Wenxiang Jiao 焦文祥

LLM Algorithm Expert / Xiaohongshu

Towards agents that reason deeply, learn from experience, and continually improve.

My research spans capability learning, deep reasoning, and self-evolving agents. I currently focus on tool and skill learning, agentic reinforcement learning, and harness engineering—how an agent’s runtime can support learning and self-improvement.

Email Google Scholar GitHub X / Twitter

Portrait of Wenxiang Jiao — 焦文祥 Wenxiang Jiao

Research agenda

From learning to self-evolution.

Explore research

Agents & Self-Evolution

How can agents learn from experience and improve themselves?

Tool & skill learning · Agentic RL · Harness Engineering 02

Deep Reasoning

How can models reason more deeply and efficiently?

Multi-agent collaboration · Efficient reasoning · LLM psychology 03

Capability Learning

How do language models acquire and update their capabilities?

Multilingual learning · Translation · Continual pretraining · Safety

A closer look

Selected work

All research

Agents & Self-Evolution

Harness-S

ArXiv 2026

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Trains a shared agent policy using complementary trajectories from sibling harnesses, transferring successful behaviors back to the original runtime.

Paper Code

Inside the paper

Harness-S framework showing directional harness evolution, complementary rescue training, and transfer back to a shared agent policy. — **Learning from sibling harnesses**Figure 2 · p. 3
A fixed sibling harness pool rescues failed tasks and transfers successful behaviors into a shared policy.

Three plots illustrating locality, complementarity, and nonstationarity in sibling harness evolution for Qwen3.5-4B on AppWorld. — **Properties of harness evolution**Figure 1 · p. 2
Sibling harnesses repair different failures, offer complementary coverage, and change in utility as the policy learns.

Harness-S results table comparing policy-only training, harness evolution, and co-evolution on three long-horizon agent benchmarks. — **Long-horizon agent benchmark results**Table 1 · p. 7
AppWorld, SWE-Gym, and DBBench results under the base harness and Rank-1 routing.

Agents & Self-Evolution

Harness-R1

ArXiv 2026

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

Learns to turn agent failure trajectories into executable runtime patches, using task reruns as reinforcement learning feedback.

Scholar citations14 GitHub stars123

Paper Code

Inside the paper

Harness-R1 framework connecting failure trajectories, a trainable harness engineer, editable runtime hooks, and environment rewards. — **Learning to edit agent harnesses**Figure 2 · p. 3
Failure evidence becomes executable runtime patches; task reruns provide feedback to train the harness engineer.

Plot comparing matched-baseline reward changes from harness editing across WebShop, ALFWorld, and DBBench, with diamonds marking the equal-weight average. — **Effect of harness editing**Figure 1 · p. 2
Matched-baseline reward changes across three benchmarks; diamonds denote the equal-weight average.

Harness-R1 main results table with task success and reward on WebShop, ALFWorld, and DBBench. — **Task success across three benchmarks**Table 1 · p. 7
Main results on WebShop, ALFWorld, and DBBench for prompt-based, supervised, and reinforcement-trained harness engineers.

Agents & Self-Evolution

DeepAgent

WWW 2026

DeepAgent: A General Reasoning Agent with Scalable Toolsets

A general reasoning agent with autonomous tool discovery, memory folding, and tool reinforcement learning across more than 16,000 tools.

Scholar citations95 GitHub stars1.1k

Paper Code

Inside the paper

DeepAgent framework diagram showing autonomous tool discovery, memory folding, and ToolPO training. — **Framework overview**Figure 3 · p. 3
Tool discovery, memory folding, and ToolPO in a unified reasoning process.

Bar and radar charts comparing DeepAgent with baseline agents on general tool-use tasks and downstream applications. — **Overall performance**Figure 1 · p. 1
General tool-use and downstream results; scores are normalized to the best result.

Main tool-use results table for ToolBench, API-Bank, TMDB, Spotify, and ToolHop, including base and reinforcement-trained DeepAgent. — **Tool-use benchmark results**Table 1 · p. 5
Pass@1 results with ground-truth tools and open-set tool retrieval.

Agents & Self-Evolution

MMSkills

NeurIPS 2026

MMSkills: Towards Multimodal Skills for General Visual Agents

Builds reusable multimodal skills from interaction trajectories, combining textual procedures, visual state cards, and multi-view keyframes.

Paper Code

Inside the paper

MMSkills framework showing multimodal skill packages, skill generation from public trajectories, and branch-loaded agent inference. — **Multimodal skills framework**Figure 2 · p. 4
Reusable skill packages, trajectory-based generation, and branch-loaded guidance for visual agents.

OSWorld results table comparing no-skill, text-only, and MMSkills conditions across six model families and desktop applications. — **OSWorld benchmark results**Table 1 · p. 6
Application-level success rates comparing no skills, text-only skills, and MMSkills across six visual models.

Deep Reasoning

Multi-Agent Debate

EMNLP 2024

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

Studies degeneration of thought in self-reflection and uses multi-agent debate to encourage divergent reasoning.

Scholar citations1815 GitHub stars621

Paper Code

Inside the paper

Multi-Agent Debate diagram showing affirmative and negative debaters and a judge, contrasted with self-reflection. — **Multi-Agent Debate framework**Figure 2 · p. 3
Debaters challenge competing solutions while a judge guides the discussion toward a final answer.

Common MT results table comparing GPT-4, GPT-3.5-Turbo, Vicuna models, self-reflection, and Multi-Agent Debate. — **Commonsense translation results**Table 1 · p. 5
Translation quality on Common MT, with automatic metrics and human assessments across three test subsets.

Deep Reasoning

REA-RL

ICLR 2026

REA-RL: Reflection-Aware Online Reinforcement Learning for Efficient Large Reasoning Models

Reflection-aware reinforcement learning for more efficient reasoning, retaining reflective ability while reducing inference cost.

Paper Code

Inside the paper

REA-RL workflow showing parallel reasoning samples, overthinking detection, sequential revision, and reward-guided policy updates. — **Reflection-aware reinforcement learning**Figure 2 · p. 5
Parallel sampling and sequential revision train more efficient reasoning with reflection-aware rewards.

REA-RL main results table comparing accuracy and token-consumption ratios on GSM8K, Math500, Gaokao23, AMC23, and AIME24. — **Reasoning accuracy and efficiency**Table 2 · p. 7
Accuracy and token-consumption ratios across five mathematical reasoning benchmarks and two token budgets.

Deep Reasoning

PsychoBench

ICLR 2024 Oral

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

A systematic evaluation of LLM psychological portrayals across personality, relationships, motivation, and emotional abilities.

Scholar citations188 GitHub stars140

Paper Code

Inside the paper

PsychoBench taxonomy connecting personality and ability tests to thirteen scales across four psychological dimensions. — **The PsychoBench assessment structure**Figure 1 · p. 3
Thirteen psychological scales covering personality, relationships, motivation, and emotional abilities.

PsychoBench personality results table comparing Big Five, Eysenck, and Dark Triad scores for language models and human reference groups. — **Psychological profiles across models**Table 3 · p. 10
Personality-scale results for language models, including jailbroken GPT-4, alongside human reference scores.

Citation and star counts checked Nov 4, 2026.

Recently

Updates

News archive

Sep 25, 2026

Three papers accepted to NeurIPS 2026, namely MMSkills, Agent2World, and VisFactor. Congratulations!

Apr 7, 2026

Three papers accepted to ACL 2026 conference (2 main + 1 findings). Congratulations!

Jan 26, 2026

Two papers (REA-RL, DeepCompress) accepted to ICLR 2026. Congratulations!

Jan 17, 2026

Gave an invited talk at Agentic AI Summit 2026 titled Scalable and Personalizable AI Agents.

Jan 13, 2026

DeepAgent accepted to WWW 2026. Congratulations!

Background

Experience & education

Experience

Sep 2025 — Present
Xiaohongshu
LLM Algorithm Expert
General agents, tool and skill learning, and self-improvement through Harness Engineering.
Nov 2021 — Aug 2025
Tencent
Senior Researcher
Hunyuan, Aug 2024–Aug 2025: long-context understanding and complex instruction following.
AI Lab, Nov 2021–Aug 2024: multilingual models, translation, reasoning, and multi-agent collaboration.

Education

Aug 2017 — Nov 2021
The Chinese University of Hong Kong
Ph.D. in Computer Science & Engineering
Advised by Prof. Irwin King and Prof. Michael R. Lyu.
Aug 2015 — Jun 2017
Nanjing University
Master’s in Optical Engineering
Sep 2011 — Jul 2015
Nanjing University
Bachelor’s in Information Engineering

Academic community

Recognition & service

Selected recognition

2024
ICLR Oral Presentation
PsychoBench · Top 1.2%
2022
Best Paper Award
Multilingual Representation Learning Workshop, EMNLP
2022
1st Place · WMT
Large-Scale African Languages Translation · Constrained track

Reviewing

Nature, Nature Machine Intelligence, NeurIPS, ICLR, ICML, ACL, and EMNLP.

Selected invited talks

Scalable and Personalizable AI Agents
Agentic AI Summit · Jan 2026

From Machine to Human-like Intelligence
CCMT tutorial · Oct 2023

🌎 Visitor Footprints

... visitors from around the world

Loading visitor map…

Wenxiang Jiao 焦文祥

From learning to self-evolution.

Agents & Self-Evolution

Deep Reasoning

Capability Learning

Selected work

Harness-S

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Harness-R1

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

DeepAgent

DeepAgent: A General Reasoning Agent with Scalable Toolsets

MMSkills

MMSkills: Towards Multimodal Skills for General Visual Agents

Multi-Agent Debate

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

REA-RL

REA-RL: Reflection-Aware Online Reinforcement Learning for Efficient Large Reasoning Models

PsychoBench

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

Updates

Experience & education

Experience

Xiaohongshu

Tencent

Education

The Chinese University of Hong Kong

Nanjing University

Nanjing University

Recognition & service

Selected recognition

Reviewing

Selected invited talks

🌎 Visitor Footprints