DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses

AI summary

Overview

Research area: AI research agents for scientific hypothesis generation and recursive self-improvement (Natural Language Processing / AI for Science). Technical level: Advanced. Scope: The paper defines what makes a research hypothesis valuable and testable, builds a multi-agent framework (DiscoveryRSI) that develops hypothesis trees and turns them into executable investigations, adds three mechanisms for learning from that research experience, and introduces DFM-Bench to measure hypothesis quality and cumulative research progress.

What This Paper Is About

Research agents can propose scientific hypotheses, but they rarely learn from the process of developing and investigating them — the record of which proposals improved, which revisions helped, and which agent actions deserved credit is usually discarded. This paper asks how an agent can turn hypothesis development and investigation into a recursive learning loop that improves its own next round of research. The authors answer with an operational definition of a valuable, testable hypothesis, a framework that preserves the full research history, and learning mechanisms that convert that history into model training, workflow edits, and new investigations.

Key Contributions

1. An operational definition and framework for hypothesis-driven research. A hypothesis is defined as a four-part object — claim, investigation plan, decision criteria, and alternative explanations — grounded in a documented evidence base and declared constraints (Definition 1). DiscoveryRSI develops these through multi-agent evolutionary search, compiles selected computational plans into executable Harbor tasks, and preserves the development histories as reusable learning experience.
2. Three coupled recursive self-improvement mechanisms. Hypothesis credit optimization attributes scientific assessments to documented contributions and converts them into bounded, weighted training supervision; harness replay diagnoses recurring failures and proposes targeted workflow edits with model weights held fixed; rediscovery applies accepted model and harness updates to fresh questions.
3. DFM-Bench, a benchmark that assesses five connected levels: individual hypothesis quality (eight-axis profile plus aggregate score), revision value relative to an actual predecessor, development-path progress, research-run quality, and scientific contribution maturity — compared in Table 1 against AgentIdeaBench, DiscoveryBench, BLADE, ScienceAgentBench, and OEB.
4. Empirical evidence that learning from research histories works, including automated quality gains across two self-improvement rounds, changes in agent behavior, expert-supervised human review with 27 participants, and a post-cutoff literature correspondence rate for DiscoveryRSI-27B.

Main Findings

• Hypothesis quality rises each round: mean automated quality goes from 4.86 (base model) to 5.18 after round one to 5.39 after round two. The second-round model beats the base model in all four life-science domains, with gains from 0.28 points (disease mechanisms) to 0.94 points (population multi-omics). The incremental gain shrinks from 0.32 in round one to 0.21 in round two.
• Gains concentrate on testability, not originality: falsifiability rises 4.00 to 6.20, alternative explanations 4.44 to 5.98, data resolvability 3.73 to 4.99, and feasibility 4.79 to 5.87. Meanwhile importance declines 6.48 to 6.04 and novelty 4.87 to 4.58.
• Behavior shifts toward developing existing hypotheses: evolve/repair operations rise from 77.5% of episodes (base) to 90.0% (round one) and 91.7% (round two), while restarts fall from 21.7% to 5.8% and 5.0%. In episodes 11–20, restarts drop from roughly 32% to 3% after round two.
• Development paths become more evidence-driven: the largest path-score gain is evidential and inferential discipline (+1.25), then technical and knowledge exploration (+0.34) and scientific continuity and synthesis (+0.24). Cumulative progress, stopping decisions, and budget efficiency change by less than 0.1 points, and disciplinary connections decline by 0.28.
• Human reviewers see more improvement per revision: within assessed research trees, 41% of DiscoveryRSI-27B revisions raise the mean human score, 47% leave it unchanged, and 12% lower it, versus 31%/37%/32% for GPT-5.6 Sol and 25%/44%/31% for GPT-6 Luna. Mean change per revision is +0.07 for DiscoveryRSI-27B, +0.03 for GPT-5.6 Sol, and −0.01 for GPT-6 Luna.
• Absolute human ratings are mixed: DiscoveryRSI-27B scores higher than GPT-6 Luna on importance (3.10 versus 2.94) and on plausible, evidence-bounded claims (3.37 versus 3.16); both receive 2.31 for advance over known work, while GPT-6 Luna scores higher on data fit (the provided content is truncated at this point in the criteria comparison).
• Hypotheses match post-cutoff published research more often: DiscoveryRSI-27B reaches a 6.3% correspondence rate with research published after the model cutoff, versus 2.5% for GPT-6 and 2.3% for Claude Opus 5.
• The battery setting replicates the pipeline: 200 episodes and 351 generated hypotheses with three AI reviews each yield a mean hypothesis quality of 5.61 on 165 complete scores, substantive progress on 214 of 366 transitions (58.47%), a mean complete-trajectory score of 81.52 across 152 paths, a research round score of 71.25, and 20 contribution families, 7 of which carry diagnostic or research evidence.

Methodology in Plain English

The researchers first wrote down what a scientific hypothesis actually needs in order to be useful: a claim tied to evidence, a concrete plan for testing it, criteria decided in advance for what results would mean, and rival explanations that could also account for the outcome. They treat a hypothesis as that whole package, not just a sentence.

Agents then build these packages as research trees. A Lead agent picks directions and operations; Generation and Evolution agents dig into evidence and publish proposals or revisions; a Blinded Critique compares scientific merit; a Progress reviewer judges whether a revision is real progress, a necessary repair, repetition, or unsupported. Each hypothesis version keeps its parents, its evidence, its drafts, and a record of which agent actions produced it. Selected computational plans are compiled into Harbor tasks — runnable packages with defined inputs, environment, required outputs, and verification checks — while unavailable data and physical experiments stay listed as explicit dependencies.

Learning happens along three routes. Hypothesis credit optimization uses pairwise critique results scored with TrueSkill to rank hypotheses, then traces credit back to the specific recorded agent invocations that contributed, with caps and duplicate suppression. That credit weights a training loss applied only to the assistant responses and tool-call arguments that the record supports; tool outputs and known failed actions stay in context but get no direct loss. This produced DiscoveryRSI-27B via full-parameter fine-tuning with a 256K-token context window: from 15,792 collected research trajectories, filtering yielded 2,397 training trajectories (634 Lead, 861 Generation, 902 Evolution), totaling 291.12 million tokens of which 6.18 million (2.12%) receive direct supervision, trained for one epoch with Adam at a learning rate of 5×10⁻⁶, global batch size 16, in BF16.

Harness replay leaves the model weights fixed and instead edits the research workflow — coordination, evidence access, review timing, work allocation — based on short diagnostic runs that expose repeated exploration, failed tool use, lost artifacts, or premature closure; candidate edits are screened with quick runs and replay, and only accepted after matched development-set evaluation. Rediscovery then runs the updated system on fresh questions, generating new experience. Model and harness updates are accepted separately or jointly on a development set, and final numbers come from untouched test families.

Evaluation combines 241 research questions judged automatically by GPT-5.6 Sol, behavioral analysis of six campaigns of 20 episodes per model variant, expert review of 292 proposals (HCC clinical cohort 52, MIMIC-IV 60, eICU 60, UK Biobank 55, GEO targets 35, GEO mechanisms 30) by 27 medical doctors, medical students, and bioscience PhD students under three senior supervisors, and a literature correspondence check against a shared reference corpus with model-specific cutoffs.

Why This Matters

Most AI-for-science systems are judged on a single output — a hypothesis, a paper draft, a plot — and throw away the reasoning trail that produced it. This paper argues that the trail itself is the training signal, and shows a workable way to convert assessed research quality into bounded supervision without letting one lucky proposal dominate training. It also reframes benchmark design: instead of only scoring ideas, DFM-Bench scores revisions, development paths, whole runs, and contribution families, which is closer to how research actually accumulates.

Real-world applications:

• Clinical and population-health study design: the paper's worked example asks whether repeated creatinine–urine output mismatch adds information about ICU readmission or in-hospital death within 48 hours of ICU discharge, using MIMIC-IV, eICU, and UK Biobank — the kind of question hospitals must triage before committing analyst time.
• Therapeutic target prioritization and disease-mechanism research, where proposals must specify comparisons, controls, and interpretation criteria before data are touched.
• Population multi-omics and public dataset reuse (GEO, ArrayExpress), where a key failure mode is proposing analyses the available data cannot resolve — which is precisely the axis (data resolvability) that improves most.
• Battery and materials diagnostics, demonstrated as a transfer setting covering electrical diagnostics, health prediction, degradation mechanisms, and usage history and aging, with checks for operating conditions, sensing, and physical plausibility.

Industry relevance: teams running automated R&D pipelines, drug-discovery or diagnostics groups evaluating AI-generated study proposals, and organizations building agentic research tools can reuse the framework's core idea — log contributions, attribute credit, and gate workflow changes on matched evaluation — independently of the specific life-science domains tested.

Future Directions

• Close the importance/novelty gap. Testability improved sharply while importance (6.48 to 6.04) and novelty (4.87 to 4.58) declined. Understanding whether recursive self-improvement systematically trades ambition for rigor is an open question the paper surfaces but does not resolve.
• Improve exploration and resource allocation. Development-path gains are concentrated in evidential discipline; cumulative progress, stopping decisions, and budget efficiency move by less than 0.1 points, and disciplinary connections decline by 0.28.
• Extend beyond the tested settings. The paper reports four life-science domains plus a battery setting; whether the credit-attribution and harness-replay mechanisms transfer to other scientific fields, and to physical experiments currently left as explicit dependencies, is untested here.
• Separate proposed from supported contributions. DFM-Bench's contribution layer distinguishes proposed ideas from contributions backed by completed investigations and independent checks; the paper notes that compiled tasks are measured for execution readiness separately, leaving full execution-grounded evaluation as a next step.

Target Audience

Researchers and engineers working on AI research agents, agentic self-improvement, and LLM post-training who want a concrete recipe for turning multi-agent research traces into supervision; benchmark designers interested in evaluating scientific process rather than single outputs; and domain scientists or research-ops leads evaluating whether AI-generated study proposals are specific and testable enough to act on.

Authors’ abstract

Scientific discovery requires developing worthwhile hypotheses into practical investigations, yet learning from this evolving research process remains challenging. We introduce an operational definition of valuable and testable research hypotheses: evidence-grounded claims about consequential uncertainties, coupled with feasible investigations, explicit decision criteria, and alternative explanations. Our DiscoveryRSI framework develops hypothesis trees through multi-agent evolutionary search, compiles selected investigations into executable Harbor tasks, and recursively learns through three mechanisms: hypothesis credit optimization attributes scientific assessments to documented research contributions; harness replay identifies productive strategies and recurring failures to refine research workflows; and rediscovery applies accepted improvements to new investigations, generating further experience for subsequent learning. We also introduce DFM-Bench, which provides comprehensive assessment of hypothesis quality and cumulative research progress at the levels of hypotheses, revisions, development paths and research runs. Across four life-science domains, automated evaluation shows that hypothesis quality rises from 4.86 for the base model to 5.39 after two rounds of self-improvement, with the large gains in falsifiability and data resolvability. We further validate hypothesis quality through expert review, comparing our DiscoveryRSI-27B with other research agents. Preliminary human assessments show that DiscoveryRSI-27B scores higher than GPT-6 Luna on hypothesis importance and whether claims are plausible and evidence-bounded. DiscoveryRSI-27B achieved a 6.3% rate of hypothesis correspondence with research published after the model cutoff, largely outperform GPT-6 with 2.5% and Claude Opus 5 with 2.3%, while delivering greater human-rated improvement along research paths.