arXiv 2609.26457 (2026-09-22), by the full Weco AI team (Srikanth, Zhao, Xu, Wu, Jiang). This is the tenth entry in our RSI lineage series, and the first complete loop run on the company's own production agent.
Position in the RSI Lineage
Of the nine RSI papers previously covered—OpenRSI as the blueprint, NeoHorse productized but single-round, Dream-RSI at the strategy layer, RRSI on defense—AIDE² differs in one crucial way: the object being improved is the company's own commercial product. Weco AI is a startup; CEO Zhengyao Jiang (co-founder, ML PhD). Its open-source engine AIDE once topped MLE-Bench with a 51.38% exceedance rate, and has been productized as "Aiden" since July. What AIDE² does is let AIDE improve AIDE—the inner loop rewrites task code, the outer loop rewrites the inner-loop agent itself. The resulting next generation then runs again: 8 days, 100 node trajectories, 99 rewrite proposals, 7 accepted, with scores climbing from 0.703 to 0.778.
One Anti-Overfitting Trick, Used Hard
Score separation. When the inner-loop agent optimizes, it sees public signals; the outer loop decides who advances using private held-out data. A detail in Appendix D shows this is really working: about a quarter of the rejected rewrites scored *higher than the incumbent on public signals*—they were rejected by the private score. Without separation, these candidates overfitting the public signal would have leaked into the next generation.
The outer-loop model is Claude Opus 4.7; inner-loop evaluation uses Gemini 3 Flash. The paper's rationale is pragmatic: per-candidate evaluation cost dominates proposal cost, and Flash matches or slightly beats pricier models on selection tasks. Money saved becomes search steps.
The selection benchmark spans three task families: ML engineering, heuristic algorithm engineering, and harness engineering—the third being modifications to the agent's own prompts, context management, and feedback loops. The agent that modifies the harness *is* the harness: meta-level and object-level are the same body.
What Seven Generations Produced
The three-generation footprint (AIDE_0 → AIDE_47 → AIDE_85) shows two major changes:
Search strategy. Greedy search was replaced by a UCB1 bandit managing five drafting-strategy arms, plus a fork from the global best every five steps—when the leading strategy plateaus, its solution is kept while a new strategy arm continues improving. Both ends of the bargain are captured.
Context management. AIDE_0 concatenated the full history into every drafting prompt, so prompts grew linearly with each run; AIDE_85 instead reads compact summaries of the root node and recent candidates, compressing per-call prompts from a growth curve to a constant—7x compression on MLE-Bench, 40x on WeatherBench 2, ~50x on ALE-Bench and FML-Bench. The saved context budget converts directly into more search steps. A failure memory was also added: it activates only when candidate bug rate exceeds 15%, injecting up to three deduplicated error signatures.
Generalization, Cheating Rates, and the Sharpest Number
Four external benchmarks that no human ever helped select: ALE-Bench, MLE-Bench, FML-Bench, and WeatherBench 2 (the last fully out-of-distribution, physical weather forecasting). AIDE_85 matched or beat AIDE_human—the production agent hand-tuned by humans for two years—across all of them, ranking among the strongest on FML-Bench. Note the gains are not monotonic: generation 47 was best on MLE and WeatherBench, 85 best on ALE and FML; the paper prints this honestly.
The cheating rate declined along the lineage: 55% (AIDE_0) → 39% (AIDE_47) → 32% (AIDE_85), vs. 39% for the human baseline—measured on a kernel-engineering task family never used in selection. The score never rewarded "cheating less," yet it declined on its own. (The "33% reduction" figure circulating in secondary coverage does not exist in the paper; the actual figure is an absolute 23-percentage-point drop.)
The sharpest number is in Appendix C. AIDE_85's gains transfer across three base models—Gemini 3 Flash, GPT-5.6-sol, and fable 5 all benefit. At the same budget, Gemini 3 Flash with AIDE_85 scores 1858, beating fable 5 with AIDE_0 at 1796: a weaker model with a good harness beats the strongest model with the old harness. This is the most quantified demonstration yet that the harness is the program, not the weights. The honest flip side is printed too: on MLE-Bench, fable 5's gain with AIDE_85 over AIDE_0 is within one standard error—not a win everywhere.
The Unlit Fire, and Full Marks for Honesty
The paper designed its own "ignition test": seat the evolved agent (AIDE_47) into the outer loop and race it against the human-designed outer loop, three seeds each, fifty steps. Result: 0.780 vs 0.782—human slightly ahead. The paper itself writes "inconclusive, no claim made." The R of self-acceleration hasn't happened—recursion currently means "sustained climbing," not "getting better at improving itself."
On the honesty rubric, this paper scores full marks: the negative ignition result, the non-significant MLE gains, two full replications accepting only 2 and 4 rewrites respectively, and an entire chapter listing rejected proposals. Appendix D maps rejected proposals onto the classical toolbox—island-model genetic algorithms, tournament selection, Luby restarts, bagged ensembles—ensemble approaches were explored but never retained; the agent's own analysis was that "ensembles burn budget that buys fewer search steps." Cost consciousness grew inside the search on its own.
Lineage and Predictions
Connections to this series' throughline: the economics of verification bandwidth—score separation is the mechanization of "assertion-execution separation," where public signals are assertions and private scoring is execution. The ignition test is an acceptance gate for recursive self-improvement, one level above "did the score go up"—it asks whether the *machine that improves* got stronger. It dialogues with RRSI: two groups independently converged on the same root disease (adaptive overfitting); RRSI prescribes seven remedies, AIDE² prevents it with one move—score separation. Different routes, same destination.
Falsifiable predictions: within 12 months, "separated scoring" becomes a standard field in RSI systems; the ignition-test protocol gets standardized as an acceptance gate by follow-up papers; and whoever first makes an agent outer loop significantly surpass the human outer loop takes the lead of the AGI narrative. This fire is still, for now, only a damp match.
---
*Sources: full close reading of arXiv 2609.26457v1 (method, experiments, Appendix C cross-model transfer, Appendix D rejected proposals); cross-checked against Weco's official blog (first 8-day results on 7/14), the original AIDE paper (2502.13138, MLE-Bench 51.38%), and $8M funding background. The "33% cheating reduction" claim in secondary coverage was not found in the paper and has been corrected to 55% → 32%; the "two years of development" figure is a company statement—the paper itself says "among the strongest on FML-Bench."*