rsi.watch

Weidong Tang et al. (20 authors) · 2026-10

The Trinity: Capability-Driven Recursive Self-Improvement for Multimodal Agents

“The first iteration brings the largest gain; iterations 2–4 add another 9.4 Train and 6.4 Unseen points.”
Who
Weidong Tang et al. (20 authors)
Date
2026-10 (month only) The page says "Technical report, October 2026" and gives no day.
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ indirect measure. All RSI variants use one frozen backbone, Claude Opus 5; Codex and DeepSeek appear on the page only as reference curves. “Every variant starts from the same Base Harness H₀ and uses a frozen Claude Opus 5 backbone.”
how we rate evidence
Replication
No rerun found. No paper or code is public yet. We found no rerun by another group. Searched up to 2026-10-26. all claims
Checked
2026-10-26, against the source page
Links
Source

Also on the page

  • Three roles (Creator, Explorer, Evolver) share one frozen backbone, Claude Opus 5, and work through separate Harnesses.
  • Base Harness: 39.06 on Train and 37.45 on Unseen benchmarks. At iteration 1 with evolved modules plus skills: 63.80 and 53.53; modules only: 56.95 and 47.26 (Table 2). There are 28 training and 18 unseen benchmarks; Figure 5's Unseen curve covers the 16 non-robot benchmarks.
  • For the two real robots, the page says the comparisons "do not isolate module evolution from skill preparation".
  • The page links no paper or code. It says the citation "will be updated with the arXiv identifier once the paper is public".

← Ledger Report an error