New benchmark tests LLMs' ability to track elapsed time

Researchers have developed ChronoState, a new benchmark designed to test how language models handle temporal decisions based on elapsed time and symbolic task state. Using a Qwen2.5-3B-Instruct model, ChronoState demonstrated that hidden elapsed time conditioning, when injected through a non-token channel, significantly improves performance on temporal-state action selection. However, the system's ability to generalize to unseen time-related concepts and its autonomous time-tracking capabilities remain limited, with prompt-injected timestamps proving more effective in some scenarios. AI

IMPACT This research explores methods for integrating temporal awareness into LLMs, potentially improving their performance in time-sensitive applications.

RANK_REASON The cluster contains a research paper introducing a new benchmark and methodology for evaluating language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ChronoState
• Qwen2.5-3B-Instruct

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →