Science

AAAR-1.0 (EqInfer)

AAAR-1.0 Equation Inference: given the LaTeX context before and after a removed equation from a peer-reviewed NLP/ML paper, pick the human-written equation among 4 candidates (3 GPT-4-synthesized negatives). Per-item results recovered from the deleted git history of the official repo cover the single released subject, Gemini Pro (multiple-choice setting, 1000-word contexts); the paper's other models have no released per-item outputs.