Searcharxiv⌕ Search

arXiv subjects

Arjun Radhakrishna

Publications and source records attributed to Arjun Radhakrishna.

At least 19 recordsLinked to original sources

An Empirical Investigation of Robustness in Large Language Models under Tabular Distortions

Distortions are a natural consequence of tables collected from real-world sources, where scraping, extraction, or manual intervention can introduce semantic and structural inconsistencies. Humans can often identify and mentally correct such errors autonomously before reasoning over the data. In this work, we investigate whether Large Language Models (LLMs) possess a similar capability by evaluating them on 506 distorted table question answering (TQA) samples derived from canonical tables. Our findings reveal systematic weaknesses in current LLM pre-training paradigms, where models exhibit a strong inductive bias toward assuming tables are well-formed, often overlooking distortions or discarding inconsistent entries instead of repairing them. Even frontier models like GPT-5.4 and Opus-4.6 achieve only 68% accuracy under distortion. While distortion-aware prompting partially mitigates these failures, errors persist. Through lightweight finetuning, we demonstrate that models can learn to autonomously detect and repair human-recoverable distorted tables.

cs.AI↗

DUET: Co-Evolving Solver and Grader Agents

Agentic workflows are increasingly used across domains such as technology, finance, and enterprise operations. As these agents become more widely deployed, continually improving them becomes increasingly important. This raises an immediate challenge: How should the agent evolve? This evolution requires effective evaluation that can assess outcomes and provide useful feedback for optimization. As the agent evolves, its behaviors and failure modes may also change, making a fixed evaluator increasingly inadequate. Another fundamental question: How should we evaluate an evolving agent? These two challenges are inherently coupled; changes in agent behavior can expose limitations of the current evaluator, while a stronger evaluator provides more informative feedback for improving the agent. Motivated by this interaction, we introduce DUET, a framework that jointly optimizes a solver agent and a grader agent to improve both. DUET iteratively selects training tasks, executes them with the solver, evaluates the resulting outcomes with the grader, and uses a tool-using update module to revise the solver and the grader, alternating between the two across rounds. By updating the grader within the optimization loop, DUET turns evaluation from a fixed source of feedback into a first-class optimization objective that adapts alongside the solver. Experiments across four agent benchmarks show that DUET improves both solver and grader performance and consistently outperforms baselines that optimize the solver with a fixed grader.

cs.LG↗

SkillAxe: Sharpening LLM-Authored Agent Skills Through Evaluation-Guided Self-Refinement

Skill documents, structured natural-language instructions that guide Large Language Model (LLM) agents, are critical to modern agent frameworks, yet LLMs struggle to write skills that actually work. On SkillsBench, human-authored skills improve pass rates by 16.2 percentage points, while LLM-authored skills provide no measurable gain. We introduce SkillAxe, a fully unsupervised framework that enables LLMs to iteratively diagnose and refine their own skills. SkillAxe decomposes skill quality into four interpretable dimensions (quality impact, trigger precision, instruction compliance with fault attribution, and solution-path coverage), producing structured improvement briefs that require no ground-truth labels, test suites, or environment rewards. On SkillsBench, SkillAxe improves pass rates by 28\% relative over unimproved LLM skills and closes 47--67\% of the gap to human-authored skills. We validate the approach as a continuous improvement engine in the wild on SpreadsheetBench, where a SkillAxe-built skill library learns from past agent trajectories and raises pass rate from 16.0\% to 52.0\% using only 22 skills.

cs.MA↗

Scaling Competence, Shrinking Reasoning: Cognitive Signatures in Language Model Learning

We analyze reasoning in language models during task-specific fine-tuning and draws parallel between reasoning tokens--intermediate steps generated while solving problem and the human working memory. Drawing from cognitive science, we align training dynamics with the Four Stages of Competence: models initially produce incorrect outputs without reasoning, then begin reasoning (but still fail), eventually reason effectively, and finally solve tasks without explicit reasoning. We find that reasoning token length expands as performance improves, peaks at the stage of conscious competence, then declines as the model internalizes the task. Notably, after training, models retain performance even when reasoning is removed--suggesting it scaffolded learning but is no longer needed. This progression offers actionable insights: reasoning token dynamics can serve as a signal for diagnosing training stage, identifying convergence, and guiding early stopping. We propose metrics to track this trajectory and argue that reasoning behavior is valuable for understanding and optimizing reasoning model training.

cs.CL↗

STACKFEED: Structured Textual Actor-Critic Knowledge Base Editing with FeedBack

Large Language Models (LLMs) often generate incorrect or outdated information, especially in low-resource settings or when dealing with private data. To address this, Retrieval-Augmented Generation (RAG) uses external knowledge bases (KBs), but these can also suffer from inaccuracies. We introduce STACKFEED, a novel Structured Textual Actor-Critic Knowledge base editing with FEEDback approach that iteratively refines the KB based on expert feedback using a multi-actor, centralized critic reinforcement learning framework. STACKFEED defines a ReACT actor agent on each document to perform structured edits based on document specific targeted instructions. Experimental results showcase that STACKFEED significantly improves KB quality and performance of the RAG system. We evaluate STACKFEED on low-resource programming problems, modified python packaged and factual question-answering tasks.

cs.AI↗

ConDABench: Interactive Evaluation of Language Models for Data Analysis

Real-world data analysis tasks often come with under-specified goals and unclean data. User interaction is necessary to understand and disambiguate a user's intent, and hence, essential to solving these complex tasks. Existing benchmarks for evaluating LLMs on data analysis tasks do not capture these complexities or provide first-class support for interactivity. We introduce ConDABench, a framework for generating conversational data analysis (ConDA) benchmarks and evaluating external tools on the generated benchmarks. \bench consists of (a) a multi-agent workflow for generating realistic benchmarks from articles describing insights gained from public datasets, (b) 1,420 ConDA problems generated using this workflow, and (c) an evaluation harness that, for the first time, makes it possible to systematically evaluate conversational data analysis tools on the generated ConDA problems. Evaluation of state-of-the-art LLMs on the benchmarks reveals that while the new generation of models are better at solving more instances, they are not necessarily better at solving tasks that require sustained, long-form engagement. ConDABench is an avenue for model builders to measure progress towards truly collaborative models that can complete complex interactive tasks.

cs.CL↗

Do Code Models Suffer from the Dunning-Kruger Effect?

As artificial intelligence systems increasingly collaborate with humans in creative and technical domains, questions arise about the cognitive boundaries and biases that shape our shared agency. This paper investigates the Dunning-Kruger Effect (DKE), the tendency for those with limited competence to overestimate their abilities in state-of-the-art LLMs in coding tasks. By analyzing model confidence and performance across a diverse set of programming languages, we reveal that AI models mirror human patterns of overconfidence, especially in unfamiliar or low-resource domains. Our experiments demonstrate that less competent models and those operating in rare programming languages exhibit stronger DKE-like bias, suggesting that the strength of the bias is proportionate to the competence of the models.

cs.AI↗

Collaboration and Conflict between Humans and Language Models through the Lens of Game Theory

Language models are increasingly deployed in interactive online environments, from personal chat assistants to domain-specific agents, raising questions about their cooperative and competitive behavior in multi-party settings. While prior work has examined language model decision-making in isolated or short-term game-theoretic contexts, these studies often neglect long-horizon interactions, human-model collaboration, and the evolution of behavioral patterns over time. In this paper, we investigate the dynamics of language model behavior in the iterated prisoner's dilemma (IPD), a classical framework for studying cooperation and conflict. We pit model-based agents against a suite of 240 well-established classical strategies in an Axelrod-style tournament and find that language models achieve performance on par with, and in some cases exceeding, the best-known classical strategies. Behavioral analysis reveals that language models exhibit key properties associated with strong cooperative strategies - niceness, provocability, and generosity while also demonstrating rapid adaptability to changes in opponent strategy mid-game. In controlled "strategy switch" experiments, language models detect and respond to shifts within only a few rounds, rivaling or surpassing human adaptability. These results provide the first systematic characterization of long-term cooperative behaviors in language model agents, offering a foundation for future research into their role in more complex, mixed human-AI social environments.

cs.AI↗

TableTalk: Scaffolding Spreadsheet Development with a Language Agent

Spreadsheet programming is challenging. Programmers use spreadsheet programming knowledge (e.g., formulas) and problem-solving skills to combine actions into complex tasks. Advancements in large language models have introduced language agents that observe, plan, and perform tasks, showing promise for spreadsheet creation. We present TableTalk, a spreadsheet programming agent embodying three design principles -- scaffolding, flexibility, and incrementality -- derived from studies with seven spreadsheet programmers and 85 Excel templates. TableTalk guides programmers through structured plans based on professional workflows, generating three potential next steps to adapt plans to programmer needs. It uses pre-defined tools to generate spreadsheet components and incrementally build spreadsheets. In a study with 20 programmers, TableTalk produced higher-quality spreadsheets 2.3 times more likely to be preferred than the baseline. It reduced cognitive load and thinking time by 12.6%. From this, we derive design guidelines for agentic spreadsheet programming tools and discuss implications on spreadsheet programming, end-user programming, AI-assisted programming, and human-agent collaboration.

cs.SE↗

TEN: Table Explicitization, Neurosymbolically

We present a neurosymbolic approach, TEN, for extracting tabular data from semistructured input text. This task is particularly challenging for text input that does not use special delimiters consistently to separate columns and rows. Purely neural approaches perform poorly due to hallucinations and their inability to enforce hard constraints. TEN uses Structural Decomposition prompting - a specialized chain-of-thought prompting approach - on a large language model (LLM) to generate an initial table, and thereafter uses a symbolic checker to evaluate not only the well-formedness of that table, but also detect cases of hallucinations or forgetting. The output of the symbolic checker is processed by a critique-LLM to generate guidance for fixing the table, which is presented to the original LLM in a self-debug loop. Our extensive experiments demonstrate that TEN significantly outperforms purely neural baselines across multiple datasets and metrics, achieving significantly higher exact match accuracy and substantially reduced hallucination rates. A 21-participant user study further confirms that TEN's tables are rated significantly more accurate (mean score: 5.0 vs 4.3; p = 0.021), and are consistently preferred for ease of verification and correction, with participants favoring our method in over 60% of the cases.

cs.CL↗

Ordered Semantically Diverse Sampling for Textual Data

The goal of diversity sampling is to select a representative subset of data in a way that maximizes information contained in the subset while keeping its cardinality small. We introduce the ordered diverse sampling problem based on a new metric that measures the diversity in an ordered list of samples. We present a novel approach for generating ordered diverse samples for textual data that uses principal components on the embedding vectors. The proposed approach is simple and compared with existing approaches using the new metric. We transform standard text classification benchmarks into benchmarks for ordered diverse sampling. Our empirical evaluation shows that prevailing approaches perform 6% to 61% worse than our method while also being more time inefficient. Ablation studies show how the parts of the new approach contribute to the overall metrics.

cs.CL↗

LLM-Guided Compositional Program Synthesis

Program synthesis from input-output examples, also called programming by example (PBE), has had tremendous impact on automating end-user tasks. Large language models (LLMs) have the ability to solve PBE tasks by generating code in different target languages, but they can fail unpredictably. To recover for failure, most approaches, such as self-reflection, use the LLM to solve the same task, but with a richer context. We introduce a novel technique that recovers from failure by constructing simpler subtasks for the LLM to solve. Our approach performs compositional program synthesis using LLMs, where LLM not only guides the decomposition of the PBE task into subtasks, but also solves the subtasks. We present different strategies for decomposing the original task. We experimentally show that our approach can solve challenging task instances that are not solved by self-reflection alone.

cs.PL↗

MetaReflection: Learning Instructions for Language Agents using Past Reflections

The popularity of Large Language Models (LLMs) have unleashed a new age ofLanguage Agents for solving a diverse range of tasks. While contemporary frontier LLMs are capable enough to power reasonably good Language agents, the closed-API model makes it hard to improve in cases they perform sub-optimally. To address this, recent works have explored ways to improve their performance using techniques like self-reflection and prompt optimization. Unfortunately, techniques like self-reflection can be used only in an online setup, while contemporary prompt optimization techniques are designed and tested to work on simple tasks. To this end, we introduce MetaReflection, a novel offline reinforcement learning technique that enhances the performance of Language Agents by augmenting a semantic memory based on experiential learnings from past trials. We demonstrate the efficacy of MetaReflection by evaluating across multiple domains, including complex logical reasoning, biomedical semantic similarity, open world question answering, and vulnerability threat detection, in Infrastructure-as-Code, spanning different agent designs. MetaReflection boosts Language agents' performance by 4% to 16.82% over the raw GPT-4 baseline and performs on par with existing state-of-the-art prompt optimization techniques while requiring fewer LLM calls.

cs.CL↗

Exploring Interaction Patterns for Debugging: Enhancing Conversational Capabilities of AI-assistants

The widespread availability of Large Language Models (LLMs) within Integrated Development Environments (IDEs) has led to their speedy adoption. Conversational interactions with LLMs enable programmers to obtain natural language explanations for various software development tasks. However, LLMs often leap to action without sufficient context, giving rise to implicit assumptions and inaccurate responses. Conversations between developers and LLMs are primarily structured as question-answer pairs, where the developer is responsible for asking the the right questions and sustaining conversations across multiple turns. In this paper, we draw inspiration from interaction patterns and conversation analysis -- to design Robin, an enhanced conversational AI-assistant for debugging. Through a within-subjects user study with 12 industry professionals, we find that equipping the LLM to -- (1) leverage the insert expansion interaction pattern, (2) facilitate turn-taking, and (3) utilize debugging workflows -- leads to lowered conversation barriers, effective fault localization, and 5x improvement in bug resolution rates.

cs.HC↗

GrACE: Generation using Associated Code Edits

Developers expend a significant amount of time in editing code for a variety of reasons such as bug fixing or adding new features. Designing effective methods to predict code edits has been an active yet challenging area of research due to the diversity of code edits and the difficulty of capturing the developer intent. In this work, we address these challenges by endowing pre-trained large language models (LLMs) of code with the knowledge of prior, relevant edits. The generative capability of the LLMs helps address the diversity in code changes and conditioning code generation on prior edits helps capture the latent developer intent. We evaluate two well-known LLMs, Codex and CodeT5, in zero-shot and fine-tuning settings respectively. In our experiments with two datasets, the knowledge of prior edits boosts the performance of the LLMs significantly and enables them to generate 29% and 54% more correctly edited code in top-1 suggestions relative to the current state-of-the-art symbolic and neural approaches, respectively.

cs.SE↗

Overwatch: Learning Patterns in Code Edit Sequences

Integrated Development Environments (IDEs) provide tool support to automate many source code editing tasks. Traditionally, IDEs use only the spatial context, i.e., the location where the developer is editing, to generate candidate edit recommendations. However, spatial context alone is often not sufficient to confidently predict the developer's next edit, and thus IDEs generate many suggestions at a location. Therefore, IDEs generally do not actively offer suggestions and instead, the developer is usually required to click on a specific icon or menu and then select from a large list of potential suggestions. As a consequence, developers often miss the opportunity to use the tool support because they are not aware it exists or forget to use it. To better understand common patterns in developer behavior and produce better edit recommendations, we can additionally use the temporal context, i.e., the edits that a developer was recently performing. To enable edit recommendations based on temporal context, we present Overwatch, a novel technique for learning edit sequence patterns from traces of developers' edits performed in an IDE. Our experiments show that Overwatch has 78% precision and that Overwatch not only completed edits when developers missed the opportunity to use the IDE tool support but also predicted new edits that have no tool support in the IDE.

cs.PL↗

AutoTSG: Learning and Synthesis for Incident Troubleshooting

Incident management is a key aspect of operating large-scale cloud services. To aid with faster and efficient resolution of incidents, engineering teams document frequent troubleshooting steps in the form of Troubleshooting Guides (TSGs), to be used by on-call engineers (OCEs). However, TSGs are siloed, unstructured, and often incomplete, requiring developers to manually understand and execute necessary steps. This results in a plethora of issues such as on-call fatigue, reduced productivity, and human errors. In this work, we conduct a large-scale empirical study of over 4K+ TSGs mapped to 1000s of incidents and find that TSGs are widely used and help significantly reduce mitigation efforts. We then analyze feedback on TSGs provided by 400+ OCEs and propose a taxonomy of issues that highlights significant gaps in TSG quality. To alleviate these gaps, we investigate the automation of TSGs and propose AutoTSG -- a novel framework for automation of TSGs to executable workflows by combining machine learning and program synthesis. Our evaluation of AutoTSG on 50 TSGs shows the effectiveness in both identifying TSG statements (accuracy 0.89) and parsing them for execution (precision 0.94 and recall 0.91). Lastly, we survey ten Microsoft engineers and show the importance of TSG automation and the usefulness of AutoTSG.

cs.SE↗

Landmarks and Regions: A Robust Approach to Data Extraction

We propose a new approach to extracting data items or field values from semi-structured documents. Examples of such problems include extracting passenger name, departure time and departure airport from a travel itinerary, or extracting price of an item from a purchase receipt. Traditional approaches to data extraction use machine learning or program synthesis to process the whole document to extract the desired fields. Such approaches are not robust to format changes in the document, and the extraction process typically fails even if changes are made to parts of the document that are unrelated to the desired fields of interest. We propose a new approach to data extraction based on the concepts of landmarks and regions. Humans routinely use landmarks in manual processing of documents to zoom in and focus their attention on small regions of interest in the document. Inspired by this human intuition, we use the notion of landmarks in program synthesis to automatically synthesize extraction programs that first extract a small region of interest, and then automatically extract the desired value from the region in a subsequent step. We have implemented our landmark-based extraction approach in a tool LRSyn, and show extensive evaluation on documents in HTML as well as scanned images of invoices and receipts. Our results show that our approach is robust to various types of format changes that routinely happen in real-world settings.

cs.SE↗