WildSci: Advancing Scientific Reasoning from In-the-Wild Literature

TL;DR

WildSci introduces a fully automated pipeline to synthesize 56K domain-specific science questions from peer-reviewed literature across 9 disciplines and 26 subdomains, framing them as multiple-choice items to enable scalable reinforcement learning with verifiable rewards. The dataset is complemented by a model-voting quality control process and a GRPO-based training regime, demonstrating improvements on in-domain and out-of-domain science benchmarks (WildSci-Val, GPQA-Aug, SuperGPQA, MMLU-Pro). The work analyzes data-driven training dynamics, domain coverage effects, and the impact of question refinement, revealing post-saturation generalization and domain-dependent performance dynamics. By releasing code and data, WildSci aims to support sustainable, scalable research in scientific reasoning for language models and facilitate cross-domain transfer including math-science integration through mixed-training strategies.

Abstract

Recent progress in large language model (LLM) reasoning has focused on domains like mathematics and coding, where abundant high-quality data and objective evaluation metrics are readily available. In contrast, progress in LLM reasoning models remains limited in scientific domains such as medicine and materials science due to limited dataset coverage and the inherent complexity of open-ended scientific questions. To address these challenges, we introduce WildSci, a new dataset of domain-specific science questions automatically synthesized from peer-reviewed literature, covering 9 scientific disciplines and 26 subdomains. By framing complex scientific reasoning tasks in a multiple-choice format, we enable scalable training with well-defined reward signals. We further apply reinforcement learning to finetune models on these data and analyze the resulting training dynamics, including domain-specific performance changes, response behaviors, and generalization trends. Experiments on a suite of scientific benchmarks demonstrate the effectiveness of our dataset and approach. We release WildSci to enable scalable and sustainable research in scientific reasoning, available at https://huggingface.co/datasets/JustinTX/WildSci.

Figures (10)

• Figure 1: Overview of the data creation pipeline. Filtering is based on heuristic rules, while refinement expands the option space and rephrases questions to increase diversity.
• Figure 2: Comparison of the number of papers and gen- erated questions across different disciplines.
• Figure 3: Distribution of questions after filtering and refinement in WildSci.
• Figure 4: Performance trends on validation and test sets during training of the 3B model on WildSci All Aligned. The model exhibits continued generalization on test sets even after overfitting on the validation set.
• Figure 5: Domain-specific accuracy trends on MMLU-Pro during training. We report mean accuracy across three runs for each domain. Shaded regions indicate standard deviation. (a) shows steady improvements in domains with higher WildSci coverage (chemistry, physics, engineering), while (b) illustrates more variable performance in domains with lower coverage (law, history, philosophy).