Research direction

LLM Reasoning and Agents

Language-model systems for reasoning, prompting, evaluation, multi-agent interaction, and specialized domain workflows.

LLM reasoning Agents Evaluation Domain adaptation

Focus

This direction studies how language models can reason more reliably, interact in agentic systems, and operate in domains where correctness, calibration, and domain knowledge matter.

Typical Questions

  • How can prompting and agent protocols improve reasoning reliability?
  • How should multi-agent systems be evaluated when outputs are strategic or uncertain?
  • How can models be adapted to specialized technical domains without losing robustness?
  • What measurements reveal failure modes before deployment?
Selected papers

LLM Highlights

3 papers
Two reasoning models cross-teaching successful hints to rescue failed attempts and reward diverse correct solutions
2026 · Collaborative Reasoning

CoRe: Collaborative Reasoning via Cross Teaching

  • Turns peer success into a training signal through a cross-teaching protocol.
  • Balances correctness, diversity, and rescue from failed attempts in the reward.
  • Shows large gains from complementarity without increasing model size.
Publication list PDF
Full fine-tuning matrix dynamics projected into low-rank adaptation and optimizer moments
2026 · Efficient Fine-Tuning

LoFT: Low-Rank Adaptation That Behaves Like Full Fine-Tuning

  • Aligns low-rank update dynamics with full fine-tuning.
  • Projects optimizer momentum and variance into the low-rank subspace.
  • Narrows the adapter/full fine-tuning gap without increasing inference cost.
Paper PDF
Multi-agent debate avoiding a majority trap and drifting toward truth through asymmetric cognitive energy
2026 · Multi-Agent Debate

Breaking the Martingale Curse: Multi-Agent Debate via Asymmetric Cognitive Potential Energy

  • Identifies why standard multi-agent debate can collapse toward erroneous consensus.
  • Uses asymmetric cognitive potential energy and peer prediction to create positive drift.
  • Frames debate as a mechanism for truth-directed convergence rather than majority reinforcement.
Publication list