My research studies how large language models reason, act as agents, refuse unsafe requests, and portray psychological traits. I develop benchmarks, open-source systems, and training/alignment methods to diagnose model failures and build more reliable AI applications.
General Agents
Multimodal, tool-using, and collaborative agents for complex, long-horizon tasks.
Abstracts interaction data into atomic decision experiences via hindsight reasoning, then retrieves them at inference through policy-driven wide- and deep-search strategies, improving fine-grained visual perception and multimodal reasoning over trajectory-level retrieval baselines.
Represents reusable multimodal procedures as compact, state-conditioned skill packages — a textual procedure paired with runtime state cards and multi-view keyframes — generated from public trajectories and consulted by a branch-loaded skill agent at runtime.
A benchmark and foundation agent for omni-modal AI assistants. OmniGAIA synthesizes multi-hop queries across video, audio, and image via an omni-modal event graph; the accompanying OmniAtlas agent uses active omni-modal perception trained with hindsight-guided tree exploration and OmniDPO.
LLM Reasoning
Mathematical, reflective, and efficient long-chain reasoning.
Tackles overthinking in large reasoning models with a small reflection model that enables parallel sampling and sequential revision during online RL, plus a reflection reward that preserves reflection ability — reducing inference cost by ~35% without sacrificing accuracy.
A dual-reward RL framework that classifies problems as Simple or Hard in real time and adaptively shortens or extends Chain-of-Thought, improving both accuracy and token efficiency on challenging math benchmarks.
LLM Safety
Risk awareness, jailbreak robustness, multilingual safety, and refusal behavior.
PaSBench evaluates proactive safety across 416 multimodal scenarios in five safety-critical domains. Top models such as Gemini-2.5-pro reach 64–71% accuracy but miss 45–55% of risks under repetition — failure analysis traces this to unstable proactive reasoning rather than missing knowledge.
Identifies a refusal position bias in safety-tuning data and proposes Decoupled Refusal Training: MLE with a harmful response prefix plus Reinforced Transition Optimization, letting LLaMA-3 and Mistral models refuse at any position throughout a harmful response without hurting performance.
Decomposes a malicious image-generation query into innocuous sub-queries and iteratively edits the output; bypasses safeguards on GPT-4V, GPT-4o, and Gemini 1.5/Pro in over 60% of cases. A companion Think Twice Prompting defense blocks more than 95%.
LLM Personality
Emotion, personality, and psychological portrayals in conversational AI.
Uses emotion appraisal theory to test how LLMs' feelings shift across 400+ situations grouped into 36 factors, benchmarked against responses from 1,200+ human subjects. Models including GPT-4, Mixtral-8x22B, and LLaMA-3.1 respond appropriately in some cases but fail to align with human emotional behavior or connect similar situations.
Across 2,500 settings per model on GPT-3.5/4, Gemini-Pro, and LLaMA-3.1, shows that LLMs respond consistently to the Big Five Inventory. Further demonstrates GPT-3.5 can emulate diverse personalities and represent specific population groups when given targeted prompts.
For the chronological list of all papers, see my Google Scholar page.