02 / Research

Research

I'm a PhD researcher in Computer Science at Arizona State University. My core research is on LLM agents — how they're trained, how they reason and act at test time, and how we evaluate and keep them safe. My work spans these stages, with papers at ICLR, ACL, NAACL, EMNLP and NeurIPS workshops.

Core research · LLM agents

The agent lifecycle

My work spans the full stack of building capable, reliable LLM agents — from generating the data that probes them, to training, to test-time methods, to evaluation & safety. Select a stage to explore the papers.

Data Generation

Creating the benchmarks and datasets that probe and stress-test agents.

ActionReasoningBench: Reasoning about Actions with and without Ramification Constraints ICLR 2025 UnSeenTimeQA: Time-Sensitive Question-Answering Beyond LLMs' Memorization ACL 2025

Training

How agents learn — post-training, RL and self-improvement.

Mid-Training with Self-Generated Data Improves Reinforcement Learning in Language Models Preprint ThinkTuning: Instilling Cognitive Reflections without Distillation EMNLP 2025

Test-Time

How agents reason and act at inference — sampling, planning and multi-agent methods.

GuidedSampling: Steering LLMs Towards Diverse Candidate Solutions at Inference-Time ICLR 2026 OptAgent: Optimizing Query Rewriting for E-Commerce via Multi-Agent Simulation Preprint Hypothesis Generation for Materials Discovery and Design Using Goal-Driven and Constraint-Guided LLM Agents Findings of NAACL 2025

Evaluation & Safety

Measuring real capability and surfacing failure modes — jailbreaks and robustness.

BuildBench: Benchmarking LLM Agents on Compiling Real-World Open-Source Software DL4C @ NeurIPS 2025 When “Competency” in Reasoning Opens the Door to Vulnerability: Jailbreaking LLMs via Novel Complex Ciphers NeurIPS 2025 Workshop (Reliable ML from Unreliable Data) Can NLP Models Correctly Reason Over Contexts That Break the Common Assumptions? Preprint

Themes

Research areas

LLM Agents

Multi-agent query rewriting for e-commerce (OptAgent), goal-driven hypothesis generation for materials discovery, and agents that compile real-world open-source software (BuildBench) — spanning how agents are trained, act, and are evaluated.

Reasoning & Planning

Reasoning about actions and change (ActionReasoningBench) and guiding models toward more diverse, higher-quality solutions at inference time (GuidedSampling).

Evaluation & Benchmarks

Benchmarks for temporal reasoning (UnSeenTimeQA), reasoning about actions (ActionReasoningBench) and agentic software building (BuildBench).

Safety & Robustness

Jailbreaking LLMs with novel complex ciphers (ACE / LACE) and studying how models fail when the common assumptions behind a context no longer hold.

Training & Self-Improvement

Teaching models to reflect without distillation (ThinkTuning) and improving RL with self-generated mid-training data.

Output

Recent publications

All publications →

[001]

Mid-Training with Self-Generated Data Improves Reinforcement Learning in Language Models

Aswin RRV, Jacob Dineen, Divij Handa, Mihir Parmar, Ben Zhou, Swaroop Mishra, Chitta Baral

Preprint arXiv

TL;DR Mid-training on self-generated data boosts downstream RL and reasoning.

arXiv ↗ LLMsTraining

Shows that mid-training language models on their own self-generated data improves downstream reinforcement learning, yielding stronger reasoning than RL alone.

[002]

GuidedSampling: Steering LLMs Towards Diverse Candidate Solutions at Inference-Time

Divij Handa, Mihir Parmar, Aswin RRV, Md Nayem Uddin, Hamid Palangi, Chitta Baral

✓ ICLR 2026 Conference

TL;DR Makes inference-time sampling produce genuinely diverse solutions — +21.6% pass@50.

OpenReview ↗ LLMsReasoning

Decouples exploration from generation at inference time so repeated sampling produces genuinely diverse candidate solutions, improving pass@50 by ~21.6% over standard repeated sampling.

[003]

BuildBench: Benchmarking LLM Agents on Compiling Real-World Open-Source Software

Zehua Zhang, A. P. Bajaj, Divij Handa, S. Liu, A. S. Raj, H. Chen, H. Wang, Y. Liu, et al.

✓ DL4C @ NeurIPS 2025 Workshop

TL;DR A realistic benchmark for LLM agents that compile real-world open-source software.

arXiv ↗ AgentsEvaluation

A challenging, realistic benchmark of diverse real-world open-source software for evaluating LLM agents on automatically compiling projects, plus OSS-Build-Agent, a strong baseline with an enhanced build-instruction retrieval module.

[004]

When “Competency” in Reasoning Opens the Door to Vulnerability: Jailbreaking LLMs via Novel Complex Ciphers

Divij Handa, Zehua Zhang, Amir Saeidi, Shrinidhi Kumbhar, Md Nayem Uddin, Aswin RRV, Chitta Baral

✓ NeurIPS 2025 Workshop (Reliable ML from Unreliable Data) Workshop

TL;DR Stronger reasoning makes LLMs easier to jailbreak — via custom ciphers (ACE/LACE).

OpenReview ↗ LLMsSafety

Shows that as LLMs get better at reasoning they become more susceptible to novel jailbreaks. Introduces ACE and LACE — attacks that encode malicious queries with custom and layered ciphers — and CipherBench to measure cipher-decoding ability.

[005]

ThinkTuning: Instilling Cognitive Reflections without Distillation

Aswin RRV, Jacob Dineen, Divij Handa, Md Nayem Uddin, Mihir Parmar, Chitta Baral, Ben Zhou

✓ EMNLP 2025 Conference

TL;DR Teaches models to self-reflect through teacher feedback, without distillation.

arXiv ↗ ReasoningTraining

A GRPO-based interactive training method where a teacher model gives corrective feedback on a student model’s rollouts, instilling self-reflective reasoning without distillation.