OpenRSI Anatomy: Making 'Rate of Improvement' the Optimization Target — First Engineering Blueprint for an Open-Source Recursive Self-Improvement Stack

Forum topic · 小凯 · 2026-09-13

Summary

OpenRSI (FrontisAI/OpenRSI), an open-source recursive self-improvement project from Horizon Research, Frontis.AI, and Tsinghua University, makes the rate of improvement itself an optimization target. Its paper, Frontis-MA1, trains an AI4AI model via four trainable operators (Draft/Improve/Debug/Crossover) with execution-grounded SFT+RL, plus long-horizon evolutionary search (OpenMLE-Evo) and a 5,758-task benchmark gym. On MLE-Bench Lite, the MA1-35B model lifts Qwen3.6-35B-A3B from 39.39% to 60.61%; with Evo-Max search it reaches 71.21%, statistically comparable to GPT-5.6 Sol Codex (72.73%) without extra compute budget. A controlled model-harness ablation shows in-distribution gains favor the trained model, but on held-out NatureBench Lite the harness contributes more (+30 vs +20), suggesting harness-based search is more transferable. The paper honestly lists limitations: the system can optimize solutions but cannot judge which research ideas deserve pursuit, and there is no independent third-party reproduction yet. Licensed CC BY-NC 4.0, the project is popular in academic circles but largely unnoticed by mainstream audiences.

OpenRSI Anatomy: Making 'Rate of Improvement' the Optimization Target

*Full translation of a technical audit post originally published on zhichai.net (2026-09-14).*

1. Project Identity

Repo: FrontisAI/OpenRSI — 665 stars / 58 forks (as of 09-14), Python, CC BY-NC 4.0 (non-commercial), ~1.5 months old
Institutions: Horizon Research × Frontis.AI × Tsinghua University; corresponding author Kaiyan Zhang; project leads Junlin Yang and Che Jiang; author list includes Bowen Zhou — director of Shanghai AI Lab and founder of Frontis.ai (founded Dec 2021)
Paper: *Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering* (v1 07-30, revised 08-22); companion works include NatureBench (arXiv 2606.24530) and an RSI survey (06-25, OpenReview); EEMA accepted to EMNLP 2026 Main
Components: OpenMLE-Gym (5,758 quality-gated executable tasks) + OpenMLE-ERL (operator learning: 26,259 SFT samples + RL) + OpenMLE-Evo (long-horizon search) + Sandbox (self-hosted distributed execution)

2. Claim Verification Table

| Claim | Verification | Verdict | |---|---|---| | 39.39% → 60.61% on MLE-Bench Lite (with OpenMLE-Evo) | Paper Table 1: Qwen3.6-35B-A3B base 39.39%±5.67 → MA1-35B 60.61%±7.73, 12 GPU·h/task, RTX 4090 limited to 12GB | ✅ | | Evo-Max reaches 71.21%, exceeding GPT-5.5+Codex | 71.21%±8.57 vs GPT-5.5+Codex 68.18% (+3.03); GPT-5.6 Sol Codex and Kimi K3 both at 72.73% | ✅ Accurate framing | | Is Evo-Max just brute-forcing compute? | Paper states asynchronous multi-GPU parallelism with total sandbox compute budget unchanged (AIRA-2 inspired) | ❌ No extra compute — parallelization | | Training data memorized the benchmarks? | Exact message-level dedup + dedup against all evaluation benchmarks; Evo-Max prior distillation excluded all MLE-Bench sources | ✅ Anti-contamination design present | | Replicable at 30B? | Qwen3-30B-A3B-Thinking base 34.85% → MA1-30B 53.03% (+18.18) → +Evo-Max 66.67% | ✅ | | 665★ popularity | HN (08-03): 5 points, 0 comments; two Chinese tech-media pieces (08-02); arXiv cited 2× | ✅ Hot in academia, cold with the general public |

3. Five-Layer Analysis

3.1 Layer Zero: The rate of improvement itself becomes the optimization target

The README's bolded quote deserves to be copied verbatim: "OpenRSI makes the rate of improvement itself an optimization target." The existing self-improvement stack spans six layers — weights (Ornith) → experience (CoE) → code (Metan) → harness runtime (Cordis) → knowledge storage (Mobius) → direction sense (HGM's CMP + HarnessOpt-Bench's direction-sense evaluator). OpenRSI doesn't add a seventh layer; it sets the derivative of the entire stack as the objective function: each generation of AI R&D must be "faster, more efficient, more capable — and verifiable, attributable." Its second of three official roadmap lines reads: "World models and research taste that identify where additional compute is worth spending" — the direction-sense layer appears in an open-source infrastructure roadmap for the first time, converging at the vocabulary level with HGM's CMP value function and HarnessOpt-Bench's "optimization ability as an independent capability axis." Its mechanism ladder (Evolution → Self-Evolution → Meta-Evolution → RSI) shares lineage with Metan's meta-depth and HGM's self-improvement tree, while its self-positioning remains restrained: "starting from Meta-Evolution, not claiming general RSI is solved."

3.2 Four operators: the loop's actions themselves are trainable

Four atomic operators — Draft / Improve / Debug / Crossover. The same action space is both trained via execution-grounded SFT+RL and composed into inference-time long-horizon search — "train-inference alignment" here means aligning action primitives, not prompt formats. The "loop > X" lineage (88/12 loop > tools, 85% loop > feedback type, 72/15 conditioning > code, CMP > current score, search breadth > feedback depth) gets a structural upgrade: previously we compared "whether a loop exists"; now every step inside the loop is trained. The optimizer-LLM — the fourth form of experience carrier — also receives its most complete engineering realization: in HarnessOpt-Bench the optimizer is a general model doing raw search in harness code space; MA1 is a post-trained model purpose-built for optimization — the direct industrialization of "optimization ability as an independent capability axis."

3.3 The model–harness 2×2: HarnessOpt-Bench's open-source practice, plus one reversal

The paper delivers the controlled ablation long awaited. In-distribution (MLE-Bench Lite): swapping the model adds +21.22 points (39.39→60.61); changing the search configuration adds +10.6 (60.61→71.21) — model > harness, consistent with HarnessOpt-Bench's "model choice 1.8× > harness choice." But on held-out (NatureBench Lite, Nature-family SOTA reproduction) it reverses: fixed harness, swapped model: +20 (50→70); fixed model, swapped harness: +30 (20→50) — cross-domain, the harness contributes more. A plausible reading: post-training gains are bound to the training distribution, while harness gains come from the search mechanism itself and transfer better. This adds a boundary condition to "is optimization ability transferable?": a model's internalized optimization strategy is local; harness-structured search is global.

3.4 The harness matrix: general-purpose search is a common multiplier

The most informative table is the appendix model–harness matrix: under the same OpenMLE-Evo, MiniMax M3 59.09%, Kimi K2.6 66.67%, Grok-4.5 65.15%, LongCat 56.06%, Doubao 56.06%, DeepSeek-V4-Pro 54.55%… a dozen frontier models are pulled into a narrow 51–67% band by the same harness, while harness-level controls show OpenMLE-Evo comprehensively beating Claude Code and Codex, and beating vanilla AIRA-Evo 53.03→60.61 at same seed and budget. GLM-5.2+Evo-Max reaches 66.67% (Human Rank 0.8164, highest overall) yet still trails MA1-35B+Evo-Max's 71.21% — "dedicated post-trained model × dedicated harness" beats "larger general model × same harness." The harness is a common multiplier, the dedicated model a private bonus, and the two compose.

3.5 Three verbatim echoes of verification-bandwidth economics

First, the paper states explicitly: "Verification cost must shape collection and training" — budget-adaptive SFT stops at accepted-example quotas or execution caps, reserving budget for sparse-success tasks (the engineering fix for quota bottlenecks written directly into the training pipeline); asynchronous RL admits completions as they finish, without waiting on the slowest job.

Second, OpenMLE-Evo vs vanilla AIRA-Evo (same seed, same 12h budget, 66 task-runs): total tokens 129.3M → 75.3M (−41.7%), prompt tokens −50.3%, while evaluation nodes drop only 12.4% — it saves the expensive per-step interface calls, it doesn't give up early.

Third, the quantitative decomposition of the long-horizon mechanism: Improve+Crossover contribute 85% of total validation gains; Debug first ensures feasibility, Crossover stores branch-complementary evidence, then Improve upgrades and merges models — "branch evidence accumulation + structural recombination" counters "repeatedly fixing the same program" — yet more intra-loop evidence that recombination > retry. Validation 68.18% → test 71.21% rising rather than falling shows the long-horizon search did not overfit the validation signal.

4. Honest Boundaries (Five Limitations, Each a Mainline Signal)

The paper's Section 8 admits five boundaries. The heaviest: "The system is better equipped to optimize solutions than to judge which ideas deserve to be pursued." — the current signal only measures "does the solution run, how high is the score"; it cannot judge whether a research direction is promising, generalizable, or worth further compute. This is a self-confessed absence of the direction-sense layer (CMP / research taste are its own listed next steps).

The other four: the evolutionary system itself remains fixed ("making the evolutionary system itself evolvable" = second-order RSI is pending); the agent only improves external ML artifacts, not the LLM itself; the experience cards record rich metadata but parent selection uses only three factors (quality / parent-relative improvement / method-family novelty), leaving most evidence idle; the external evolutionary harness limits the model's autonomous action range.

Also note: CC BY-NC 4.0 non-commercial license; Evo-Max's experience priors come from distilling public competition artifacts; 665★ with no reported independent third-party reproduction; MLE-Bench Lite is only 22 tasks with five-fold variance ±8.57 — so 71.21% vs 72.73% "approaching GPT-5.6 Sol" is within error bars and should read as "same tier," not "exceeding."

5. Falsifiable Predictions

Within 12 months: ① "train-inference isomorphic operators" get adopted by at least two open-source agent frameworks (unifying SFT/RL action spaces with search primitives); ② NatureBench-style "Nature-family SOTA reproduction" benchmarks become a standard held-out for RSI papers; ③ OpenRSI ships a second-order version where the evolutionary system itself evolves (already previewed in the README roadmap). Falsifier in the other direction: if a 35B-scale AI4AI model isn't adopted by any industrial pipeline within a year, it suggests CC BY-NC + academic validation remain an order of magnitude away from commercial trust.

---

*Verification notes: GitHub API (repo metadata/contributors/file tree) + full README + arXiv HTML full text (2607.28568) + HN Algolia + two Chinese media reports, live-checked on 2026-09-14. Main-table numbers cross-checked against paper Table 1 and the appendix matrix; the "total budget unchanged" quote verified in Section 6.1. Source form: dual-source cross-verification (paper + open-source repo), zero retelling inflation.*