Pinned Tweet
Recursive self-improvement and continual learning research drops every day, and people keep asking me what those concepts actually mean.
@TuXinming and I tried to unwrap the popular papers and products a bit: WHAT agents update, and WHEN agents update.
⬇️Blog: xinmingtu.cn/blog/2026/self-…
[1/9] There’s been incredible discussion and progress lately around Self-Evolving Agents, and Recursive Self-Improvement.
In our new blog post, @TongChen_ and I propose a 3×3 Framework to map this fast-growing landscape along two core axes:
👉 WHAT does the agent update? (The Substrates)
👉 WHEN do those updates persist? (The Persistence Horizons)
Here is the map. 🧵👇
🔗Read our full blog deep dive, interactive 3×3 diagrams, our expanded appendix mapping 30+ work xinmingtu.cn/blog/2026/self-…
Tong Chen retweeted
So excited that this finally came out!! Studying reward hacking at scale in autoresearch agents is hard, and we ended up injecting hacks into long agent trajectories, so that we can do a controlled study on how monitors catch them.
HackBench serves as a stress test of agent monitor’s long context recall, understanding of the research task, and reasoning over multiple turns.
Introducing *HackBench* 🤡💀⏱️📊 ⏲️ 💪🔧 🔥
- Detecting & localizing reward hacking in autoresearch agents
- Frontier models fail
- Dataset on HuggingFace
- Recipe to collect natural & synthetic hacks
Blog post: facebookresearch.github.io/R…
🧵1/5
Tong Chen retweeted
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-opt…
💻 Try it on your task: github.com/xingjian-bai/opal
🧵
Tong Chen retweeted
Agent evals grade thousands of runs, then use each once for a score.
We gave them a second life: hide the result, ask another agent "did it work?"
Models are bad judges but good choosers: picking 1 of 5 runs lifts Fable 5.1 from 58% → 70% on Terminal-Bench 4.0. 🧵
Tong Chen retweeted
Claim: teaching AIs to be better collaborators with humans improves the quality of their work!
🤖🤝Introducing *AutoInteract* 🔁💬🧑💻🧑🤝🧑✨🌱
- creates train data to improve AI's collaborative abilities
- simulates the way humans act & constructs challenging multi-turn interaction tasks
- Improves LiveCodeBench-Pro, Multi-turn LCB-Pro & SWE-together
We expect the recipe generalizes to much more!
Blog post: facebookresearch.github.io/R…
🧵1/5
Tong Chen retweeted
I'm on the faculty job market. At #COLM, let's chat!
I work on compositional intelligence for participatory AI: modular AI systems built bottom-up by diverse stakeholders.
Also: knowledge, structured data, pluralism, multi-agent, and safety.
More: bunsenfeng.github.io
Tong Chen retweeted
Introducing Gacha Decoding 🎲: we show how to diversify LM responses, beating prior work by >11x in sample efficiency.
Reroll your LM for varied ideas, data, RL envs, {your use here}!
We get diversity from instruction following + external RNG, not token entropy. Here's how 🧵
Tong Chen retweeted
Harness evolution doesn’t beat test-time scaling on Terminal-Bench. But on long-horizon, out-of-distribution games, where harness evolution can help with adaptation, it shines: +80% on ARC-AGI-3 and +11% on EdgeBench games under matched budgets.
Check out our newest findings: arxiv.org/abs/2607.12227
Automatic harness evolution appears to be a promising path toward AI self-improvement, but we find that its gains still largely come from repeated sampling and show limited generalization.
Blog post: yikee.github.io/harnessevolu…
Code: github.com/rethinking-harnes…
Tong Chen retweeted
‼️The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA!
Introducing 🩵Context Language Models (CLMs)🩵
- Natively manage their own context
- Treat context as a file
- Learn policies in CLM weights, no harness
Tong Chen retweeted
Image generation models CANNOT follow instructions (yet)🚨
We introduce Visual Verifiable Rewards (VVR) to solve it✅
VVRBench asks the model to "generate three yellow circles"🟡
Surprisingly, SOTA models, including GPT-Image-2.5-Sunburst, fail on complex VVR tasks.
RL with VVR (RLVVR), and mixing VVR intro standard image post-training rewards, improves instruction following in natural prompts.
Tong Chen retweeted
Claim: we've solved the AI slop problem (!) 💩🧹✨
Blog post: facebookresearch.github.io/R…
🧵1/5
Key idea: take *expert* human writing and learn rubrics that find the gap between experts and models. Train with those rubrics.
We train with RL-XAR (RL with eXpert Aligned Rubrics) & see large performance gains on writing scientific paper sections, Pulitzer prize novel continuations and high quality Wikipedia pages.
Tong Chen retweeted
Introducing ApprenticeBench: computer use + continual learning on a real job.
We show Fable 5.1 and GPT-6 Astra can now continually learn on a job and surpass human professionals. A decisive step change in AI's job readiness.
No FDEs. Agents deploy themselves into the job. 🧵
Tong Chen retweeted
Muse Spark 1.3 is out, with strong performance. My team contributed to its long context & IF abilities (@tesatory, @shehzaadzd, Jack Lanchantin) & coding ability (@ping_iris_yu, @EasonNie) amongst other things!
Tong Chen retweeted
We share our early research on building Software World - a "GitHub" run by agents. We deploy agents for Python packages in a dependency chain, and each agent is tasked with collaborating with others and optimizing the package it owns.
Tong Chen retweeted
Harness evolution brings a 0.8B model to 100% on ALFWorld. But how much is generalizable improvement vs. overfitting or test-time scaling?
Our method, Harness Delta Attribution, analyzes gains and attributes them to these categories. On 4 benchmarks, lots of the gain is overfitting or due to higher compute usage. Future evaluation should study this more!
Post: wenwen-d.github.io/blog/harn…
Tong Chen retweeted
[1/9] There’s been incredible discussion and progress lately around Self-Evolving Agents, and Recursive Self-Improvement.
In our new blog post, @TongChen_ and I propose a 3×3 Framework to map this fast-growing landscape along two core axes:
👉 WHAT does the agent update? (The Substrates)
👉 WHEN do those updates persist? (The Persistence Horizons)
Here is the map. 🧵👇
🔗Read our full blog deep dive, interactive 3×3 diagrams, our expanded appendix mapping 30+ work xinmingtu.cn/blog/2026/self-…
Tong Chen retweeted
Automatic harness evolution appears to be a promising path toward AI self-improvement, but we find that its gains still largely come from repeated sampling and show limited generalization.
Blog post: yikee.github.io/harnessevolu…
Code: github.com/rethinking-harnes…
Tong Chen retweeted
Agentic audit + expert reviewer would be the requirement for future benchmarks.
Our paper BenchGuard just got accepted to COLM 2026.
Will share more updates on the camera-ready version! arxiv.org/pdf/2604.24955
We audited SWE-Bench Pro, one of the most widely used AI coding benchmarks, and found it no longer reliably measures frontier coding capability.
We find 30% of SWE-Bench Pro tasks to be broken, and are retracting our previous recommendation that the research community use it as a leading coding eval.
openai.com/index/separating-…
Tong Chen retweeted
Presenting our paper on the social world model today.
⏰Wed, Jul 8, 2026 • 2:30 PM – 4:15 PM KST
📍HALL A #815
Come and chat if interested. Also happy to chat about agentic RL, test-time learning and world modeling with LLMs.
Tong Chen retweeted
I’ll be at ICML next week🇰🇷, July 4–10.
Would love to hang out for fun or chat with folks about research such as social simulation, world models, and agentic RL, as well as broader experiences around industry positions and job seeking.
Feel free to reach out if you’ll be there — happy to meet up!