DiscoveryRSI applies recursive self-improvement to scientific hypothesis trees across four life-science domains

Most science agents still invent a claim and hope a bench agrees. DiscoveryRSI keeps the whole tree of revisions and teaches the next round from that history.

The paper's bet is simple enough to argue with: recursive self-improvement belongs on hypothesis development itself, including how investigations are planned and scored, before anyone celebrates a new binder or DNA language model.

Hypothesis trees plus Harbor tasks become the RSI substrate

DiscoveryRSI defines a valuable, testable hypothesis as a joint object: an evidence-grounded claim about a consequential uncertainty, a feasible investigation, explicit decision criteria, and alternative explanations. Multi-agent evolutionary search grows trees of those objects. Selected computational investigations compile into Harbor tasks with inputs, environments, required outputs and verification checks. Research histories then drive three coupled learning mechanisms. Hypothesis credit optimisation attributes scientific assessments to documented contributions and supplies bounded training signal for DiscoveryRSI-27B. Harness replay keeps the model fixed while diagnosing recurring failures and testing workflow edits. Rediscovery applies accepted model and harness updates to fresh investigations so the loop has new experience.

That package is the novelty claim. The system learns from how hypotheses were revised, which controls were added, and which plans failed, rather than from a final abstract alone.

Quality scores, expert path gains and post-cutoff literature match rates

Automated evaluation across four life-science domains lifts mean hypothesis quality from 4.86 for the base model to 5.39 after two RSI rounds. The authors highlight falsifiability and data resolvability among the largest component gains, and they report evolve/repair actions rising from 77.5% to 91.7% of episodes as quality climbs.

Expert-supervised review finds a mean human-rated improvement per revision of +0.07 for DiscoveryRSI, against +0.03 for GPT-5.6 Sol and −0.01 for GPT-6 Luna. DiscoveryRSI-27B reaches a 6.3% rate of hypothesis correspondence with research published after the model cutoff, where correspondence means a linked post-cutoff publication that supplies a concrete design change, comparator, measurement or testable evidence gap. GPT-6 sits at 2.5% and Claude Opus 5 at 2.3% on that same metric. DFM-Bench scores hypotheses, revisions, development paths, research runs and contributions, with Harbor execution readiness measured separately.

Same domains, automated plus expert lenses, and a narrow match definition

The 4.86 to 5.39 lift is within-system across RSI rounds, which is the right shape for a self-improvement claim. Human path-gain comparisons put DiscoveryRSI beside Sol and Luna under expert review rather than against an anonymous leaderboard. The 6.3% post-cutoff correspondence rate uses a strict linking rule described in the paper's appendix; it is a signal of actionable overlap with later literature, and it is still a small absolute percentage. Readers should avoid treating it as a replication rate or a claim that the agent predicted those papers. Domain coverage is four life-science areas chosen by the authors; fairness across chemistry, materials or clinical trial design is outside the reported tables.

Wet-lab closure, regulatory grade evidence and open-ended grant competition

DiscoveryRSI does not show that Harbor-compiled investigations replace institutional review boards, wet-lab confirmation or journal peer review. It does not measure whether recursive loops remain stable when hypothesis credit is noisy or gamed. Cost, wall-clock time and human supervisor hours per accepted improvement are only partly characterised for buyers who need a staffing model. Cross-domain transfer from the four evaluated life-science settings into unrelated fields remains unsettled. The paper also leaves open how often rediscovery proposes investigations that are ethically or biosafety constrained even when they are computationally well-formed.

arXiv paper, Gen-Verse repository and DFM-Bench

Primaries are arXiv:2610.23839 and the Gen-Verse DiscoveryRSI repository named in the paper. DFM-Bench and any released trajectories or Harbor task packs should be taken from those links as published. Licence and weight availability follow the repository; if a checkpoint named DiscoveryRSI-27B is gated or delayed, say so from the repo state rather than assuming Hugging Face presence.

Use DFM-Bench language before adopting RSI as a lab process

Research-engineering leads who already run multi-agent literature scouts can adopt DFM-Bench's hypothesis checklist (claim, investigation, decision criteria, alternatives) as a review rubric this month, then decide whether Harbor compilation and credit optimisation are worth the harness investment. Labs comparing "AI scientist" vendors should ask for post-cutoff correspondence methods and human revision-gain numbers in the DiscoveryRSI shape, rather than only accepting fluency demos. Skip a full RSI deployment if you cannot retain contribution-level histories; without them, credit optimisation has nothing honest to train on.

Distinct from Rosalind agents, Carbon-X and RFD4

Frontier has already covered Rosalind-style scientific agents, Hugging Face Carbon-X DNA modelling and NVIDIA/UW RFD4 protein binder work. Those pieces are about binders, sequences or agent products in the wet-facing stack. DiscoveryRSI's object is the hypothesis tree and the learning signal inside research management. It sits closer to Co-Scientist and AI Scientist-v2 lineage on evolutionary hypothesis search, then adds Harbor compilation plus the three RSI mechanisms. File it under how agents choose what to investigate, and keep the binder and DNA stories on their own spines.