arXiv subjects

Vinayak Arannil

Publications and source records attributed to Vinayak Arannil.

7 recordsLinked to original sources

arXiv2026

From Self-Distillation to Self-Practice: Privileged Information for Multi-Turn Agents

On-policy self-distillation (OPSD) has become a popular recipe for post-training LLM agents. It supervises the agent model at the token level with a stronger teacher view of the same model, obtained by conditioning on privileged information (PI). In this work, we show that in multi-turn agents, this paradigm teaches the student to act with confidence but without the information behind it. The trained agent behaves as if it had privileged information it never observed, and its performance falls well short of plain RL, in the worst case below the untrained base model. Therefore, we propose Privileged Self-Practice (PSP), which keeps the PI and moves it from the loss to the sampler. When the student's rollouts on a task mostly fail, we inject a short per-task instruction written by an analyzer model, sample the task again with the instruction in context, and train on the result with an unchanged GRPO objective. The privileged information stays in the prompt and never enters the loss. Across AppWorld and SWE-bench Verified, with three different student models, PSP obtains the best average score in every setting and is the only method that consistently outperforms plain GRPO, improving task-goal completion by up to 65% on AppWorld and the resolved rate by up to 61% on SWE-bench Verified.

cs.AI↗

arXiv2026

An Empirical Study of Automating Agent Evaluation

Agent evaluation requires assessing complex multi-step behaviors involving tool use and intermediate reasoning, making it costly and expertise-intensive. A natural question arises: can frontier coding assistants reliably automate this evaluation process? Our study shows that simply prompting coding assistants is insufficient for this task. Without domain-specific evaluation knowledge, frontier coding assistants achieve only a 30% execution success rate and produce over-engineered evaluations averaging 12+ metrics per agent, indicating that strong coding ability does not automatically translate to reliable agent evaluation. We introduce EvalAgent, an AI assistant that automates the end-to-end agent evaluation pipeline. EvalAgent encodes evaluation domain expertise as evaluation skills (procedural instructions, reusable code and templates, and dynamically retrieved API documentation) that compose into a trace-based pipeline producing complete evaluation artifacts including metrics, executable code, and reports. To systematically assess generated evaluations, we introduce a meta-evaluation framework alongside AgentEvalBench, a benchmark comprising 20 agents, each paired with evaluation requirements and test scenarios. We further propose the Eval@1 metric to measure whether generated evaluation code both executes and yields meaningful results on the first run. Our experiments show that EvalAgent produces focused evaluations, improving Eval@1 from 17.5% to 65%, and achieving 79.5% human expert preference over baseline approaches. Further ablation studies show that evaluation skills are critical for handling complex evaluation: removing them causes Eval@1 to drop significantly from 65% to 30%.

cs.CL↗

arXiv2026

PERSONAJUDGE: Simulating Individual Human Preference Judgments with Evaluator-Specific Demonstration Data

Large language models increasingly serve as judges in AI evaluation, but current approaches rely on consensus preferences that ignore individual evaluator variation. We propose a novel simulation approach that combines categorical judgments with evaluator-specific auxiliary data--retrospective reasoning traces and interface telemetry--to enable LLM-based simulation of individual evaluators via in-context learning. We conduct a systematic empirical study of this approach using multi-facet data from 32 trained annotators across 4,200 preference judgments in a 4 x 4 x 4 factorial design. Our key findings: (1) The simulation approach achieves up to 9.9 percentage point improvements over the Base Judge; (2) Reasoning traces provide the largest gains with higher collection efforts, while interface telemetry often hurts rather than helps performance despite being cheaper to collect. (3) Simulation difficulty is systematic, predicted by an evaluator's neutral usage (most clearly on Helpfulness) and divergence from consensus; the neutral-usage tendency--rather than simulatability itself--is the cross-task-stable property (r = 0.728). These results establish both the potential and limits of evaluator-specific auxiliary data for personalized evaluation, offering methodological insights for scaling individual aware AI assessment.

cs.HC↗

arXiv2025

PILOT: Steering Synthetic Data Generation with Psychological & Linguistic Output Targeting

Generative AI applications commonly leverage user personas as a steering mechanism for synthetic data generation, but reliance on natural language representations forces models to make unintended inferences about which attributes to emphasize, limiting precise control over outputs. We introduce PILOT (Psychological and Linguistic Output Targeting), a two-phase framework for steering large language models with structured psycholinguistic profiles. In Phase 1, PILOT translates natural language persona descriptions into multidimensional profiles with normalized scores across linguistic and psychological dimensions. In Phase 2, these profiles guide generation along measurable axes of variation. We evaluate PILOT across three state-of-the-art LLMs (Mistral Large 2, Deepseek-R1, LLaMA 3.3 70B) using 25 synthetic personas under three conditions: Natural-language Persona Steering (NPS), Schema-Based Steering (SBS), and Hybrid Persona-Schema Steering (HPS). Results demonstrate that schema-based approaches significantly reduce artificial-sounding persona repetition while improving output coherence, with silhouette scores increasing from 0.098 to 0.237 and topic purity from 0.773 to 0.957. Our analysis reveals a fundamental trade-off: SBS produces more concise outputs with higher topical consistency, while NPS offers greater lexical diversity but reduced predictability. HPS achieves a balance between these extremes, maintaining output variety while preserving structural consistency. Expert linguistic evaluation confirms that PILOT maintains high response quality across all conditions, with no statistically significant differences between steering approaches.

cs.CL↗

arXiv2025

MetaSynth: Meta-Prompting-Driven Agentic Scaffolds for Diverse Synthetic Data Generation

Recent smaller language models such Phi-3.5 and Phi-4 rely on synthetic data generated using larger Language models. Questions remain about leveraging synthetic data for other use cases, such as adapting LLMs to specific domains. A key limitation of synthetic data is low diversity, which negatively impacts its downstream applicability for improving other models. To address this, we propose MetaSynth, a method for generating synthetic data that enhances diversity through meta-prompting, where a language model orchestrates multiple "expert" LLM agents to collaboratively generate data. Using only 25 million tokens of synthetic data generated with MetaSynth, we successfully adapt a well-trained LLM (Mistral-7B-v0.3) to two specialized domains-Finance and Biomedicine-without compromising the capabilities of the resulting model in general tasks. In addition, we evaluate the diversity of our synthetic data using seven automated metrics, and find that it approaches the diversity of LLM pre-training corpora. Continually pre-training Mistral-7B-v0.3 with MetaSynth notably outperforms the base LLM, showing improvements of up to 4.08% in Finance and 13.75% in Biomedicine. The same model shows degraded performance when trained on data generated using a template prompt, even when the template includes prior generations and varying In-Context exemplars of real data. Our findings suggest that a few million tokens of diverse synthetic data without mixing any real data, is sufficient for effective domain adaptation when using MetaSynth.

cs.CL↗

arXiv2024

ADEQA: A Question Answer based approach for joint ADE-Suspect Extraction using Sequence-To-Sequence Transformers

Early identification of Adverse Drug Events (ADE) is critical for taking prompt actions while introducing new drugs into the market. These ADEs information are available through various unstructured data sources like clinical study reports, patient health records, social media posts, etc. Extracting ADEs and the related suspect drugs using machine learning is a challenging task due to the complex linguistic relations between drug ADE pairs in textual data and unavailability of large corpus of labelled datasets. This paper introduces ADEQA, a question-answer(QA) based approach using quasi supervised labelled data and sequence-to-sequence transformers to extract ADEs, drug suspects and the relationships between them. Unlike traditional QA models, natural language generation (NLG) based models don't require extensive token level labelling and thereby reduces the adoption barrier significantly. On a public ADE corpus, we were able to achieve state-of-the-art results with an F1 score of 94% on establishing the relationships between ADEs and the respective suspects.

cs.CL↗

arXiv2024

DoPAMine: Domain-specific Pre-training Adaptation from seed-guided data Mining

Large Language Models (LLMs) have shown remarkable ability to generalize effectively across numerous industry domains while executing a range of tasks. Many of these competencies are obtained from the data utilized during the pre-training phase of the Language Models (LMs). However, these models exhibit limitations when tasked with performing in specialized or low-resource industry domains. More recent approaches use LLMs for generating domain-specific synthetic data but most often they lack in truthfulness and complexity. Alternatively, in cases where domain data is available like healthcare and finance most of the LMs are proprietary necessitating the need for a scalable method to curate real world industry specific pre-training data. In this work, we propose an automated and scalable framework - DoPAMine:Domain-specific Pre-training Adaptation from seed-guided data Mining, to mine domain specific training data from a large data corpus for domain adaptation of a LM. The framework leverages the parametric knowledge of a LLM to generate diverse and representative seed data tailored to a specific domain which is then used to mine real world data from a large data corpus like Common Crawl. We evaluated our framework's performance in the continual pre-training (CPT) setting by training two domain specific 7B parameter LMs in healthcare and finance with data mined via DoPAMine. Our experiments show that DoPAMine boosts the performance of pre-trained LLMs on average by 4.9% and 5.1% in zero-shot and 5-shot settings respectively on healthcare tasks from MMLU, MedQA, MedMCQA and PubMedQA datasets, and 2.9% and 6.7% for zero-shot and 5-shot settings respectively on finance tasks from FiQA-SA, FPB and Headlines datasets when compared to the baseline.

cs.CL↗