rsi.watch

Haojin Wang et al. (9 authors) · 2026-10-25

ParetoRSI-Gym: Can AI Agents Improve Language Models Across Multiple Domains?

“We also observe that the agent systems tend to peak before they stop: in seven of the nine configurations the proxy-best round is not the last one, indicating current agents lack the capability to robustly continue improving the model.”
Who
Haojin Wang et al. (9 authors)
Date
2026-10-25
Evidence
Moderate certainty Starts at High: measured by someone other than the developer.
  • ↓ imprecise. The average gain is 2.3 points. Standard deviations across seeds are reported only on Qwen3-4B-Base, where most lie between one and three points. “Most standard deviations lie between one and three points”
how we rate evidence
Replication
No rerun found. The code is on GitHub. 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

  • Agents pick training data for five domains (math, code, finance, medicine, science) and fine-tune a separate base model over five rounds. Final scores come from benchmarks the agent never sees during the run.
  • The best system, Claude Code with Opus-5, raised generalization scores at its proxy-best checkpoint by 9.2% relative to the starting model, +2.3 absolute points, averaged over three base models and five domains.
  • With one backbone (DeepSeek V4.1 Flash Max) on Qwen3-4B-Base, one run per harness, changing only the harness moved the outcome from +4.3 points to +0.2.

In the self-improving AI catalog

  • ParetoRSI-Gym: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that ParetoRSI-Gym cites, oldest first. Under each: where that work sits on the map, then where ParetoRSI-Gym sits (changes in bold), and what ParetoRSI-Gym'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 · does not improve the improver

    “For instance, PostTrainBench (Rank et al., 2026) explores how well LLM agents can perform post-training autonomously under bounded compute constraints.”
    “This differentiates ParetoRSI-Gym from existing benchmarks such as PostTrainBench (Rank et al., 2026) and RSI-Bench Data (Meng et al., 2026) which evaluate the ability of agents to perform post-training on the same benchmark.”

  2. 2026 ParetoRSI-Gym

    Prompts, memory · gone when the run ends · does not improve the improver

Also cites The AI Scientist 2024, AlphaEvolve 2025, Darwin Gödel Machine 2025, Hyperagents 2026, RRSI 2026

← Ledger Report an error