Wenxiang Jiao 焦文祥

Harness-R1

Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories

Trains a 9B harness engineer via RL to patch runtime harnesses from agent failure trajectories, lifting task success from 44.3% to 53.6%.

OmniGAIA

OmniGAIA: Towards Native Omni-Modal AI Agents

Benchmark and foundation agent for omni-modal AI assistants — multi-hop queries across video, audio, and image, with OmniAtlas trained via hindsight-guided tree exploration and OmniDPO.

MMSkills

MMSkills: Towards Multimodal Skills for General Visual Agents

Reusable multimodal procedures encoded as state-conditioned skill packages, generated from public trajectories and consulted by a branch-loaded skill agent at runtime.

DeepAgent

DeepAgent: A General Reasoning Agent with Scalable Toolsets

A reasoning agent that tackles general tasks by searching for and using appropriate tools from over 16,000 RapidAPIs in an end-to-end agentic reasoning process.

Multi-Agent Debate

Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate

The MAD framework addresses the Degeneration-of-Thought problem in self-reflection and explores divergent chains of thought through structured agent interaction.

CipherChat

GPT-4 Is Too Smart To Be Safe: Stealthy Chat with LLMs via Cipher

Examines whether safety alignment generalizes to non-natural languages such as ciphers; GPT-4 understands ciphers well enough to produce unsafe outputs.

PsychoBench

On the Humanity of Conversational AI: Evaluating the Psychological Portrayal of LLMs

Evaluates diverse psychological aspects of LLMs, including personality traits, interpersonal relationships, motivational tests, and emotional abilities.

Is ChatGPT A Good Translator

Is ChatGPT A Good Translator? A Preliminary Study

ChatGPT performs competitively with commercial translation systems on high-resource European languages; GPT-4 further bridges the gap for low-resource and distant languages.