ChronoState: Hidden Time Conditioning for Frozen-Backbone Language Models

A new study has introduced ChronoState, a benchmark aimed at evaluating how well frozen-backbone language models can use elapsed wall-clock time for making decisions based on timing. This research, found on arXiv with ID 2608.09124, looks at scenarios where decisions depend on symbolic task states and time passing—think things like cache expirations or deadlines. The authors propose adding elapsed time as a hidden scalar, which isn't visible in the token sequence but can aid model calculations. Using the Qwen2.5-3B-Instruct as a frozen bf16 backbone, they implemented a 31-dimensional sine-plus-log time encoding and achieved impressive accuracy of 0.9305 ± 0.0134 on the benchmark, suggesting that these models could improve temporal reasoning without needing fine-tuning.