Agent Trajectory Evaluation — References

Sources on the loop structure these traces record, on self-correction as a measurable behaviour, and on environments where a run can be reproduced.

The structure being scored

• ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., 2022), arXiv:2210.03629 — the thought, action, observation cycle that a trajectory record is a serialisation of, and the reason the observation has to be captured for any step verdict to be meaningful.

Error and recovery

• Reflexion: Language Agents with Verbal Reinforcement Learning (Shinn et al., 2023), arXiv:2303.11366 — carrying a written critique of a failed attempt into the next one, which is the clearest published treatment of recovery as a first-class behaviour rather than as evidence of a defect.

Reproducible environments

• SWE-bench: Can Language Models Resolve Real-World GitHub Issues? (Jimenez et al., 2023), arXiv:2310.06770 — cited here for the environment design: a task whose world can be reset and whose outcome is checkable without a judgement, which is the precondition for replaying a trajectory at all.
• The agent benchmark environments built around interactive websites and simulated operating systems — where a task is specified in natural language and success is read from the resulting environment state — are the reference designs for reproducible trajectory measurement. Described rather than listed by title, since several of these projects share names across distinct papers and versions.

Credit assignment

• The reinforcement-learning literature on temporal credit assignment — attributing a delayed outcome to the actions that caused it — is the formal statement of the central problem here, though its standard solutions assume a reward signal and a number of episodes that language agent evaluation does not have. Described as a body of work rather than cited to one title.

No verified reference is recorded for checkpoint-segmented trajectory scoring, for the error-and-recovery cross-tabulation used here, or for mid-run replay from a divergence point.