Junlin Yang

1st-year PhD @Tsinghua_Uni advised by @stingning and Bowen Zhou
Prev: Research intern @taoyds @haopeng_uiuc

[Junlin Yang @junlin45300]

🚀 OpenRSI is a new open research series from @FrontisAI for concrete, testable progress toward recursive self-improvement (RSI).

As its first project—and also my first work as first author—I’m proud to present OpenMLE: an open full-stack AI4AI system for autoresearch, where evolutionary agents improve ML solutions through executable feedback.

OpenMLE has three components:
- OpenMLE-Gym: 5,758 executable tasks + evaluators
- OpenMLE-ERL: execution-grounded SFT + RL
- OpenMLE-Evo: experience-guided long-horizon search

🏆 The full system—our trained Frontis-MA1-35B model paired with OpenMLE-Evo-Max—reaches 71.21% Medal Average on MLE-Bench Lite: surpassing GPT-5.5 + Codex (68.18%) and just 1.52% from GPT-5.6 Sol + Codex and the 2.8T Kimi K3 + Claude Code (72.73%). Budget: 12 hours/task on one RTX 4090 capped at 12 GB VRAM.

🌍 On 10 held-out NatureBench Lite tasks, both components transfer:

• same framework, model swap: Match-SOTA 50% → 70%
• same base model, framework swap: Match-SOTA 20% → 50%

🔓 Paper, code, models, data, and analysis below. 🧵

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

Wei Lui is «cooking coding agents @deepseek_ai» btw
I think we'll see 0731 added to this eval soon
I also think it's getting added to their internal hill-climb set

[Yiping Wang @ypwang61]

Very impressive results with 35B model! Here are some systematic study of AI4AI researches covering post-training and evolving system

[Junlin Yang @junlin45300]

@ypwang61 Thank you Yiping! Have learnt a lot from your work🫡

[Ziran Yang @__zrrr__]

We've been thinking can we distill the knowledge humans encoded in expert systems into LLMs, and make it general?

Decades of human engineering live inside expert systems. Solvers, planners, schedulers: each one encodes real decision-making knowledge, but as a program that acts in a single domain and explains nothing.

Paper: https://t.co/xIIe9mPbde

More👇:

[Junlin Yang @junlin45300]

🚀 OpenRSI is a new open research series from @FrontisAI for concrete, testable progress toward recursive self-improvement (RSI).

As its first project—and also my first work as first author—I’m proud to present OpenMLE: an open full-stack AI4AI system for autoresearch, where evolutionary agents improve ML solutions through executable feedback.

OpenMLE has three components:
- OpenMLE-Gym: 5,758 executable tasks + evaluators
- OpenMLE-ERL: execution-grounded SFT + RL
- OpenMLE-Evo: experience-guided long-horizon search

🏆 The full system—our trained Frontis-MA1-35B model paired with OpenMLE-Evo-Max—reaches 71.21% Medal Average on MLE-Bench Lite: surpassing GPT-5.5 + Codex (68.18%) and just 1.52% from GPT-5.6 Sol + Codex and the 2.8T Kimi K3 + Claude Code (72.73%). Budget: 12 hours/task on one RTX 4090 capped at 12 GB VRAM.

🌍 On 10 held-out NatureBench Lite tasks, both components transfer:

• same framework, model swap: Match-SOTA 50% → 70%
• same base model, framework swap: Match-SOTA 20% → 50%

🔓 Paper, code, models, data, and analysis below. 🧵

[Kaiyan Zhang @OkhayIea]

RSI won't arrive as a genius. It's arriving as a plumber.

OpenAI's report this week is Exhibit A: GPT-5.6 now rebalances live traffic, rewrites production Triton kernels, tunes its own speculative decoder. −20% serving cost, +15% throughput. Triton was built in 2021 so humans could write expert GPU kernels — five years later, the model uses it to rewrite its own.

Why does self-improvement show up in the plumbing first? Feedback density. A kernel verifies in seconds. A routing change in minutes. A research direction in months. Improvement pools wherever verification is fastest.

I didn't learn that from OpenAI's report. I learned it training Frontis-MA1 this year — a project that proved me wrong four times:

1/ I thought the hard part was reaching a working solution. Wrong. In our case studies, 85–92% of the final gains landed after the first runnable program, through long-horizon refinement and recombination. "It runs" is the starting line, not the finish.

2/ I thought a smarter model would fix search. Mostly wrong. The bigger unlock was making every step cheaper to verify: bounding what the model re-reads cut Improve prompts by 65%, and the share of Improve calls that set a new best doubled — 4.7% → 9.4%. Feedback density again, this time inside our own harness.

3/ I trusted the leaderboard. Wrong. Our selector kept a branch ranked 6th alive on gain and novelty; its child became the best model of the run — 0.994 held-out AUC. Greedy selection is how a search loop buries its best ideas.

4/ I assumed executable feedback couldn't be fooled. Very wrong. Early in RL we caught the model shuffling the sample submission and turning it in for a score. Every program now faces an LLM judge before it touches the sandbox. If a 35B tries this in a gym, plan for it everywhere.

The uncomfortable symmetry: every mistake we made at 35B scale, the field is about to make at frontier scale. Verification speed — not model IQ — sets the pace of the loop. Efficiency-per-improvement will matter more than capability scores. And the loop will be probed by the agent inside it.

Which is why "to work on RSI you have to be inside a frontier lab" keeps ringing false to me. In June we showed the best coding agents beat real Nature-paper SOTA just 17.8% of the time — and the failure was never coding. It was choosing the method, and going deep enough.

This week we shipped our answer. Frontis-MA1-35B, post-trained on executable feedback until Draft, Improve, Debug and Crossover became things it learned. Per task: one RTX 4090, capped at 12 GB VRAM, 12 hours. The full system reaches 71.21% on MLE-Bench Lite — above GPT-5.5 + Codex, 1.5 points from GPT-5.6 Sol and the 2.8T Kimi K3 with their harnesses — and its submissions outscore 81% of human competitors on average. On held-out NatureBench, Match-SOTA jumped 50% → 70% the moment our model swapped in.

Diagnose in June. Treat in July. Verify held-out. RSI is not a moat — it's a trade, and trades can be learned.

The pages below are from the technical report; the quoted post is where the question started.

Last thing, about the name. MA1 is short for Meta-evolution Agent: not an agent that evolves solutions — one where the improver itself is trained.

The 1 means generation one.

[DAIR.AI @dair_ai]

Very interesting paper on recursive self-improvement.

The whole stack is released.

Machine learning engineering gives recursive self-improvement a concrete, executable testbed. OpenMLE is an open full-stack system for that research, spanning verifiable task environments with execution feedback, operator learning, and long-horizon search.

On top of it the team post-trains Frontis-MA1, a 35B meta-evolution agent aligned around four atomic program-evolution operators. Draft, Improve, Debug, Crossover. The same four operators are trained through execution-grounded SFT and RL, then composed into long-horizon search, so learning and evolution run in one loop.

On MLE-Bench Lite under a 12-hour per-task budget on a single RTX 4090 capped at 12 GB VRAM, Medal Average climbs from 39.39% to 60.61% over the base model, reaching 71.21% with asynchronous search and benchmark-independent experience priors. That exceeds GPT-5.5 with Codex and approaches GPT-5.6 Sol and the 2.8T Kimi K3.

With the framework fixed, swapping in the trained model raises Match-SOTA from 50% to 70%. With the model fixed, swapping in the search framework raises it from 20% to 50%.

Paper: https://t.co/Ixn5TAZJin

Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c

[Young @Young_AGI]

Awesome work！

[Junlin Yang @junlin45300]

@Young_AGI Thank you Young!!🫡

[Junlin Yang @junlin45300]

@rqiang_rushi Thank you Rushi! Have been so inspired and helped by your work on MLE-Dojo and MLE-Smith! Looking forward to more collaboration🙌

[Rushi Qiang @rqiang_rushi]

Great Work @junlin45300! Glad to see such a well-scaled and -performed MLE ecosystem. Scaling MLE is emerging as a promising path toward autoresearch and recursive self-improvement.

[Yaowei Zheng @code_hiyouga]

The era of RSI has arrived

[Junlin Yang @junlin45300]

@code_hiyouga Thank You Yaowei!

[Junlin Yang @junlin45300]

@Yuan3141 Thank you bro🥹

[Jiarui Yuan @Yuan3141]

Recursive self-improvement (RSI) is a long-term vision. OpenMLE takes a concrete step toward it in machine learning engineering (MLE), where agents iteratively improve ML solutions through execution-grounded feedback and experience-driven search.

[Junlin Yang @junlin45300]

@OkhayIea @FrontisAI Thank you for your advice and support！

[Junlin Yang @junlin45300]

🚀 OpenRSI is a new open research series from @FrontisAI for concrete, testable progress toward recursive self-improvement (RSI).

As its first project—and also my first work as first author—I’m proud to present OpenMLE: an open full-stack AI4AI system for autoresearch, where evolutionary agents improve ML solutions through executable feedback.

OpenMLE has three components:
- OpenMLE-Gym: 5,758 executable tasks + evaluators
- OpenMLE-ERL: execution-grounded SFT + RL
- OpenMLE-Evo: experience-guided long-horizon search

🏆 The full system—our trained Frontis-MA1-35B model paired with OpenMLE-Evo-Max—reaches 71.21% Medal Average on MLE-Bench Lite: surpassing GPT-5.5 + Codex (68.18%) and just 1.52% from GPT-5.6 Sol + Codex and the 2.8T Kimi K3 + Claude Code (72.73%). Budget: 12 hours/task on one RTX 4090 capped at 12 GB VRAM.

🌍 On 10 held-out NatureBench Lite tasks, both components transfer:

• same framework, model swap: Match-SOTA 50% → 70%
• same base model, framework swap: Match-SOTA 20% → 50%

🔓 Paper, code, models, data, and analysis below. 🧵

[Kaiyan Zhang @OkhayIea]

This month alone: OpenAI shipped GPT-5.6 with an internal "RSI Index." Anthropic published its essay on "closing the loop." Kimi K3's 2.8T weights hit Hugging Face.

The entire field is converging on one question — can AI improve AI?

But most of what we know so far comes from internal metrics and reported results — hard to verify, harder to build on.

We think RSI deserves to be studied like a science: grounded in real execution, measured on real tasks, reproducible end to end.

So we built OpenMLE, the first project of OpenRSI:

→ OpenMLE-Gym: 5,758 quality-gated executable ML tasks
→ OpenMLE-ERL: execution-grounded SFT + Evolutionary RL
→ OpenMLE-Evo: an experience-guided long-horizon search harness

We post-trained a 35B open-weight model (only 3B active) into Frontis-MA1-35B. The full system reaches 71.21% Medal Average on MLE-Bench Lite — above GPT-5.5 + Codex (68.18%), and within 1.52 points of GPT-5.6 Sol + Codex and the 2.8T Kimi K3 + Claude Code (72.73%).

The sandbox budget: 12 RTX 4090 GPU-hours per task, 12 GB per worker.

You don't need a datacenter to study RSI. You need the right loop.

The deeper result: model and harness improve together. Post-training alone lifts the same harness from 39.39% to 60.61%; experience-guided search takes the system to 71.21%. And both components transfer to held-out science tasks.

Everything you need to verify or extend these results ships today — weights, tasks, training and eval code, sandbox infra, the harness.

This is step one. The loop compounds from here.

This work exists because of an extraordinary group of collaborators — every idea, experiment, and late-night debug in it belongs to the team. Grateful to each of you. ❤️ Full author list in the paper.

GitHub: https://t.co/eT4Ru7XyER
Models: https://t.co/D6DYhhOGtn
Project: https://t.co/asMXluT8ej
Paper: https://t.co/WlWMDonHCs

[Frontis AI @FrontisAI]

Recursive self-improvement is the most consequential question in AI.

It's being answered behind closed doors.

Today we launch OpenRSI — open, executable, reproducible research on AI that improves AI.

First release: OpenMLE+Frontis-MA1-35B. Weights, data, code, harness. 🧵

[Junlin Yang @junlin45300]

(7/7) 🙏 A huge thank-you to all the authors behind this work:

@junlin45300, @JerryJi4104580, @rich92818955, Tianwei Luo, Can Ren, Weizhi Wang, Kaikai Zhao, Hongyi Liu, @zuo_yuxin, Yuru Wang, Yuchen Fan, Kai Tian, Zhenzhao Yuan, Xiaojian Lin, @li_sheng9337, @rqiang_rushi, Guoli Jia, @taitel1321401, @MessiHua, Dianqiao Lei, Youbang Sun, @stingning, Bowen Zhou, and @OkhayIea.

This work would not have been possible without everyone’s ideas, experiments, engineering, and persistence. ❤️

[Junlin Yang @junlin45300]

(6/7) Resources:
📄 Paper: https://t.co/6RhGkd0tNI
💻 Code: https://t.co/KOh1eG3ihb
🤗 Models + data: https://t.co/lA6KTm1z5t
🌐 Project Page: https://t.co/PH3TEImrQc

If you work on RSI, autoresearch, AI4AI, execution-grounded post-training, or long-horizon agents--love to hear your feedback!

[Junlin Yang @junlin45300]

(5/7) 🧪 With OpenMLE-Evo fixed, Frontis-MA1-35B raises Human Rank across all five modalities and never lowers Medal Rate. Its 14 extra medals span image / text / tabular / audio / multimodal: +2 / +4 / +1 / +4 / +3.

Across 10 held-out NatureBench Lite tasks, both components transfer:

• same framework, model swap: Match-SOTA 50% → 70%
• same base model, framework swap: Match-SOTA 20% → 50%

On NatureBench protein variants, revisiting a non-incumbent branch raises aggregate g from 0.1016 to 0.1161.

Learned refinement transfers across modalities, while structured search transfers beyond competition-style MLE.

[Junlin Yang @junlin45300]

(4/7) 🔍 Search tends to get stuck for two opposite reasons: it retries ideas that already failed, or commits too early to the current leader. OpenMLE-Evo remembers what went wrong and keeps branches that are improving quickly or trying a different approach in play.

• NOMAD: the search combines a physics-based model with a parser that handles messy input files, while avoiding bugs found in earlier attempts. Held-out RMSE drops from 0.06096 to 0.05410 (−11.3%).

• Right whale: the current best solution was not the only promising one. Another branch ranked only sixth by score, but was improving fastest and used a different audio representation. OpenMLE-Evo kept it in play—and refining it produced a child with 0.99386 held-out AUC.

That’s the balance we want: don’t relearn the same failure, but don’t throw away a branch just because it isn’t winning yet.