rsi.watch

Cheikh Ahmed et al. (5 authors) · 2026-10-17

Organon: From Evolutionary Search to Recursive Self-Improving Agents for Scientific Discovery

“On ALE-Bench Lite, Organon achieves the best performance among competing search frameworks on all 10 tasks, improving mean held-out performance by 20.5% over EvoX.”
Who
Cheikh Ahmed et al. (5 authors)
Date
2026-10-17
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ imprecise. One Organon run per task. ALE-Bench values are means of three evaluations of one selected program, with no spread reported. On EPLB and Cloudcast the margins over EvoX are about 0.136% and 0.0045%, and the authors say variability is an important part of reading the ranking. “The analysis covers one run per task, and revisions within a run were not applied in isolation.”
  • ↓ indirect measure. The comparison is with other search frameworks. The paper reports no ablation with self-revision turned off; it lists repeated runs and controlled ablations as future work. “so temporal association between a revision and later progress does not identify its causal contribution.”
how we rate evidence
Replication
No rerun found. Code is linked from the paper. We found no rerun by another group. Searched up to 2026-10-28. all claims
Checked
2026-10-28, against the source page
Links
Source Paper Code

Also on the page

  • One agent runs the search for a task. Between bounded search calls it revises its own evolutionary loop, parent selection, variation, survival rule, diagnostic tools, memory and worker coordination, and later calls use the revised versions. The task evaluator and the budget of 100 candidate attempts stay fixed.
  • All four compared methods used DeepSeek-V4.1-Flash with at most 100 attempts per task; on three tasks Organon's agent sent 38 of 300 candidates to DeepSeek-V4-Pro, "which ranks below DeepSeek-V4.1-Flash on independent evaluations". On third-order autocorrelation and AHC046 the final best came from a V4-Pro worker. On ALE-Bench Lite the mean private score is 2478.6, against 2056.3 for EvoX. On KernelBench it is fastest on three of four workloads; on ConvMax EvoX leads, 2.75x against 2.29x.
  • On the three math tasks its values fall short of the reference constructions released with AlphaEvolve, for example 2.635314 against 2.635863 on circle packing. Among the four search methods it is best on third-order autocorrelation, while EvoX and ShinkaEvolve do better on circle packing and all three others do better on Heilbronn.
  • The authors' limits: "The analysis covers one run per task, and revisions within a run were not applied in isolation." Carrying capabilities from one task into the next is listed as future work.

In the self-improving AI catalog

  • Organon: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that Organon cites, oldest first. Under each: where that work sits on the map, then where Organon sits (changes in bold), and what Organon's own paper says about it. Its whole family on the citation map

  1. 2026 Dream-RSI sets itself apart

    Tools, own code · becomes the next version → gone when the run ends · improves the task agent → improves the improver

    “EvoX evolves strategies for parent selection, variation, and inspiration-context construction (Liu et al., 2026b), while Dream-RSI refines exploration policies through replay over recorded discovery trees while keeping its discovery agent fixed (Zheng et al., 2026).”
    “Because these approaches confine adaptation to a predetermined subset of components and fixed update procedures, their adaptive scope is bounded at design time.”

  2. 2026 Organon

    Tools, own code · gone when the run ends · improves the improver

Also cites STOP 2023, The AI Scientist 2024, AlphaEvolve 2025, Darwin Gödel Machine 2025, Hyperagents 2026, Weco AIDE² 2026, Recursive Harness Self-Improvement 2026, RRSI 2026

← Ledger Report an error