Systems that automate scientific discovery must repeatedly decide which experiment to run, which hypothesis to test, which tool to build, and when to stop. Many systems make these decisions by maximizing a myopic score such as expected information gain per unit cost or a learned plausibility score. We identify a structural limitation of this approach. Some actions are constructive: they acquire an epistemic capability (an instrument, assay, pipeline, simulator, or abstraction) whose value lies not in the information returned immediately but in the future actions it makes available. When the least-cost route to a confident answer requires a chain of such constructions, a planner that scores actions only by information obtainable within a bounded horizon cannot value the first construction: it yields no information within the horizon and is dominated by any measurement with positive information, however small. We formulate goal-directed discovery as a stochastic shortest-path problem in belief space in which constructive experiments change the downstream action graph, and prove that for every lookahead depth d there is an instance on which every myopic information-maximizing planner has an unbounded approximation ratio, and a related instance on which it never reaches the goal. The mechanism is a capability-indistinguishability lemma: within the horizon, acquiring a capability can be observationally indistinguishable from paying for a null action. This establishes capability gating as a reachability axis of difficulty distinct from curvature (submodularity) and information order (adaptivity gaps). We introduce CG-Plan, an incremental replanner with a capability-aware cost-to-go heuristic h = h_cap + h_exp. In a controlled testbed, the performance gap appears only under gating, persists for every fixed horizon, and arises when near-miss hypotheses come from a data-consistent proposer.
Current discussions of AI in scientific discovery are often dominated by two visible capabilities: search over existing knowledge and execution through optimization, simulation, and automation. Both are important, but neither fully captures the central act of discovery: the formation and evolution of models. This paper proposes a three-layer view of AI in discovery. Layer 1 is search and retrieval by large language models. Layer 2, as the main innovation of this paper, is model formation through qualitative reasoning: the capacity to recognize when a current framework is structurally inadequate and to understand the problem within a broader representational space, not through trial and error, but through structural insight into what is missing and where it can be found. Layer 3 is execution, optimization, and refinement. The main claim is that Layer 2 is both the most important and the least developed. Search without model formation remains confined to inherited frameworks, while execution without conceptual revision only amplifies an existing formulation. We illustrate Layer 2 reasoning through three case studies: S. S. Chern's intrinsic proof of the Gauss-Bonnet theorem, the resolution of the Nesterov Accelerated Gradient convergence problem via Lyapunov functions, and the autonomous disproof of the Erdos unit distance conjecture by OpenAI in 2026. Each case exhibits the same structural signature: a framework that had become inadequate, a missing conceptual object, and a resolution found in an unexpected neighboring field.
Large language models (LLMs) are increasingly deployed as autonomous agents in scientific tasks. Yet whether these systems can effectively engage in forms of inductive reasoning relevant to scientific discovery remains an open question. In this work, we introduce FALSIFYBENCH, an evaluation framework for hypothesis-driven reasoning inspired by the classic Wason 2-4-6 task, in which agents must discover hidden semantic properties by iteratively proposing examples and receiving feedback. This task captures key elements of scientific reasoning: hypothesis generation, evidence gathering, and belief revision in response to both confirming and disconfirming evidence. Our evaluation of 12 LLMs across model families and scales shows that reasoning models are generally stronger scientific reasoners than instruction-tuned models, although no model comes close to optimal performance. The primary driver of success is the capacity for negative testing: models that actively seek to falsify their hypotheses consistently outperform those that primarily seek confirmation. Moreover, a fine-grained turn-level analysis, neglected in previous work, reveals that failure is tied to identifiable patterns in how models navigate the hypothesis space.