Maksym Andriushchenko Posts Study on LLM Agent Time Awareness
Maksym Andriushchenko shares ongoing work testing agents on long-horizon benchmarks.
Maksym Andriushchenko posted a new LessWrong blog post asking whether LLM agents are time-aware. The post examines if agents can predict wall-clock time for tasks and estimate time already spent. Andriushchenko described the project as work in progress and said a full paper will follow with expanded related work. Replies from researchers such as Niloofar Mireshghallah and Andreas Kirsch discussed the results and suggested adding timestamps by default. Other posts noted similar failure modes observed in InferenceBench.
💥New blog post: Are LLM agents time-aware? Can they predict wall-clock time of tasks? Can they estimate time they spent? We study this on a range of tasks, including long-horizon ones (ProgramBench, PaperBench, DeepSWE, etc). Joint work with my MATS mentee @MOfengenden!
Reactions from ranked influencers
23 postsTime awareness is an essential skill required from a digital worker and a key part of broader situational awareness. Our LessWrong post:
This is work in progress and we would appreciate any feedback! We will follow-up with a full paper, which would include, among other things, a much more detailed overview of related work.
@maksym_andr @MOfengenden Banger after banger Maksym!!🥳
agents are hilariously bad at estimating time. one of many reasons for this is, obviously, that they do not experience time like we do this is why i periodically check in on models and ask them about what time means for them. for example, this is the new qwen3.8 27b:
Time awareness is an essential skill required from a digital worker and a key part of broader situational awareness. Our LessWrong post:
agents are hilariously bad at estimating time. one of many reasons for this is, obviously, that they do not experience time like we do this is why i periodically check in on models and ask them about what time means for them. for example, this is the new qwen3.8 27b:
Time awareness is an essential skill required from a digital worker and a key part of broader situational awareness. Our LessWrong post:
one of the failure modes of agents found in InferenceBench is a lack of time awareness, where agents fail to adjust their strategy or prioritize getting a working solution as the deadline approaches despite being given (and using) a timer another banger post from @maksym_andr💥
💥New blog post: Are LLM agents time-aware? Can they predict wall-clock time of tasks? Can they estimate time they spent? We study this on a range of tasks, including long-horizon ones (ProgramBench, PaperBench, DeepSWE, etc). Joint work with my MATS mentee @MOfengenden!
