A-B-D/wild-trajectory-annotations

Wild-Trajectory Annotations

Functional behavior annotations of the 27,803 trajectories in A-B-D/wild-trajectory: 947,383 agent steps, each coded with up to three verbatim quotes labelled with the Act·ONOMY action taxonomy (10 action groups, 46 sub-actions). 1,862,737 labelled spans in total.

This is stage 1 of the paper On the Behavioral Traits of LLM Agents. The code that produced these files and turns them into per-trajectory features is at github.com/JoeZhao527/A-B-D (behavior-annotation/), together with the codebook given to the annotator (behavior-annotation/codebook/).

Files

file	rows	content
turns.parquet	947,383	one row per annotatable agent step: the text shown to the annotator
annotations.jsonl	947,383	one record per step: the annotator's labelled spans

The two files join on (trajectory_id, step_index).

turns.parquet

An agent step is one source=agent ATIF step (one model call). It is annotatable when it is not a demo (few-shot) step and carries reasoning, a message or a tool call.

column	content
trajectory_id, model, task_type, source, harness	the trajectory and its grouping variables
step_index, n_steps	position of the step among the trajectory's annotatable steps, and their number
step_text	the step as shown to the annotator: [think] … (reasoning), [think:summary] … (API reasoning summary), [msg] … (message), one [action] <tool>: <arguments> per tool call
obs_before	what the agent received since its previous annotatable step (tool results, user messages, environment events), truncated to 1,200 characters
obs_before_full_len	length of that context before truncation
tools	names of the tools called in the step
has_thought, has_action	whether the step has reasoning or a message / a tool call
step_id	the ATIF step_id of the step in the trajectory

annotations.jsonl

{"trajectory_id": "...", "step_index": 0, "model_under_study": "alibaba/qwen3.7-plus",
 "task_type": "tool-api",
 "spans": [{"quote": "Let me start by understanding the workspace. I need to:",
            "grp": "Reasoning", "sub": "Analysing", "leaf": null},
           {"quote": "search: {\"query\": \"\"}",
            "grp": "Retrieval", "sub": "Retrieve from local corpus", "leaf": null}]}

field	content
quote	a verbatim phrase from step_text
grp	Act·ONOMY action group (10)
sub	Act·ONOMY sub-action (46)
leaf	always null: annotation stopped at the sub-action level

Annotation setup

• Annotator: openai/gpt-oss-120b served with vLLM 0.11.2, temperature 0, reasoning effort low.
• Input per step: step_text, the obs_before context, and the text of the previous 8 annotatable steps of the trajectory. Only step_text may be quoted.
• One call selects up to three quotes from the step and gives each an action group; quotes that are not found in step_text are discarded. Then one call per quote picks its sub-action from the sub-actions of that group.

License

The quoted text comes from the trajectories in A-B-D/wild-trajectory, each under its source's license.

Citation

@article{zhao2026behavioral,
  title   = {On the Behavioral Traits of LLM Agents},
  author  = {Zhao, Haokai and Gao, Jie and Xiao, Yunze and Wang, Xintao and Xuan, Weihao and
             Joshi, Aditya and Dredze, Mark and Huang, Jen-tse},
  journal = {arXiv preprint arXiv:2609.32776},
  year    = {2026}
}