Researcher @OpenAI | Training GPT-5.x and 6.x thinking models | Ex. Tencent, MSR , AI2
San Francisco, CA
Joined December 2021
- Tweets385
- Following1K
- Followers5.2K
- Likes1.4K
Wenhao Yu retweeted
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
Wenhao Yu retweeted
Here is our current plan for OpenAI:
openai.com/index/built-to-be…
Wenhao Yu retweeted
What should AI generate in order to improve itself?
Not just more questions, traces, or answers.
We believe it should learn to generate environments.
Excited to share my first work after joining Tencent Hunyuan LLM. We study how models can construct reusable, verifiable environments that provide stable training signals for self-improvement.
This is only a first feasibility step, but we see environment construction as a necessary path toward truly self-improving AI.
Paper: arxiv.org/abs/2605.14392
Wenhao Yu retweeted
You've been asking for this one...
Now in preview: Codex in the ChatGPT mobile app.
Start new work, review outputs, steer execution, and approve next steps, all from the ChatGPT mobile app. Codex will keep running on your laptop, Mac mini, or devbox.
Wenhao Yu retweeted
Introducing GPT-5.5
A new class of intelligence for real work and powering agents, built to understand complex goals, use tools, check its work, and carry more tasks through to completion. It marks a new way of getting computer work done.
Now available in ChatGPT and Codex.
Wenhao Yu retweeted
Update: We have released our Benchmark at huggingface.co/datasets/elai… and Train dataset at huggingface.co/datasets/elai… !
🤔🤔Tired of static, lifeless image edits?
Not anymore! 🤗
🚀🚀We introduce MotionEdit, a framework supporting image editing that understands action, motion, interaction beyond static changes! 🤩🤩
🔗Full paper: arxiv.org/abs/2512.10284
✨Project page: motion-edit.github.io
Wenhao Yu retweeted
R-Zero is accepted by #ICLR26 !!!
𝑳𝑳𝑴𝒔 can really 𝑺𝒆𝒍𝒇-𝑬𝒗𝒐𝒍𝒗𝒆, 𝒘𝒊𝒕𝒉𝒐𝒖𝒕 𝑯𝒖𝒎𝒂𝒏 𝑫𝒂𝒕𝒂!
-- One LLM, two roles: Challenger creates tasks, Solver answers them.
-- No data, no labels, just a base model that learns and improves itself!
We name it 𝑹-𝒛𝒆𝒓𝒐: arxiv.org/abs/2508.05004
Wenhao Yu retweeted
For LLM RL, Entropy is not equal to exploration. Semantic diversity is also a lie.
Most RL exploration methods push LLMs to "look" different, but they ignore how the model actually learns.
We propose G²RL: Gradient-Guided Exploration for RL
arxiv.org/pdf/2512.15687
(1/n)
Wenhao Yu retweeted
Beautiful Tencent paper.
Shows a language model that keeps improving itself using only 1% to 5% human labeled questions while reaching the level of systems trained on about 20 times more data.
Earlier self play systems let a model write and solve its own questions, but over time it drifts, repeats narrow patterns, and can even perform worse.
Their method runs a challenger copy that generates questions and a solver copy that answers them, turning training into a question answer game between 2 agents.
When the challenger writes, it sometimes sees a few real human question answer pairs, which pull its synthetic questions toward realistic tasks instead of strange, off topic puzzles.
For each question, the solver tries several answers, the system estimates its success rate, and training keeps mainly mid difficulty questions where the solver is uncertain but not lost.
Because both human and synthetic questions pass this filter, the solver trains on focused, non trivial problems, avoids cheap tricks like inflating question length, and gains stronger math and general reasoning scores than earlier self play methods.
----
Paper Link – arxiv. org/abs/2512.02472
Paper Title: "Guided Self-Evolving LLMs with Minimal Human Supervision"
Wenhao Yu retweeted
Introducing Multi-Agent Evolve 🧠
A new paradigm beyond RLHF and RLVR:
More compute → closer to AGI
No need for expensive data or handcrafted rewards
We show that an LLM can self-evolve — improving itself through co-evolution among roles (Proposer, Solver, Judge) via RL — all without external supervision.
On Qwen2.5-3B-Instruct, Multi-Agent Evolve boosts average accuracy from 55% → 58% across 22 benchmarks.
Remarkably, the model automatically learns balance among roles:
- The Proposer first generates easy tasks.
- The Judge refines the difficulty metric.
- The Proposer then raises the challenge, forcing the Solver to improve.
- The system co-evolves until reaching equilibrium.
Multi-Agent Evolve: LLMs self-improve through co-evolution.
📄 Paper: arxiv.org/abs/2510.23595
💻 Code (coming soon): github.com/ulab-uiuc/Multi-a…
Wenhao Yu retweeted
🔍Aligned LMs are better at reasoning/safety, but lose out on skills like calibration and generation diversity, where pretrained models are better.
🤝How about Don't Throw Away your Pretrained Model, and use multiple model stages of LLM training pipelines in collaboration!
Wenhao Yu retweeted
@wyu_nd and I are recruiting 2026 Spring/Summer Research Interns at Tencent AI Lab 🚀
Topics include self-evolving, Agent Systems, Complex Reasoning, etc.
We are also hiring full-time researchers with PhD degrees, fully publication-driven. Please DM or email.
Wenhao Yu retweeted
It’s so much fun working with the other 39 community members on this project!
Start to try out various frontier models in BigCodeArena today.
Wenhao Yu retweeted
🚀Excited to share our latest research that introduces Reinforcement Learning on Pre-training Data (RLPT)—a breakthrough that frees LLM scaling from the bottleneck of limited high-quality text.
🌟RLPT lets models learn powerful reasoning autonomously, without relying on costly human annotations.
Here’s how RLPT works and why it matters:
1️⃣ Training-Time Scaling
Instead of relying only on supervised learning, RLPT lets the policy autonomously explore trajectories on pre-training data — unlocking more powerful reasoning across broader contexts and fostering more generalizable reasoning skills.
2️⃣ No Human Annotations Needed
Unlike RLHF or RLVR, RLPT derives reward signals directly from pre-training data. No extra labeling or annotation overhead.
3️⃣ Next-Segment Rewards
Models predict the next segment conditioned on prior context, and are rewarded by a GRM — enabling RL to scale directly on pre-training data.
✅ Stronger Performance: On Qwen3-4B-Base, RLPT improves by +3.0 (MMLU), +5.1 (MMLU-Pro), +8.1 (GPQA-Diamond), +6.0 (KOR-Bench), +6.6 (AIME24), +5.3 (AIME25).
✅ Better Scaling: Gains continue to grow with more compute — promising long-term scaling.
✅ Foundation for RLVR: Extends reasoning boundaries and strengthens further RLVR performance.
📄 Paper: arxiv.org/pdf/2509.19249
Wenhao Yu retweeted
Parallel thinking can make LLMs better reasoners—not just imitators!
Introducing Parallel-R1, the first RL framework that trains LLMs to think in parallel for complex reasoning tasks.
🔑 How it works:
- Start with SFT on easier tasks → instill parallel thinking
- Switch to RL → explore & generalize on harder problems
- Curriculum avoids the cold-start trap
🚀 Results:
- +8.4% accuracy over sequential RL models on math (MATH, AMC23, AIME)
- On AIME25, yields +42.9% over baseline
- Models shift from using parallel paths for exploration → to multi-perspective verification
Wenhao Yu retweeted
Label-free RL often collapses into low-entropy certainty.
A small diversity boost triggers an "aha moment" in a label-free setting: entropy rebounds, reasoning deepens, and pass@16/OOD jump!
Details in our NEW paper: arxiv.org/abs/2509.15194
Wenhao Yu retweeted
Thank @_akhaliq for introducing our Parallel-R1! 😀
Parallel thinking has recently emerged as a promising direction for enhancing model performance. However, most existing inference-time approaches rely on heuristic rules to aggregate results, without enabling the model to truly understand or utilize parallel thinking itself. In contrast, our Parallel-R1 introduces the first RL framework that empowers the model to actively explore and internalize parallel thinking. We hope that our work will lead to further studies in this novel direction.
Wenhao Yu retweeted
🚀 Hunyuan-Game 2.0 is here!
We've updated our game content engine with four major features, empowering artists to solve key challenges in game art production.
✨ What's New in 2.0:
🔹 Image-to-Video Generation: Turn static images into dynamic animations, including 360° character rotations and skill previews.
🔹 Custom LoRA Training: Train your own custom models with just a few images. This new capability helps studios and individual creators maintain a consistent art style and is perfect for building IP-specific assets—no coding required.
🔹 One-Click Refinement: This feature offers two modes: a high-consistency mode to enrich textures and lighting, and a high-creativity mode to transform characters into different styles with rich details.
🔹 SOTA Image Generation: In addition to the new features, we've also upgraded our image generation model and it now delivers SOTA performance for the gaming industry. It's deeply optimized for game-specific assets, providing outstanding quality and composition.
Hunyuan-Game 2.0 is now fully open to all creators, designed to empower community-driven innovation and accelerate the future of game art.
Try it now: hunyuan.tencent.com/game/hom…
Wenhao Yu retweeted
Introducing VerlTool - a unified and easy-to-extend tool agent training framework based on verl.
Recently, there's been a growing trend toward training tool agents with reinforcement learning algorithms like GRPO and PPO. Representative works include SearchR1, ToRL, ReTool, and ToolRL. While these achieve impressive performance, their training codes are either not fully open-sourced or too difficult to modify and customize with new tools, creating unexpectedly high engineering costs for the community when exploring new ideas.
To address these issues and reduce engineering overhead, we propose verl-tool.
Key Features:
1. 🔧 Complete decoupling of actor rollout and environment interaction - We use verl as a submodule to benefit from ongoing verl repo updates. All tool calling is integrated via a unified API, allowing you to easily add new tools by simply adding a Python file and testing independently.
2. 🌍 Tool-as-environment paradigm - Each tool interaction can modify the environment state. We store and reload environment states for each trajectory. For each training, you can launch
3. ⚡ Native RL framework for tool-calling agents - verl-tool natively supports multi-turn interactive loops between agents and their tool environments.
4. 📊 User-friendly evaluation suite - Launch your trained model with OpenAI API alongside the tool server. Simply send questions and get final outputs with all interactions handled internally.
We've successfully reproduced ToRL results using our verl-tool framework, demonstrating its correctness and demonstrating comparable performance on mathematical benchmarks.
VerlTool is an active ongoing project! We aim to incorporate more tools covering a wide range of use cases and expect they can be trained together in a single framework. Suggestions and contributions are highly welcomed!
Check out our GitHub: github.com/TIGER-AI-Lab/verl…
More details: 👇 (0/4)
