Projects

SWARM

Multi-agent RL for deep research · Snowflake AI Research & Yale · May 2026 – Present

Designed an RL training recipe for multi-agent deep research systems, in which an orchestrator delegates sub-tasks to one or more executors, collects their findings, and issues new tasks until it is confident in the answer.
Jointly trained the orchestrator and executors under a shared policy, sampling executor trajectories from each orchestrator rollout based on its outcome (successful executors from correct rollouts, failed executors from incorrect ones).
Prevented model collapse from invalid tool calls by ending orchestrator rollouts with zero reward and truncating executor trajectories at their first invalid call.
Extended PRIME-RL to schedule executor requests alongside orchestrator requests, balancing load across inference GPUs.
Trained Qwen3.5-4B on BrowseComp-Plus with LLM-as-a-Judge rewards and observed the orchestrator learning parallel workflows.

Attack-and-Defense CTF benchmark for LLM agents · Yale · May 2026 – Present

Built a game-like benchmark modeled on Attack-and-Defense CTF to evaluate LLM agents’ offensive and defensive cybersecurity capabilities, pairing one model as a red (attacker) agent against another as a blue (defender) agent over an isolated Docker match network.
Curated 300+ intentionally vulnerable services from open-source Docker environments; agents play black-box without knowing the underlying vulnerability.
Built the game infrastructure on the open-source A&D CTF framework ForcAD, running the service and each agent in separate Docker containers with role-based access, and isolating each match so that matches run in parallel.
Built a global checker that periodically plants flags, verifies service availability, and scores both sides independently of agent steps, plus automatic recovery that restores a service snapshot when a faulty patch crashes the service.
Developed ReAct-style red and blue agents that play asynchronously.

Digital forensics benchmark for LLM agents · Yale · Jan 2026 – Present

Built a benchmark of 14 labs (237 questions) curated from open-source forensic challenges to evaluate LLM agents on end-to-end digital forensics investigations, spanning network, memory, and disk forensics, log analysis, and malware analysis.
Developed a ReAct-style LLM agent that investigates evidence inside per-lab Docker environments, supporting multiple LLM APIs via LiteLLM.
Designed hybrid grading that combines string matching for factual questions with rubric-based LLM-as-a-Judge for analytical ones, plus post-hoc trajectory analysis.

Multiagent framework for science automation · Yale · Aug 2025 – Oct 2025 · GitHub · arXiv

Developed the ReviewerAgent, which evaluates papers generated by the freephdlabor multiagent system.
Developed a resume mechanism that lets the system restore memory from a previous run, and a non-blocking mechanism that allows real-time human intervention.