AI Night-Scientist uses reinforcement learning to teach models when to depart from predictable reasoning, widening research-direction coverage by 27.8% and contribution types by 14.9%

The framework explicitly models creativity at three levels of the reasoning trajectory: action-level creativity assigns each action an ordered set of creativity levels described in natural language from conventional to exploratory, process-level creativity decides when to shift between lower- and higher-creativity actions, and outcome-level creativity captures the novelty and usefulness of the resulting idea. Prior LLM research agents generally treat actions such as searching, debating, and writing as fixed behaviors and assess creativity mainly in the final idea; this work places creativity both in how an action is performed and in the timing of reasoning, and gives users direct semantic control over the degree of deviation. The paper provides formal definitions of the three levels (Definitions 3.1-3.3) and a concrete table of five creativity levels each for search, debate, and spark, with write fixed at level 1 to preserve credit assignment.

After GRPO training, AI Night-Scientist-8B improves predicted citation impact by 29.43 percentage points and originality by 53.79 points over Qwen3-8B; the 14B version improves by 32.03 and 66.15 points over Qwen3-14B. Scaling the zero-shot model from 8B to 14B yields almost no improvement, whereas scaling AI Night-Scientist adds another 3.29 points in citation and 12.36 points in originality, indicating that model capacity becomes substantially more useful once paired with a learned creative reasoning policy. Pairwise comparisons against matched reconstructed reference proposals on 491 NSF test awards with randomized order, reported with 95% confidence intervals; direct head-to-head comparisons against GPT-4.1 and Qwen3-8B give win rates of 74.95%/86.96% and 82.48%/97.56%.

The gains cannot be reproduced by raising decoding temperature: optimizing the same proposal-level reward, Night-8B exceeds the temperature-controlled variant by 36.50 points in originality and 18.37 points in citation; ReAct closes much of the gap, but semantic creativity levels still add 9.72 points in originality and 4.95 points in citation. This separates 'explore more' from 'specify how an action should deviate', showing that the learnable signal is semantic creativity guidance rather than sampling stochasticity; restricting the action space to search and write reduces originality by 22.99 points, indicating that spark and debate provide additional ways to redirect reasoning. The temperature baseline maps the five creativity levels only to decoding temperature and shares the same proposal-level reward; ablations are compared under the same training and evaluation pipeline, with a reward ablation table.

Diversity also improves: relative to zero-shot Qwen3-8B, Night-8B raises normalized research-paradigm coverage from 0.738 to 0.943 (roughly 1.4 additional effective categories out of seven) and contribution-type coverage from 0.370 to 0.425; removing early serendipitous action swaps reduces paradigm coverage by 0.099. This distinguishes 'more original' from 'repeatedly relying on the same creative strategy' and shows that early exposure to less familiar behaviors helps the policy discover a broader set of useful trajectories; process rewards instead shift the model toward higher per-proposal originality at the cost of citation impact and breadth. Diversity is measured with GPT-5.4-mini classification into seven contribution types and paradigm categories, reported as a normalized effective number of categories with 100,000 bootstrap resamples; across 36 evaluated domains Night-8B improves over zero-shot Qwen3-8B on both citation and originality.

The framework targets early-stage scientific ideation, when directions are still speculative and not yet fully testable; the authors explicitly position it as a creative collaborator for human-led research rather than an autonomous researcher, and generated proposals should be treated as candidate ideas requiring expert evaluation, refinement, and validation. It applies to generating proposals of a scope similar to long-term research grants from an award title as input, with training and evaluation data drawn from NSF Computer Science, Engineering, and Mathematics awards from 2018 onward and retrieval grounded in open-access arXiv literature. For readers, it offers a reusable idea: treat 'when to explore and how to explore' as a trainable control variable rather than equating randomness with creativity; for interdisciplinary exploration, the authors suggest it may help surface directions beyond a researcher's local community.

Predicted citation impact is proxied by SciJudge-30B and originality by GPT-5.1 making pairwise judgments against a retrieved nearest-neighbor paper; agreement with human judgment is in the low-to-mid 70s, so these numbers are best read as relative comparisons under evaluation proxies rather than measures of real long-term influence. The reference proposals are not original NSF proposals but reconstructions from award abstracts, project outcomes reports, and PI-related papers, described by the authors as plausible pre-award research plans used only for pairwise evaluation. Process rewards raise originality but lower citation impact and breadth, indicating that different reward combinations point to different goals that readers should choose according to their own aims. The authors also note that training can become unstable under high-entropy objectives and that infrastructure support for structured multi-step agents is limited. In addition, although this evidence bundle is a full-text parse, some tables and appendix values appear as placeholders or omissions in the text, so a few specific numbers cannot be individually verified here.