Daily AI Paper

Scientific discovery is not just about generating ideas. It is about turning a plausible idea into a testable investigation, learning from what happens along the way, and using that experience to make the next idea better. That is the problem this paper tackles. The authors argue that for research agents, the hard part is not only proposing hypotheses, but preserving the full development history behind those hypotheses so the system can learn which research moves were actually useful. Why does that matter? In real science, a good hypothesis is rarely produced in one shot. Researchers gather evidence, revise the question, add controls, think through alternative explanations, and decide whether a result would really change what we believe. Those intermediate steps contain valuable learning signal. But if an AI system only keeps the final draft of a hypothesis, it loses the trail of reasoning and the connection between scientific quality and the actions that produced it. That makes it difficult to improve the agent in a principled way. This paper sits in the growing area of AI for scientific research, alongside systems that generate hypotheses, search over research trees, or turn scientific tasks into executable workflows. The authors build on that landscape, but their emphasis is different: they want the agent to learn recursively from the research process itself. In other words, not just to do research, but to get better at doing research because of the research it has already done. To make that possible, they first define what a valuable and testable research hypothesis should look like. Their definition is operational, meaning it is meant to be used by a system, not just discussed philosophically. A hypothesis here has four parts. First is an evidence-grounded claim: what system or population is being studied, what uncertainty is being addressed, and what is being claimed. Second is an investigation plan: what data, measurements, comparisons, controls, and analyses would be used. Third is explicit decision criteria: before seeing results, what outcomes would support the claim, weaken it, or leave it unresolved. Fourth is a set of alternative explanations: competing accounts, confounders, or measurement limits that could explain the result differently. This is an important point. The paper is not saying a hypothesis is valuable only if it is true. A hypothesis can be valuable if resolving it would change scientific understanding or influence a justified decision. It can also be valuable if it leads to a useful negative result, a replication, or a measurement check. And it is testable only if the proposed investigation is actually feasible under the available data, experimental, and resource constraints. So the framework keeps scientific ambition tied to practical execution. The system they introduce is called DiscoveryRSI. At a high level, it works in three stages. First, it develops hypothesis trees through multi-agent evolutionary search. That means hypotheses are revised, combined, repaired, or restarted across episodes, with each version linked to its predecessor and to the evidence and decisions behind it. The result is not just a sequence of drafts, but a structured research history. Second, when a computational investigation is selected, DiscoveryRSI compiles it into an executable Harbor task. Harbor tasks have concrete inputs, environments, required outputs, and verification checks. This matters because it turns a research plan into something that can actually be run and evaluated. The paper is careful to note that execution validity and scientific validity are not the same thing. A task can satisfy its computational contract, while the scientific question still remains open. Likewise, some dependencies, such as unavailable data or physical experiments, remain explicit rather than being silently ignored. Third, the system learns recursively from these histories in three coupled ways. The first is hypothesis credit optimization. Here, scientific assessments are traced back to the documented contributions that produced them, and bounded credit is assigned to those contributing histories. Those credit assignments become supervision for training the model, which in this paper is used to train DiscoveryRSI-27B. The key idea is attribution: if a useful revision, control, or explanation improved the hypothesis, the system should learn which actions deserve reinforcement. The second mechanism is harness replay. This looks across the accumulated research histories to identify productive strategies and recurring failures. In practical terms, that can inform changes in how agents coordinate, how they access evidence, and how they allocate research effort. So the learning is not only about the model’s internal parameters; it also updates the research workflow around the model. The third mechanism is rediscovery. After accepted improvements are made to the model or the harness, the system applies them to fresh investigations. That creates new hypotheses and new experience, which can then feed the next round of learning. So the process is recursive: research experience improves the agent, which then produces better research experience. To evaluate all of this, the authors introduce DFM-Bench, a benchmark designed to assess both individual hypothesis quality and cumulative research progress. It looks at hypotheses, revisions, development paths, and research runs. Importantly, it evaluates not only whether a hypothesis is good, but also what the revisions contributed and how progress accumulates across a path. Execution readiness of compiled Harbor tasks is measured separately from the scientific quality of the hypothesis itself. The experimental setup spans four life-science domains. The paper does not frame the benchmark as a single narrow task; instead, it uses it to examine whether the framework generalizes across different kinds of research problems. They compare the base model to versions improved through self-learning, and they also compare DiscoveryRSI-27B with other research agents using expert-supervised review and controlled model comparisons. The main quantitative result is that automated evaluation shows mean hypothesis quality rising from 4.86 for the base model to 5.39 after two rounds of self-improvement. The paper highlights that the biggest gains are in falsifiability and data resolvability. That is meaningful because those are exactly the properties that make a hypothesis more scientifically actionable: can it actually be tested, and can the available data distinguish the claim from alternatives? The expert-supervised assessments add another layer. They show that DiscoveryRSI-27B produces greater human-rated improvement along research paths than the comparison systems, with an average gain per revision of +0.07, compared with +0.03 for GPT-5.6 Sol and -0.01 for GPT-6 Luna. In other words, the revisions generated by DiscoveryRSI appear to improve hypotheses more consistently than the baselines in the expert review. The paper also reports a post-cutoff correspondence measure: DiscoveryRSI-27B achieved a 6.3% rate of correspondence with research published after the model cutoff, compared with 2.5% for GPT-6 and 2.3% for Claude Opus 5. This should be interpreted cautiously. It is not a claim that the model is discovering new science on its own in a strong causal sense. Rather, it suggests that its hypotheses more often align with later published research than the comparison systems do, which the authors present as supportive evidence that the framework is generating more relevant and scientifically grounded ideas. Across the different evaluations, the pattern is consistent: the recursive learning setup seems to improve not just the surface quality of hypotheses, but the properties that matter for scientific usefulness. The model becomes better at making claims that are evidence-bounded, more falsifiable, and more resolvable with data. It also appears to improve the quality of the research paths leading to those hypotheses, which is crucial because science is often about the path, not just the endpoint. There are, however, some limitations and things to watch next. First, the framework is centered on computationally tractable investigations, even though it explicitly acknowledges that some scientific questions depend on unavailable data or physical experiments. That means the current approach is strongest where hypothesis development can be linked to executable tasks and structured evidence, and less directly applicable where experimentation is slow, expensive, or highly manual. Second, the evaluation of scientific quality still depends on benchmark design and expert judgment. Those are valuable, but they are not the same as proving that a hypothesis is truly important in the broader scientific sense. The paper’s operational definition is useful because it makes the process measurable, but it also means that some aspects of scientific value remain context-dependent. Third, the recursive learning loop depends on correct attribution. If the system misidentifies which contribution mattered, or if the research record is incomplete, the credit assignment could reinforce the wrong behavior. That is a general challenge for any learning-from-process system. Still, the central contribution is clear. DiscoveryRSI turns hypothesis development into a structured, learnable process. It preserves the reasoning history, compiles selected investigations into executable tasks, and uses the resulting records to improve both the model and the research workflow. The paper’s broader message is that scientific agents may improve most when they are trained not just on final answers, but on the full, revisable path that produces a worthwhile hypothesis.