LED: Latent Error Dynamics — prediction records and evaluation scripts
Description
Supporting data for the paper "LED: Modeling Latent Error Dynamics for First-Error Localization in LLM Reasoning" by Z. Pang, Y. Li, R. Zhang, J. Wang and Z. Miao. This deposit contains the per-chain prediction records from which every table and interval in the paper is derived, together with the scripts that reproduce them. No model weights or hidden-state caches are needed to recompute the reported numbers. Records are provided for three backbones (DeepSeek-R1-Distill-Llama-8B, Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3) under three protocols (main chain-level split, problem-disjoint grouped split, extended-protocol controls). Each record carries the chain identifier, difficulty tier, chain length, annotated first-error step, outcome bit, and the per-seed predictions and scores of every method compared in the paper. The ProcessBench benchmark is publicly available from its own distribution and is not redistributed here. Trained probe weights and representation caches are released with the camera-ready version. See README.md inside the archive for the full field schema.