Teaching Language Models to Forecast Research Success Through Comparative Idea Evaluation

Srujan P Mule, Aniketh Garikaparthi, Manasi Patwardhan

arXiv:2605.21491 · 2026-05-23 공개 · arXiv · PDF

reinforcement-learning rlvr verifiable-rewards scientific-discovery small-language-models autonomous-research benchmark-performance comparative-forecasting

Abstract

As language models accelerate scientific research by automating hypothesis generation and implementation, a new bottleneck emerges: evaluating and filtering hundreds of AI-generated ideas without exhaustive experimentation. We ask whether LMs can learn to forecast the empirical success of research ideas before any experiments are run. We study comparative empirical forecasting: given a benchmark-specific research goal and two candidate ideas, predict which will achieve better benchmark performance. We construct a dataset of 11,488 idea pairs grounded in objective outcomes from PapersWithCode. While off-the-shelf 8B-parameter models struggle (30% acc.), SFT dramatically boosts performance to 77.1%, outperforming GPT-5 (61.1%). By framing evaluation as a reasoning task via Reinforcement Learning with Verifiable Rewards (RLVR), we train models to discover latent reasoning paths, achieving 71.35% acc. with interpretable justifications. Through additional ablations and out-of-distribution tests, we show robustness to surface-level heuristics and transfer to both a cross-domain time-split test set and an independently constructed test set. Our results demonstrate that compute-efficient small language models can serve as effective, objective verifiers, offering a scalable path for autonomous scientific discovery.

한국어 요약

한 줄 요약

**[Research Forecasting / Comparative Idea Eval]** SFT로 8B LM이 PapersWithCode 기반 11,488 idea pair에서 77.1% 정확도로 연구 아이디어 비교 forecasting, GPT-5(61.1%) 능가하며 RLVR로 71.35% 정확도·interpretable justification 동시 제공.

핵심 기여도

핵심 아이디어

연구 아이디어의 empirical 성공 예측은 정답 정의가 가능한 PapersWithCode 결과를 활용하면 small LM도 SFT·RLVR로 GPT-5보다 더 정확하고 interpretable한 verifier가 될 수 있으며, 이것이 autonomous 과학 발견의 scalable 경로다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: Small LM이 GPT-5 능가하는 verifier 가능성 입증, compute-efficient autonomous science 경로 제시, RLVR로 interpretable·정확도 trade-off 균형. **한계**: PapersWithCode에 grounded되어 다른 평가 metric·도메인 일반화 별도 검증, comparative pair에 한정(단일 아이디어 절대 평가 미커버), 새 분야·새 metric에 대한 추론 한계.

실용적 활용