Search papers, labs, and topics across Lattice.
InfiniteScienceGym is introduced as a procedurally-generated benchmark for evaluating scientific reasoning in LLMs, addressing limitations of existing benchmarks like publication bias and label noise. The benchmark generates realistic scientific repositories with verifiable question-answering tasks, enabling controlled evaluation of evidence-grounded reasoning, abstention, and tool use. Experiments with various models reveal that even strong models struggle with recognizing unanswerable questions and achieve only moderate accuracy (<=45%), highlighting the need for improved reasoning and abstention capabilities.
LLMs still fail basic science: even the best models struggle to answer questions grounded in procedurally-generated, noise-free scientific data, achieving only 45% accuracy.
Large language models are emerging as scientific assistants, but evaluating their ability to reason from empirical data remains challenging. Benchmarks derived from published studies and human annotations inherit publication bias, known-knowledge bias, label noise, and substantial storage requirements. We present InfiniteScienceGym, a procedurally generated benchmark of scientific repositories paired with a verifiable question-answering task. From a seed, the simulator deterministically generates a self-contained repository with realistic directory structure, files, and tabular data, and a privileged QA generator produces both answerable and unanswerable questions with exact ground truth. This makes it possible to evaluate evidence-grounded reasoning, abstention, and tool-mediated analysis in a controlled setting without distributing a large static corpus. InfiniteScienceGym complements real scientific benchmarks by targeting blind spots and failure modes that are hard to evaluate using published datasets alone. Evaluating both proprietary and open-weight models, we find that none achieve more than 45% accuracy overall, that recognizing unanswerable questions remains a major weakness, and that stronger models tend to use tools more effectively rather than simply consuming more tokens.
