Nice work! I like the ablations, and the harness-level comparison is an important investigation. However, the post presents its main findings without considering very similar results from other papers. A lot in your analysis is genuinely new, but the headline has been shown before.
Can LLMs Perceive Time? An Empirical Investigation (ICLR 2026 workshop) tests wall-clock prediction and finds LLMs overestimate their own runtime by ~4–7x, similar to the 3x/7x you report. They use a ReAct scaffold though, instead of Claude Code or Codex.
Your LLM Agents are Temporally Blind (Findings of ACL 2026) tests LLMs’ alignment with human’s judgments in time-sensitivity scenarios. Like you, they test adding timestamps and also find that alignment rises, though not dramatically.
Discrete Minds in a Continuous World (Findings of EMNLP 2025) tests where a sense of time could come from and finds that models can map conversation token counts to elapsed wall-clock time (their “token-time hypothesis”, cf. your transcript-length correlation of r = 0.91), and adapt behaviour under expressed urgency (loosely similar to your human/AI ablations).
Important work, thank you for putting this together! With respect to human baselines, there is some work on it and it is similar to what you find with LLMs. Reviewer scores predict citations poorly: near-zero correlation for accepted NeurIPS papers over 7 years (Cortes & Lawrence 2021) and roughly zero among ICLR spotlights/orals (Tran et al. 2020).
Notably, scores do predict citations for rejected papers. Review separates weak papers from average ones, not good from great. Since TastyBench pre-filters to papers with ≥20 citations (probably not bad papers), it sits in the range where human experts have poor predictive power. The process is also noisy enough that re-running reviews flips ~half the acceptance decisions (Beygelzimer et al. 2023).
Meanwhile authors ranking their own papers do beat review scores at predicting citations (Su et al. 2025). So the signal exists, but apparently requires insider knowledge that neither reviewers nor a model reading abstracts has. So a plausible human-expert baseline here is near zero, and low LLM scores don’t establish subhuman taste.