Recursive self-improvement (RSI) keeps hitting two very concrete walls: building a batch of heterogeneous, verifiable environments for agents is expensive, and the Harness — the engineering shell wrapped around the model — cannot be iterated by hand fast enough to match exploding task diversity. The Trinity (arXiv:2610.36077, cs.AI, posted October 26) offers a heavily automated answer: leave the model weights completely frozen, and let the framework grow its own "capability tree". The paper shot to #1 on the Hugging Face Daily Papers board the same day (233 upvotes).
A frozen Claude Opus 5, everything else moves
The backbone is a frozen Claude Opus 5 (Appendix D: adaptive thinking at low effort throughout); all improvement happens at the Harness layer, which the paper defines as six control-flow protocols around the frozen model. Three roles split the work: Explorer runs the training tasks once with the base Harness and logs trajectories; Evolver extracts shared requirements from verified trajectories and reorganizes the task tree — originally grouped by surface category — into a hierarchical capability tree; Creator abstracts tasks at the same node into simplified, verifiable simulated worlds for capability practice, explicitly excluding target-task layouts and solutions to prevent leakage. The final tree has 11 capability nodes and 10 Harnesses, and Creator's shared module library accumulates 25 modules. When a new task arrives, the Capability-Aware Harness Router picks a compatible configuration from public task information. The paper calls this Capability-Driven Real2Sim2Real: train in simulation, measure transfer on the original tasks.
The scoreboard across 44 benchmarks
The benchmark inventory has 44 entries across 5 domains and 15 subclasses, covering rule discovery, games, GUI manipulation, and embodied interaction; the abstract's accounting is 28 training plus 18 unseen benchmarks. The base Harness H0 averages 39.06/37.45 (Train/Unseen); after four RSI iterations the paper reports a mean +34 on training and +26 on unseen, including two real-robot embodied tasks (SOON E01, CaP-X C-01). Per-benchmark: between iterations 1→4, ARC-AGI-3 climbs from 41.0 to 53.6, SciGym from 45.2 to 57.1, BuildArena from 93.3 to 97.8; the final unseen mean beats the Codex (47.5) and DeepSeek (46.2) reference Harnesses.
Two structural findings deserve their own paragraph. First, organization: among five schemes, six flat capability groups are cheapest to train (136.94M tokens) but score lowest (49.25/39.43); organizing capabilities into a tree and practicing inside Creator-generated worlds scores highest (56.95/47.26) at the cost of 405.70M training tokens — trading "where to practice" against "how well it sticks". Second, routing: the full self-routing setup reaches 71.27/60.96, closing in on the retrospective-best Oracle (73.16/63.61), meaning the router almost always picks the right branch.
Three buckets of cold water
First, the gains are uneven: Mastermind sits at 88.9 across all four iterations, and the rule-discovery game Eleusis slips from 82.2 to 81.1; Crafter, after rising to 60, still trails the Codex reference at 82.7 — the paper itself names Crafter and DiG-bench as below both comparators. Second, the evaluation protocol runs five preparation trials per benchmark–Harness pair, summarizes their trajectories into skills, and only then freezes scoring — the gains embed a "cram-before-the-exam" adaptation component, a point HF community discussion also flags. Third, reproducibility is currently on paper only: the GitHub repo holds 3 commits and 14 stars, with the README promising code "within one week of October 28, 2026"; the paper is signed by the team name The Trinity Research Team with no affiliations listed (the HF paper page tags National University of Singapore, first author Weidong Tang, 21 authors).
So what
The Trinity's real claim is not the scores but shifting the optimization target of RSI from "tasks" to "capabilities": tasks expire, capabilities compound across tasks. Anthropic has publicly discussed the governance of recursive self-improvement, and a frozen-weights, self-evolving shell is precisely a gradual path that requires no model retraining. For engineering teams the direct lesson: before pouring money into fine-tuning, making your Harness modular and well-routed may be the cheaper first-stage rocket. Code lands within a week; whether third parties can reproduce this capability tree is the real threshold between a technical report and a methodology.