TBG-CoT-Bench: A Benchmark for Evaluating Temporal Belief Tracking over Chain-of-Thought Evidence Sequences

We introduce TBG-CoT-Bench, a benchmark for evaluating how reliably language models track a temporal belief — whether Event A occurred before Event B — as evidence is presented sequentially in Chain-of-Thought form. Unlike existing temporal-reasoning benchmarks, which are largely static question-answering tasks, TBG-CoT-Bench tracks belief step by step across direction, strength, source, and source_weight, and evaluates five methodologies (a rule-based baseline, scenario-level CoT, step-wise evidence extraction, order-classification, and cumulative belief tracking) on ten synthetic scenarios using a locally deployable EXAONE model. We find that step-wise evidence extraction, which constrains the model to a small, fixed output schema at each evidence step, is the most reliable of the five architectures examined. This repository accompanies the dataset released at https://huggingface.co/datasets/CHML-real/tbg-cot-bench.