rsi.watch

Yue Huang et al. (18 authors) · 2026-10-24

RSI Arena: Towards Recursive Self-Improvement with Human Feedback

“All ten agents produce working assistants, with the strongest model, trained by GPT-6 Astra, closing a median of 76% of the base-to-Instruct gap across 12 capability benchmarks.”
Who
Yue Huang et al. (18 authors)
Date
2026-10-24
Evidence
Low certainty Starts at High: measured by someone other than the developer.
  • ↓ indirect measure. The agents trained a separate model, on one base model. The step where an improved model does the next round of research was not tested. “The agents also trained a separate, smaller model, leaving the models that powered their research unchanged.”
  • ↓ imprecise. One run per agent. “Each agent ran once, on a single base model”
how we rate evidence
Replication
No rerun found. The agents' checkpoints are on Hugging Face. 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

  • Ten frontier models each post-trained the same base model, NVIDIA Nemotron 3.5 Lightning 30B-A3B Base, for six days with 1,000 GPU-hours and $300 in API credit. The models that ran the agents were not changed.
  • In Stage 2 each agent received the arena battles its model had played, plus three more days. Eight of the ten narrowed their model's style-controlled gap to NVIDIA Instruct, and the median benchmark change was +2 points of the gap. Stage 2 also gave more time and resources and used a separate arena, so in the authors' words "these comparisons do not isolate the effect of human feedback."
  • In Stage 1, without style control, people rated NVIDIA Instruct highest (1,182, against 1,161); "with answer length and formatting controlled, the two are level at 1,095 and 1,094." Three Stage 1 models gave more harmful help than the base model on StrongREJECT.
  • NVIDIA Instruct, the reference for the 76% figure and the ratings, was run "with reasoning off".
  • The authors: "Each agent ran once, on a single base model, so the experiment establishes feasibility more clearly than reliability."

In the self-improving AI catalog

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

How it got here

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

  1. 2026 PostTrainBench sets itself apart

    Prompts, memory · gone when the run ends → saved for later runs · does not improve the improver

    “Closest to our work are agents that post-train models (Yano et al., 2025; Ma et al., 2026; Yu et al., 2026; Chen et al., 2026), above all PostTrainBench, where an agent post-trains a small base model for one named benchmark in 10 hours on one GPU (Rank et al., 2026).”
    “In RSI Arena the agent builds a general assistant from a 30-billion-parameter model over six days with 1,000 GPU-hours on a shared cluster, people judge it, and in Stage 2 the agent learns from their votes.”

  2. 2026 RSI Arena

    Prompts, memory · saved for later runs · does not improve the improver

Also cites Good, the intelligence explosion 1965, The AI Scientist 2024, Gödel Agent 2024, SICA 2025, AlphaEvolve 2025, Darwin Gödel Machine 2025, Ren, self-improvement 2026, RRSI 2026, Sakana MASS 2026

← Ledger Report an error