DiscoveryRSI: Framework for Scientific Hypothesis Discovery

• DiscoveryRSI is a framework for recursively generating, evaluating, and improving testable scientific hypotheses, characterized by structured research objects and multi-agent system.
• The method uses a combination of evolutionary search, evidence grounding, and feasibility assessment to enhance the quality and-testability of hypotheses, with practical applications in life sciences and beyond.
• DiscoveryRSI distinguishes between scientific progress and execution success, offering a comprehensive approach to hypothesis generation across multiple research runs

DiscoveryRSI is a framework for recursively learning to discover valuable and testable scientific hypotheses. It represents a hypothesis as a structured research object comprising an evidence-grounded claim, an investigation plan, explicit decision criteria, and relevant alternative explanations, denoted by $\mathcal H=(c,\pi,d,a)$. The framework combines multi-agent evolutionary search, persistent research records, executable Harbor tasks, hypothesis-level credit assignment, research-harness replay, and rediscovery on new research questions. Its central premise is that scientific progress depends not only on generating plausible ideas, but also on preserving the development process through which claims are grounded, revised, tested, assessed, and improved (Zhong et al., 16 Oct 2026).

1. Conceptual foundations and objectives

DiscoveryRSI defines a valuable hypothesis relative to an evidence base $\mathcal E$ and a set of constraints $\mathcal B$. A hypothesis is valuable when resolving uncertainty in its claim could make a consequential contribution to knowledge or change a justified research decision. It is testable when its investigation plan is feasible under the available data, measurements, computational resources, and experimental conditions, and when possible outcomes can support, weaken, or bound the claim.

The framework operationalizes five requirements:

1. Evidence grounding: the claim must be connected to documented literature, observations, data distributions, measurement limitations, or prior results.
2. Consequential uncertainty: the unresolved issue must matter for scientific understanding or a research decision.
3. Feasible investigation: the plan must specify the population or system, data sources, measurements, comparisons, controls, analysis methods, dependencies, and resources.
4. Explicit decision criteria: the proposal must identify outcomes that would support, weaken, or leave the claim unresolved.
5. Alternative explanations: competing mechanisms, confounding factors, measurement artifacts, and model limitations must be identified.

These criteria assess the quality of a research opportunity rather than the truth of the hypothesis. Confirmation, execution, and independent validation are recorded separately. A negative result, replication, measurement check, or distinction between competing explanations can therefore constitute a valuable research direction even when it does not support the original claim.

The framework distinguishes ordinary within-run hypothesis evolution from recursive learning across research runs. A restart opens a new scientific question within an existing run; rediscovery applies accepted model and workflow improvements to future investigations. The intended cycle is:

$$\text{research experience} \rightarrow \text{credit assignment and workflow diagnosis} \rightarrow \text{model and harness updates} \rightarrow \text{new research}.$$

The paper gives a MIMIC-IV example involving whether repeated mismatch between creatinine and urine output adds information about ICU readmission or in-hospital death after ICU discharge. The proposed investigation compares a clinical baseline model, a current-mismatch model, a repeated-mismatch model, and a model augmented with hidden Markov summaries. The example illustrates that DiscoveryRSI evaluates the boundedness and executability of the question rather than asserting a clinical outcome in advance (Zhong et al., 16 Oct 2026).

2. System architecture and research objects

DiscoveryRSI consists of a multi-agent hypothesis-development loop, a persistent research-record layer, hypothesis assessment and credit assignment, Harbor task compilation and execution, harness replay, model updating, and rediscovery on fresh questions.

The persistent record stores versioned hypothesis nodes, parent and crossover links, literature and evidence records, data dependencies, scripts, results, plots, artifacts, drafts, notes, scientific assessments, agent contributions, execution records, and delivery status. The resulting research structure is technically a directed acyclic graph because crossover nodes can have multiple parents. It is distinct from an individual agent’s action trajectory.

Evidence reuse and ancestry are represented separately. A shared artifact may inform several hypotheses without making those hypotheses scientific descendants of one another. This distinction prevents common evidence, shared context, or co-occurrence in an episode from being treated as proof of scientific lineage.

The principal agent roles are:

• Lead: selects research directions, parents, operations, assignments, and candidate hypotheses.
• Generation: proposes new hypotheses and investigations.
• Evolution: develops, repairs, or refines existing hypotheses.
• Critique: performs blinded pairwise scientific comparisons.
• Progress: evaluates whether a child substantively improves on its actual predecessor.
• Compiler: translates selected investigation plans into Harbor tasks.

The principal operations are:

• Evolve: continue developing an existing direction.
• Repair: correct an identified scientific or methodological defect.
• Crossover: combine contributions from multiple parents.
• Restart: open a new question, possibly using prior context but without assigning ancestry.

This architecture treats research as a persistent, versioned process rather than a sequence of isolated generated answers. It allows the system to retain not only final hypotheses but also controls, alternative explanations, failed approaches, execution artifacts, and the specific contributions associated with revisions.

3. Hypothesis development, selection, and execution

For each fresh question, DiscoveryRSI assembles an evidence pack containing relevant literature, data descriptions, available measurements, and task constraints. Generation agents formulate candidate hypotheses in the structured form $\mathcal H=(c,\pi,d,a)$. Each node retains its claim, investigation plan, decision criteria, alternatives, supporting evidence, parent identifiers, agent provenance, and episode provenance.

Evolution agents perform targeted revisions. Typical modifications include clarifying a study population, adding a control, revising an outcome measure, identifying a confounder, specifying a threshold, or making interpretation criteria explicit. Critique agents compare candidates without producer identity, while Progress agents classify parent–child changes as substantive progress, necessary repair, repetition, or insufficient evidence.

Selection considers evidence grounding, importance, novelty, falsifiability, data resolvability, feasibility, alternative explanations, actionability, predecessor-relative progress, and execution readiness. It is not based solely on a final scalar score. A hypothesis may be selected because it constitutes a consequential and testable opportunity even if it remains unconfirmed.

Selected computational investigations are compiled into executable Harbor tasks. A Harbor task contains concrete inputs, an execution environment, required outputs, validation and verification checks, source-hypothesis links, the population or system, the comparison, discriminating outcomes, and interpretation criteria. Compilation preserves the scientific commitments of the hypothesis; it does not treat unavailable wet-lab measurements or inaccessible data as available resources.

Harbor execution checks whether a computational package satisfies its execution contract. Scientific assessment remains separate: successful execution does not establish the hypothesis, and a failed task may reflect infrastructure or packaging failure rather than scientific invalidity. Delivery maturity is tracked through stages such as:

$$\text{proposed} \rightarrow \text{selected} \rightarrow \text{compiled} \rightarrow \text{readiness-passed} \rightarrow \text{executed} \rightarrow \text{independently reviewed}.$$

This separation is central to the framework. A scientifically important hypothesis can be operationally immature, while an executable task can be scientifically weak.

4. Recursive learning mechanisms

DiscoveryRSI recursively improves through three mechanisms: hypothesis credit optimization, harness replay, and rediscovery.

Hypothesis credit optimization

The framework addresses the attribution problem that arises when different agents contribute different parts of a final hypothesis. A Lead may select the direction, a Generation agent may propose the central claim, and an Evolution agent may add a control or repair the analysis plan. A final scientific assessment alone does not identify which actions deserve reinforcement.

Blinded Critique generates preference, tie, or abstention outcomes. TrueSkill-style inference estimates latent rubric-conditioned preference strengths $(\mu_h,\sigma_h)$ for hypotheses. These parameters represent relative assessment under the rubric, not probabilities that the scientific claims are true.

DiscoveryRSI uses a conservative hypothesis score,

$s_h=\mu_h-\kappa\sigma_h,$

and combines within-stratum rank, assessment coverage, uncertainty, and progress or repair indicators into a bounded hypothesis weight $q_h$. Missing ratings are treated as unknown rather than as negative evidence.

Documented contributor provenance determines credit allocation. For hypothesis $h$, verified Lead coordination histories receive a coordination fraction $\mathcal E$0, while direct production or supporting histories receive the remaining fraction. The implemented value is $\mathcal E$1. Credit is then constrained by local scientific-family caps, global scientific-family caps, and a trajectory-level maximum weight. The implemented local and global budgets are both $\mathcal E$2, with $\mathcal E$3.

Only assistant responses and tool-call arguments supported by documented contributions receive direct supervised loss. Tool observations, failed actions, repeated context, and unsupported outputs remain contextual records but are not directly reinforced. Domain balancing assigns 90% of objective mass to life-science trajectories and 10% to battery trajectories.

Harness replay

Harness replay improves the research workflow while holding model parameters fixed. The harness includes session and context management, filesystem access, tool execution, evidence retrieval, agent coordination, review timing, handoffs, retries, stopping and completion checks, and research-budget allocation.

Replay consumes contribution records, decisions, assignments, execution traces, delivery records, tool errors, lost artifacts, repeated exploration, premature closure, incomplete handoffs, recorded observations, and continuations. It distinguishes scientific-plan failures from tool-use, coordination, evidence-access, artifact-management, and completion-check failures.

Candidate harness changes may address evidence reuse, agent handoffs, retry logic, review scheduling, memory, task completion checks, or allocation of research effort. Recorded replay is not automatically a counterfactual experiment. If an alternative continuation was not recorded, it must be executed from a compatible restored state. Candidate harnesses and the incumbent are evaluated under matched resource limits using scientific quality, exploration breadth, reliability, execution success, cost, and development performance.

Rediscovery

Rediscovery applies accepted model and harness updates to fresh questions. The next system may accept neither update, only the model update, only the harness update, or both. The complete recursive process is:

1. Research fresh questions.
2. Build hypothesis trees.
3. Compile selected plans into Harbor tasks.
4. Assess hypotheses and progress.
5. Allocate contribution credit.
6. Fine-tune a model candidate.
7. Diagnose and replay harness behavior.
8. Evaluate candidate updates on development questions.
9. Accept or reject updates.
10. Run rediscovery on new questions.

The distinction between model learning and workflow learning allows scientific reasoning failures and research-process failures to be diagnosed separately.

5. DFM-Bench and evaluation methodology

DFM-Bench evaluates scientific objects and cumulative research behavior at five connected levels.

Hypothesis quality

Each hypothesis is assessed on importance, novelty, grounding, falsifiability, data resolvability, feasibility, alternative explanations, and actionability. The weights are $\mathcal E$4, summing to 100, with scores on a 1–10 scale. A complete aggregate score requires adequate review coverage and medians for all eight dimensions.

Three automated reviewer roles assess the same evidence: a rubric judge, a domain-scientist role, and a methods-reviewer role. These are separate prompts from the same model family rather than independent human experts.

Revision value

Each revision is evaluated against its actual predecessor. Outcomes are substantive progress, necessary repair, repetition, or insufficient evidence. When both parent and child have quality scores, the numerical change is

$\mathcal E$5

Categorical judgments and numerical changes are reported separately because a higher score does not automatically establish that a revision made a scientifically meaningful contribution.

Development-path progress

A development path follows ordered hypothesis ancestry. It is evaluated for disciplinary connections, technical and knowledge exploration, cumulative scientific progress, evidential and inferential discipline, continuity and synthesis, adaptive continuation and stopping, and budget efficiency. Paths are not inferred from arbitrary agent action sequences; they follow actual hypothesis ancestry.

Research-run quality

A research run evaluates the complete graph and its decisions. Its dimensions are disciplinary coverage and integration, technical and knowledge exploration, cumulative scientific progress, evidential and inferential discipline, portfolio synthesis, adaptive scientific decisions, and budget-aware exploration.

Scientific contribution and delivery maturity

Related proposals are grouped into contribution families. Evaluators record what was added beyond the starting evidence, including a new distinction, an informative negative result, a supported finding, or a reduction in uncertainty. Contribution maturity ranges from proposed and source-grounded through diagnostically examined and study-completed to independently checked.

Delivery maturity is tracked separately as proposed, selected, compiled, readiness-passed, executed, and independently reviewed. Family membership does not imply confirmation, and delivery does not establish scientific validity.

The main experiments cover 241 research questions in four life-science domains: clinical and population health, therapeutic target prioritization, disease mechanisms, and population multi-omics. Data and evidence sources include HCC clinical cohort, MIMIC-IV, eICU, UK Biobank, GEO, and ArrayExpress. A separate battery-domain setting covers electrical diagnostics, health prediction, degradation mechanisms, and usage history and aging.

DiscoveryRSI-27B was trained with full-parameter fine-tuning using a 256K-token context window, Adam, a learning rate of $\mathcal E$6, BF16 precision, global batch size 16, one training epoch, 64 GPUs, tensor parallelism 4, context parallelism 8, data parallelism 2, and 150 optimizer steps. The training corpus contained 15,792 collected research trajectories, of which 2,397 were retained: 634 Lead, 861 Generation, and 902 Evolution trajectories. Final training contained 291.12 million tokens, with 6.18 million directly supervised tokens, or 2.12%.

Human evaluation covered 292 proposals, 27 medical doctors, medical students, and bioscience PhD students, and three senior supervisors. Reviewers received anonymized proposal packets containing claims, plans, data requirements, alternatives, and citations while model identity, execution provenance, automated scores, and revision history were hidden.

6. Results, limitations, and significance

Automated hypothesis quality increased across two recursive rounds:

$\mathcal E$7

The total increase from the base model to round two was 0.53 points. The strongest changes concerned testability and investigation design. Falsifiability increased from 4.00 to 6.20, alternative explanations from 4.44 to 5.98, data resolvability from 3.73 to 4.99, and feasibility from 4.79 to 5.87. Importance declined from 6.48 to 6.04, and novelty declined from 4.87 to 4.58. The results therefore indicate improved formulation and testability rather than increased truth of the generated claims.

Behavior shifted toward sustained refinement. The share of episodes using evolve or repair increased from 77.5% to 90.0% and then 91.7%, while restarts declined from 21.7% to 5.8% and then 5.0%. Development-path gains were largest in evidential and inferential discipline, technical and knowledge exploration, and scientific continuity and synthesis. Changes in cumulative progress, stopping, and budget efficiency were below 0.1, while disciplinary connections declined by 0.28.

For DiscoveryRSI-27B, 41% of revisions improved the mean human score, 47% were unchanged, and 12% became worse, producing a mean change of $\mathcal E$8. The corresponding comparison figures were 31%, 37%, and 32% with a mean change of $\mathcal E$9 for GPT-5.6 Sol, and 25%, 44%, and 31% with a mean change of $\mathcal B$0 for GPT-6 Luna. In absolute criterion scores, DiscoveryRSI-27B exceeded Luna on importance and plausible, evidence-bounded claims, while Luna scored higher on data fit, alternatives, and informative results.

In a post-cutoff literature-correspondence evaluation, DiscoveryRSI-27B obtained a rate of $\mathcal B$1. The comparison rates were 5.4% for DeepSeek, 2.8% for GLM, 2.5% for GPT, and 2.3% for Claude Opus 5. These figures indicate retrospective correspondence with later literature under the specified retrieval and judgment protocol; they do not show that researchers adopted or were causally influenced by the hypotheses. Retrieval coverage differed substantially among systems, creating important formatting and selection confounds.

The framework has several limitations. Most large-scale quality results rely on automated evaluation by related model prompts rather than independent expert assessment. The human study is limited in size and contains dependent parent–child comparisons. Contributor weights reflect documented provenance and fixed responsibility priors rather than causal marginal contribution. The training artifacts and private research records are not distributed, limiting reproducibility. Full-parameter fine-tuning of the 27B model with a 256K context window required 64 GPUs, while research collection required extensive multi-agent execution, tool calls, assessment, and task compilation.

The observed decline in importance and novelty alongside improved falsifiability and feasibility suggests a possible trade-off between disciplined specification and originality. Additional risks include overstatement of evidence, unsafe biomedical recommendations, demographic or clinical bias, exposure of sensitive patient data, confusion between correlation and mechanism, and premature translation. Explicit alternatives, feasibility constraints, decision criteria, and separation of execution from scientific validity are safeguards, but they do not replace expert review, ethics review, data governance, or laboratory validation.

DiscoveryRSI’s principal significance is methodological. It treats scientific discovery as a continuing process in which claims, evidence, controls, alternatives, revisions, failures, execution artifacts, and contributions remain linked. Its central contribution is not a claim that automated systems have independently produced validated scientific discoveries, but an architecture for preserving and learning from the research processes that generate, evaluate, and improve hypotheses.