liuqi6777/bcp_search_agent_trajectory

BrowseComp-Plus search-agent trajectories

Full ReAct trajectories for six search agents on BrowseComp-Plus, released with Diagnosing Search Behavior and Failure Modes in Long-Horizon Search Agents.

Every reasoning trace, tool call, tool return and final answer, for all 830 questions × 6 agents. ~2.1 GB.

Code, harness and documentation: https://github.com/liuqi6777/search_agent

Setup the trajectories were produced under

All six agents ran through the same harness against the same corpus and retriever, so behavioural differences are not confounded with setup:

• Corpus / retriever — the official BrowseComp-Plus corpus (~100K docs), indexed with Qwen/Qwen3-Embedding-8B. The index is published separately at liuqi6777/Browsecomp-Plus-Indexes.
• Tools — two. search returns the top K=5 documents, each with a docid, retrieval score, and a snippet truncated to 512 tokens. visit returns the full text of a document given its id.
• Budget — 128 turns and 150 minutes per question. Hitting either cap without an answer is recorded as incomplete.
• Sampling — each model's recommended parameters; harness defaults are temperature 1.0, top-p 0.95.

Runs

Directory	Agent	Acc (%)	Gold Rec. (%)	Incomplete (%)
gpt-oss-120b-high	gpt-oss-120b	38.0	52.0	12.8
tongyi-dr	Tongyi-DR	52.2	64.2	38.0
qwen35-35b-a3b	Qwen3.5-35B-A3B	55.2	63.0	1.3
ds4pro	Deepseek V4 Pro	68.6	75.4	0.7
kimi26	Kimi K2.6	69.2	78.2	16.5
glm51	GLM 5.1	74.1	78.7	9.8

gpt-oss-120b-high is the run at high reasoning effort, which is the one the paper reports. A default-effort run exists and is not part of this release.

Accuracy comes from gpt-4o-2024-08-06 following the official BrowseComp judge protocol; every retrieval-side metric in the paper is computed from qrels and is independent of the judge.

Format

One predictions.jsonl per run, one JSON object per line:

{
  "question": "...",
  "answer": "...",
  "prediction": "...",
  "termination": "answer",
  "messages": [ ... the full history, including every tool observation ... ],
  "metadata": {
    "model_name": "...",
    "model_calls": [
      {"turn": 1, "response": {...}, "tool_calls": [{"name": "search", "arguments": {"query": "..."}}]}
    ]
  }
}

termination is answer or incomplete.

messages holds every search result and every full document an agent read, and is most of the size. If you only need behaviour — what was called, in what order, with what arguments — metadata.model_calls carries the same actions without the observations:

import json

with open("glm51/predictions.jsonl") as handle:
    for line in handle:
        record = json.loads(line)
        actions = [
            call["name"]
            for turn in record["metadata"]["model_calls"]
            for call in turn["tool_calls"]
        ]

Recovering retrieved docids (rather than visited ones) means parsing the search observations in messages. Local search results always render a hit as:

N. Document ID: <docid>
Score: <float>
Snippet: <text>

manifest.json records per-run row counts, termination breakdowns, byte sizes and SHA-256 digests.

Preprocessing

Nothing inside messages was rewritten — every observation is exactly as the agent saw it. Three things were done, by scripts/prepare_trajectories.py:

• Deduplicated. A run resumed after a rate-limit storm can hold several rows for one question; one survives, preferring the attempt that answered over the one that errored. Only kimi26 was affected (967 rows → 830).
• Scrubbed. Error records carried the serving stack's raw failure message, including internal hostnames and deployment identifiers. The exception type is kept, the message dropped.
• Sorted into dataset order, so the files are byte-reproducible.

Every run covers all 830 questions with no duplicates.

License

Apache 2.0. The underlying questions, corpus and qrels are from BrowseComp-Plus and carry their own terms.