Scientific feedback reveals productive regimes beyond training conventions

Many of the most important problems in science and engineering are inverse problems: given a desired outcome, find a design that achieves it. Evaluating whether a candidate meets the spec is often routine; a binding energy can be computed, a reactor yield simulated, a pharmacokinetic profile predicted. But searching a combinatorial design space for inputs that satisfy those targets is fundamentally harder. We introduce SciDesignBench, a benchmark of 520 simulator-grounded tasks across 14 scientific domains and five settings spanning single-shot design, short-horizon feedback, long-horizon refinement, and seed-design optimization. On the 10-domain shared-core subset, the best zero-shot model reaches only 29.0% success despite substantially higher parse rates. Simulator feedback helps, but the leaderboard changes with horizon: Sonnet 4.5 is strongest in one-turn de novo design, whereas Opus 4.6 is strongest after 20 turns of simulator-grounded refinement. Providing a starting seed design reshuffles the leaderboard again, demonstrating that constrained modification requires a fundamentally different capability from unconstrained de novo generation. We then introduce RLSF, a simulator-feedback training recipe. An RLSF-tuned 8B model raises single-turn success rates by 8-17 percentage points across three domains. Together, these results position simulator-grounded inverse design as both a benchmark for scientific reasoning and a practical substrate for amortizing expensive test-time compute into model weights.

Scientific ideation is increasingly mediated by large language models, but current ideation systems are usually trained and evaluated on immediately judgeable proxies such as novelty, clarity, and feasibility. This leaves open whether delayed signals of scientific uptake can be used as feedback for steering models toward research directions with higher expected impact. We study this question using citation-normalized impact as a noisy but scalable proxy for scholarly uptake. We construct a large-scale dataset from over 100K computer science papers by extracting goal-conditioned idea descriptions and assigning each paper an ordinal, year-normalized citation label. We then train a goal-conditioned reward model to predict citation-impact labels from research goal and idea pairs, and use this reward to align an idea generator through supervised fine-tuning followed by reinforcement learning. To reduce circularity, we evaluate generated ideas with a held-out, reference-grounded protocol that compares model outputs against historical ideas under the same research goal and weights judgments by the reference idea's citation-impact label. Experiments show that our RL-tuned model consistently produces ideas with higher estimated impact than both the base model and supervised fine-tuning baselines. Our findings position scientific impact as a practical, outcome-grounded feedback signal for aligning LLMs in open-ended scientific discovery.

Recent work has questioned whether large language models (LLMs) can perform genuine in-context learning (ICL) for scientific experimental design, with prior studies suggesting that LLM-based agents exhibit no sensitivity to experimental feedback. We shed new light on this question by carrying out 800 independently replicated experiments on iterative perturbation discovery in Cell Painting high-content screening. We compare an LLM agent that iteratively updates its hypotheses using experimental feedback to a zero-shot baseline that relies solely on pretraining knowledge retrieval. Access to feedback yields a $+53.4\%$ increase in discoveries per feature on average ($p = 0.003$). To test whether this improvement arises from genuine feedback-driven learning rather than prompt-induced recall of pretraining knowledge, we introduce a random feedback control in which hit/miss labels are permuted. Under this control, the performance gain disappears, indicating that the observed improvement depends on the structure of the feedback signal ($+13.0$ hits, $p = 0.003$). We further examine how model capability affects feedback utilization. Upgrading from Claude Sonnet 4.5 to 4.6 reduces gene hallucination rates from ${\sim}33\%$--$45\%$ to ${\sim}3$--$9\%$, converting a non-significant ICL effect ($+0.8$, $p = 0.32$) into a large and highly significant improvement ($+11.0$, $p=0.003$) for the best ICL strategy. These results suggest that effective in-context learning from experimental feedback emerges only once models reach a sufficient capability threshold.