better or at par with PPO, GRPO, and DPO across diverse tasks. Thus, HAEPO provides a stable and interpretable framework by explicitly leveraging full-trajectory history while balancing exploration and stability.

cs.LG↗

arXiv2025

Disentangling Polysemantic Neurons with a Null-Calibrated Polysemanticity Index and Causal Patch Interventions

Neural networks often contain polysemantic neurons that respond to multiple, sometimes unrelated, features, complicating mechanistic interpretability. We introduce the Polysemanticity Index (PSI), a null-calibrated metric that quantifies when a neuron's top activations decompose into semantically distinct clusters. PSI multiplies three independently calibrated components: geometric cluster quality (S), alignment to labeled categories (Q), and open-vocabulary semantic distinctness via CLIP (D). On a pretrained ResNet-50 evaluated with Tiny-ImageNet images, PSI identifies neurons whose activation sets split into coherent, nameable prototypes, and reveals strong depth trends: later layers exhibit substantially higher PSI than earlier layers. We validate our approach with robustness checks (varying hyperparameters, random seeds, and cross-encoder text heads), breadth analyses (comparing class-only vs. open-vocabulary concepts), and causal patch-swap interventions. In particular, aligned patch replacements increase target-neuron activation significantly more than non-aligned, random, shuffled-position, or ablate-elsewhere controls. PSI thus offers a principled and practical lever for discovering, quantifying, and studying polysemantic units in neural networks.

cs.LG↗

arXiv2025

Rubric Is All You Need: Enhancing LLM-based Code Evaluation With Question-Specific Rubrics

Since the emergence of Large Language Models (LLMs) popularized by the release of GPT-3 and ChatGPT, LLMs have shown remarkable promise in programming-related tasks. While code generation using LLMs has become a popular field of research, code evaluation using LLMs remains under-explored. In this paper, we focus on LLM-based code evaluation and attempt to fill in the existing gaps. We propose multi-agentic novel approaches using \emph{question-specific rubrics} tailored to the problem statement, arguing that these perform better for logical assessment than the existing approaches that use \emph{question-agnostic rubrics}. To address the lack of suitable evaluation datasets, we introduce two datasets: a Data Structures and Algorithms dataset containing 150 student submissions from a popular Data Structures and Algorithms practice website, and an Object Oriented Programming dataset comprising 80 student submissions from undergraduate computer science courses. In addition to using standard metrics (Spearman Correlation, Cohen's Kappa), we additionally propose a new metric called as Leniency, which quantifies evaluation strictness relative to expert assessment. Our comprehensive analysis demonstrates that \emph{question-specific rubrics} significantly enhance logical assessment of code in educational settings, providing better feedback aligned with instructional goals beyond mere syntactic correctness.

cs.SE↗

arXiv2025

Loop Invariant Generation: A Hybrid Framework of Reasoning optimised LLMs and SMT Solvers

Loop invariants are essential for proving the correctness of programs with loops. Developing loop invariants is challenging, and fully automatic synthesis cannot be guaranteed for arbitrary programs. Some approaches have been proposed to synthesize loop invariants using symbolic techniques and more recently using neural approaches. These approaches are able to correctly synthesize loop invariants only for subsets of standard benchmarks. In this work, we investigate whether modern, reasoning-optimized large language models can do better. We integrate OpenAI's O1, O1-mini, and O3-mini into a tightly coupled generate-and-check pipeline with the Z3 SMT solver, using solver counterexamples to iteratively guide invariant refinement. We use Code2Inv benchmark, which provides C programs along with their formal preconditions and postconditions. On this benchmark of 133 tasks, our framework achieves 100% coverage (133 out of 133), outperforming the previous best of 107 out of 133, while requiring only 1-2 model proposals per instance and 14-55 seconds of wall-clock time. These results demonstrate that LLMs possess latent logical reasoning capabilities which can help automate loop invariant synthesis. While our experiments target C-specific programs, this approach should be generalizable to other imperative languages.

cs.LO↗

arXiv2025

Automated Type Annotation in Python Using Large Language Models

Type annotations in Python enhance maintainability and error detection. However, generating these annotations manually is error prone and requires extra effort. Traditional automation approaches like static analysis, machine learning, and deep learning struggle with limited type vocabularies, behavioral over approximation, and reliance on large labeled datasets. In this work, we explore the use of LLMs for generating type annotations in Python. We develop a generate check repair pipeline: the LLM proposes annotations guided by a Concrete Syntax Tree representation, a static type checker (Mypy) verifies them, and any errors are fed back for iterative refinement. We evaluate four LLM variants: GPT 4oMini, GPT 4.1mini (general-purpose), and O3Mini, O4Mini (reasoning optimized), on 6000 code snippets from the ManyTypes4Py benchmark. We first measure the proportion of code snippets annotated by LLMs for which MyPy reported no errors (i.e., consistent results): GPT 4oMini achieved consistency on 65.9% of cases (34.1% inconsistent), while GPT 4.1mini, O3Mini, and O4Mini each reached approximately 88.6% consistency (around 11.4% failures). To measure annotation quality, we then compute exact-match and base-type match accuracies over all 6000 snippets: GPT 4.1mini and O3Mini perform the best, achieving up to 70.5% exact match and 79.1% base type accuracy, requiring under one repair iteration on average. Our results demonstrate that general-purpose and reasoning optimized LLMs, without any task specific fine tuning or additional training can be effective in generating consistent type annotations.They perform competitively with traditional deep learning techniques which require large labeled dataset for training. While our work focuses on Python, the pipeline can be extended to other optionally typed imperative languages like Ruby

cs.PL↗

arXiv2025

The Indian Pulsar Timing Array Data Release 2: I. Dataset and Timing Analysis

The Indian Pulsar Timing Array (InPTA) employs unique features of the upgraded Giant Metrewave Radio Telescope (uGMRT) to monitor dozens of the International Pulsar Timing Array (IPTA) millisecond pulsars (MSPs), simultaneously in the 300-500 MHz and the 1260-1460 MHz bands. This dual-band approach ensures that any frequency-dependent delays are accurately characterized, significantly improving the timing precision for pulsar observations, which is crucial for pulsar timing arrays. We present details of InPTA's second data release that involves 7 yrs of data on 27 IPTA MSPs. This includes sub-banded Times of Arrival (ToAs), Dispersion Measures (DM), and initial timing ephemerides for our MSPs. A part of this dataset, originally released in InPTA's first data release, is being incorporated into IPTA's third data release which is expected to detect and characterize nanohertz gravitational waves in the coming years. The entire dataset is reprocessed in this second data release providing some of the highest precision DM estimates so far and interesting solar wind related DM variations in some pulsars. This is likely to characterize the noise introduced by the dynamic inter-stellar ionised medium much better than the previous release thereby increasing sensitivity to any future gravitational wave search.

astro-ph.IM↗

arXiv2025

Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval Augmented Generation Across Learning Style

Effective teaching requires adapting instructional strategies to accommodate the diverse cognitive and behavioral profiles of students, a persistent challenge in education and teacher training. While Large Language Models (LLMs) offer promise as tools to simulate such complex pedagogical environments, current simulation frameworks are limited in two key respects: (1) they often reduce students to static knowledge profiles, and (2) they lack adaptive mechanisms for modeling teachers who evolve their strategies in response to student feedback. To address these gaps, \textbf{we introduce a novel simulation framework that integrates LLM-based heterogeneous student agents with a self-optimizing teacher agent}. The teacher agent's pedagogical policy is dynamically evolved using a genetic algorithm, allowing it to discover and refine effective teaching strategies based on the aggregate performance of diverse learners. In addition, \textbf{we propose Persona-RAG}, a Retrieval Augmented Generation module that enables student agents to retrieve knowledge tailored to their individual learning styles. Persona-RAG preserves the retrieval accuracy of standard RAG baselines while enhancing personalization, an essential factor in modeling realistic educational scenarios. Through extensive experiments, we demonstrate how our framework supports the emergence of distinct and interpretable teaching patterns when interacting with varied student populations. Our results highlight the potential of LLM-driven simulations to inform adaptive teaching practices and provide a testbed for training human educators in controlled, data-driven environments.

cs.AI↗

arXiv2025

ReviewEval: An Evaluation Framework for AI-Generated Reviews

The escalating volume of academic research, coupled with a shortage of qualified reviewers, necessitates innovative approaches to peer review. In this work, we propose: 1. ReviewEval, a comprehensive evaluation framework for AI-generated reviews that measures alignment with human assessments, verifies factual accuracy, assesses analytical depth, identifies degree of constructiveness and adherence to reviewer guidelines; and 2. ReviewAgent, an LLM-based review generation agent featuring a novel alignment mechanism to tailor feedback to target conferences and journals, along with a self-refinement loop that iteratively optimizes its intermediate outputs and an external improvement loop using ReviewEval to improve upon the final reviews. ReviewAgent improves actionable insights by 6.78% and 47.62% over existing AI baselines and expert reviews respectively. Further, it boosts analytical depth by 3.97% and 12.73%, enhances adherence to guidelines by 10.11% and 47.26% respectively. This paper establishes essential metrics for AIbased peer review and substantially enhances the reliability and impact of AI-generated reviews in academic research.

cs.CL↗

arXiv2025

The Impact of Large Language Models on K-12 Education in Rural India: A Thematic Analysis of Student Volunteer's Perspectives

AI-driven education, particularly Large Language Models (LLMs), has the potential to address learning disparities in rural K-12 schools. However, research on AI adoption in rural India remains limited, with existing studies focusing primarily on urban settings. This study examines the perceptions of volunteer teachers on AI integration in rural education, identifying key challenges and opportunities. Through semi-structured interviews with 23 volunteer educators in Rajasthan and Delhi, we conducted a thematic analysis to explore infrastructure constraints, teacher preparedness, and digital literacy gaps. Findings indicate that while LLMs could enhance personalized learning and reduce teacher workload, barriers such as poor connectivity, lack of AI training, and parental skepticism hinder adoption. Despite concerns over over-reliance and ethical risks, volunteers emphasize that AI should be seen as a complementary tool rather than a replacement for traditional teaching. Given the potential benefits, LLM-based tutors merit further exploration in rural classrooms, with structured implementation and localized adaptations to ensure accessibility and equity.

cs.CY↗

arXiv2025

Sakshm AI: Advancing AI-Assisted Coding Education for Engineering Students in India Through Socratic Tutoring and Comprehensive Feedback

The advent of Large Language Models (LLMs) is reshaping education, particularly in programming, by enhancing problem-solving, enabling personalized feedback, and supporting adaptive learning. Existing AI tools for programming education struggle with key challenges, including the lack of Socratic guidance, direct code generation, limited context retention, minimal adaptive feedback, and the need for prompt engineering. To address these challenges, we introduce Sakshm AI, an intelligent tutoring system for learners across all education levels. It fosters Socratic learning through Disha, its inbuilt AI chatbot, which provides context-aware hints, structured feedback, and adaptive guidance while maintaining conversational memory and supporting language flexibility. This study examines 1170 registered participants, analyzing platform logs, engagement trends, and problem-solving behavior to assess Sakshm AI's impact. Additionally, a structured survey with 45 active users and 25 in-depth interviews was conducted, using thematic encoding to extract qualitative insights. Our findings reveal how AI-driven Socratic guidance influences problem-solving behaviors and engagement, offering key recommendations for optimizing AI-based coding platforms. This research combines quantitative and qualitative insights to inform AI-assisted education, providing a framework for scalable, intelligent tutoring systems that improve learning outcomes. Furthermore, Sakshm AI represents a significant step toward Sustainable Development Goal 4 Quality Education, providing an accessible and structured learning tool for undergraduate students, even without expert guidance. This is one of the first large-scale studies examining AI-assisted programming education across multiple institutions and demographics.

cs.HC↗

arXiv2025

Roamify: Designing and Evaluating an LLM Based Google Chrome Extension for Personalised Itinerary Planning

In this paper, we present Roamify, an Artificial Intelligence powered travel assistant that aims to ease the process of travel planning. We have tested and used multiple Large Language Models like Llama and T5 to generate personalised itineraries per user preferences. Results from user surveys highlight the preference for AI powered mediums over existing methods to help in travel planning across all user age groups. These results firmly validate the potential need of such a travel assistant. We highlight the two primary design considerations for travel assistance: D1) incorporating a web-scraping method to gather up-to-date news articles about destinations from various blog sources, which significantly improves our itinerary suggestions, and D2) utilising user preferences to create customised travel experiences along with a recommendation system which changes the itinerary according to the user needs. Our findings suggest that Roamify has the potential to improve and simplify how users across multiple age groups plan their travel experiences.

cs.HC↗

arXiv2025

Multi-Agent Inverse Q-Learning from Demonstrations

When reward functions are hand-designed, deep reinforcement learning algorithms often suffer from reward misspecification, causing them to learn suboptimal policies in terms of the intended task objectives. In the single-agent case, inverse reinforcement learning (IRL) techniques attempt to address this issue by inferring the reward function from expert demonstrations. However, in multi-agent problems, misalignment between the learned and true objectives is exacerbated due to increased environment non-stationarity and variance that scales with multiple agents. As such, in multi-agent general-sum games, multi-agent IRL algorithms have difficulty balancing cooperative and competitive objectives. To address these issues, we propose Multi-Agent Marginal Q-Learning from Demonstrations (MAMQL), a novel sample-efficient framework for multi-agent IRL. For each agent, MAMQL learns a critic marginalized over the other agents' policies, allowing for a well-motivated use of Boltzmann policies in the multi-agent context. We identify a connection between optimal marginalized critics and single-agent soft-Q IRL, allowing us to apply a direct, simple optimization criterion from the single-agent domain. Across our experiments on three different simulated domains, MAMQL significantly outperforms previous multi-agent methods in average reward, sample efficiency, and reward recovery by often more than 2-5x. We make our code available at https://sites.google.com/view/mamql .

cs.MA↗