rsi.watch

Binghang Lu, Yige Sun and Guang Lin · 2026-10-23

RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training

“With Opus 5.5, RSI-Opt raises the training-suite score from +10.48 to +14.33, a summary-score difference of +3.85.”
Who
Binghang Lu, Yige Sun and Guang Lin
Date
2026-10-23
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ risk of bias. The quoted score is on the three tasks whose scores guided evolution and picked the final policy. Fresh seeds separate the reported runs from the search, but the tasks are the same. “Neither step makes the evolution tasks unseen.”
  • ↓ imprecise. Three final seeds per cell. The comparison reports differences of means, with no confidence intervals. “With three final seeds per cell, uncertainty estimates have limited precision.”
  • ↓ indirect measure. The gain is measured against the fixed-prompt agent. No control separates inheritance across generations from independent proposals at the same budget. “We do not isolate the effect of recursive inheritance against a matched-budget search that always proposes from generation 0”
how we rate evidence
Replication
No rerun found. The paper says the harness, policies, transcripts and results are released but gives no link. 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

Also on the page

  • An in-run agent edits the optimizer code while a model trains. A meta-agent, Opus 5.5 during evolution, rewrites the policy that agent starts each run with: a prompt, a playbook, a code library and a decision cadence. The LLM weights, the meta-agent's instructions and the scoring rule stay fixed.
  • Evolution ran 5 generations of four children on three training tasks: ResNet-18 on CIFAR-10, a 34M-parameter GPT on FineWeb and a flow-matching UNet on MNIST. Final numbers use fresh seeds. Six other tasks, three ML and three PDE, are held out.
  • On the six held-out tasks, Opus 5.5's aggregate score is lower with the evolved policy than with the fixed prompt, +4.87 against +5.22, though "It improves the mean metric on four of the six held-out tasks". The same policy, used unchanged with GPT-5.5, takes the held-out score from -0.36 to +3.11, and with GPT-6.1 from +0.80 to +4.30.
  • Before filesystem restrictions were added, one meta-agent read the description of a held-out task; that policy and its descendant were excluded from parent selection and final selection. The authors add that "direct-read and ancestry checks alone do not establish the absence of indirect exposure", calling this "a limitation of the current held-out evaluation claim, not evidence that such indirect propagation occurred."
  • The authors: "We do not isolate the effect of recursive inheritance against a matched-budget search that always proposes from generation 0".

In the self-improving AI catalog

  • RSI-Opt: what it changes, whether the change lasts, and how it was tested

How it got here

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

  1. 2025 Darwin Gödel Machine uses it

    Tools, own code · becomes the next version · improves the improver → improves the task agent

    “An archive supports continued exploration and inheritance, following the archive-based approach of the Darwin Gödel Machine [35].”
    “The outer loop maintains an archive of evaluated policies, following the archive-based search of the Darwin Gödel Machine [35].”

  2. 2026 RSI-Opt

    Tools, own code · becomes the next version · improves the task agent

Also cites Reflexion 2023, Voyager 2023, Promptbreeder 2023, STOP 2023, ADAS 2024, AlphaEvolve 2025

← Ledger Report an error