Frontis AI

"EEMA: Experience-Evolving Meta Agents for AutoML" has been accepted to #EMNLP2026 Main Conference. EEMA, the experience layer behind OpenMLE-Evo-Max in our Frontis-MA1 report, distills public competition artifacts and failure traces into compact Skills, then routes them across draft / debug / improve rounds. On MLE-bench Lite (22 tasks, 12h/task, Codex + GPT-5.4-medium, single 24GB GPU): → 77.3% medal rate (base agent alone: 50.0%) → 20.1% avg leaderboard grade (base: 38.0%) → 96.0% execution success (base: 81.4%) Same base agent, no extra training — and it beats GPT-5.5-xhigh (68.2%). Code will land in OpenRSI, alongside OpenMLE and Frontis-MA1: 🔗 github.com/FrontisAI/OpenRSI Congrats to everyone who worked on it 👏

🎉 EnterpriseClawBench has been accepted to the EMNLP 2026 Industry Track!

Automating AI R&D is also what we pursue, and we believe it will bring tremendous value to the industry, accelerating AI solutions for the trillion scenarios of millions of enterprises.

A checkpoint worth writing down. This June, I completed my PhD at Tsinghua University and joined @FrontisAI as CTO. What we’re building: products that put AI4AI and recursive self-improvement (RSI) to work on real, high-value tasks — across enterprise workflows and scientific discovery. Frontis Horizon already serves 100+ enterprises. We’re building the RSI product layer on that real operating ground: real workflows, real constraints, and real outcome signals. For enterprise work, our architecture is evolving around three layers (ME-WE-MA): ME manages each employee’s context, memory, and intent. WE is the network of expert AI agents that ME calls to get work done. MA builds and evolves those experts using outcomes from real tasks. ME holds the context. WE does the work. MA improves WE. Along the way, we’ve been sharing the technical foundations openly — a survey, two benchmarks, and a model: 01 · A Survey “Self-Improving Agents in the Era of Experience: A Survey of Self- to Meta-Evolution” Introduces Meta-Evolution: improvement that transfers across tasks, rather than self-evolution confined to a single task. We believe cross-task evolution is the form that matters for real-world deployment. GitHub: github.com/FrontisAI/Awesome… Paper: openreview.net/pdf?id=IUltZS… 02 · A Workplace Benchmark “EnterpriseClawBench: Benchmarking Agents from Real Workplace Sessions” Built from real workplace sessions through a privacy-preserving pipeline: 852 reproducible tasks spanning files, tools, and usable business-artifact delivery. It evaluates complete agent systems — harness and model together — across quality, cost, runtime, and skill transfer. The reusable construction and evaluation protocol is open, while proprietary enterprise data remains protected. GitHub: github.com/FrontisAI/Enterpr… Paper: arxiv.org/abs/2606.23654 03 · A Scientific Benchmark “NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?” NatureGym turns Nature-family papers into executable, high-value AI4AI tasks. NatureBench then measures whether coding agents can reproduce their published SOTA. GitHub: github.com/FrontisAI/NatureB… Paper: arxiv.org/abs/2606.24530 04 · A Model “Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering” An AI4AI model trained at scale for machine learning engineering, with the full stack released — data, weights, training, and search code. Frontis-MA1-35B + OpenMLE-Evo-Max reaches 71.21% Medal Average on MLE-Bench Lite, with 12 hours on a single RTX 4090 per task — above GPT-5.5 + Codex and within 1.5 points of GPT-5.6 Sol and the 2.8T Kimi K3. GitHub: github.com/FrontisAI/OpenRSI Paper: arxiv.org/abs/2607.28568 These are not four disconnected research outputs. Together, they form a product stack taking shape: transferable improvement, evaluation grounded in real science and real work, and models that improve machine learning workflows. The product layer goes on top of these foundations. More on that soon. We are still early, but we’ve seen the first signs of what RSI and AI4AI can do on real tasks — and we’re building toward it. If you want to use what we’re building, contribute to OpenRSI, or build with us, my DMs are open.

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.

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: github.com/FrontisAI/OpenRSI Models: huggingface.co/collections/F… Project: frontisai.github.io/OpenRSI/ Paper: huggingface.co/papers/2607.2…