RevengeBench Traces

Data for RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments (NeurIPS 2026, Evaluations and Datasets Track)

• Project Page
• Code
• Croissant metadata (Croissant 1.1 with Responsible AI and provenance fields)
• Provenance record

Authors: Babak Rahmani, Sebastian Dziadzio, Joschka Strüber, Sergio Hernández-Gutiérrez, Matthias Bethge (Tübingen AI Center, University of Tübingen). Published and maintained by the Bethge Lab.

Per-message traces and per-round simulation outcomes for RevengeBench tournaments, normalized into four Parquet tables joined by tournament_id and config_id. Each tournament is one LLM coding agent trying to reconstruct one of 75 hidden target policies (15 per arena: BattleSnake, Halite, HuskyBench, RoboCode, RobotRumble) as executable code over five refinement rounds. The release covers 1,725 tournaments from 19 learner configurations: 12 LLMs, a five-seed Gemma 4 31B subset, no-probe and description ablations for DeepSeek v4 Pro and Gemma 4 31B, and a Codex-harness ablation for GPT-5.5.

License

The dataset structure (schemas, identifiers, derived metrics) is released under CC-BY-4.0 (https://creativecommons.org/licenses/by/4.0/).

Per-row model output content is governed by the upstream model provider's terms of service; see experiment_configs.output_license_notice for the per-config notice. Downstream re-use of message content is the user's responsibility.

Limitations and intended use

Intended uses: research on agent code-editing behavior, strategy recovery, scientific reasoning analysis, evaluation methodology, and behavioral ablations of LLMs.

Out of scope: training models on the contents (subject to provider TOS); claims about "true" model capabilities (the dataset is a snapshot in time per model); use as evidence in adversarial-evaluation contexts beyond what the paper claims.

Sensitive content: message content is LLM reasoning about programming tasks; the data involves no human subjects. Tool outputs occasionally quote public metadata of the open-source game engines and CodeClash code the agents inspected (maintainer e-mail addresses in package manifests and license headers, a developer home-directory path in upstream files). A scan of all message text found no API keys or access tokens.

Not included: some experiments reported in the paper are not part of this release, e.g. the Bayesian Program Inference baseline, the passive and description ablations for GPT-5 and GPT-5.4-mini, and the player-versus-player evaluation.

The full Responsible AI documentation (limitations, biases, use cases, social impact, sensitive information, collection and preprocessing, maintenance) is in the rai:* fields of croissant.json.

Schema

The release contains four Parquet tables. Map- and list-valued columns in tournaments are stored as JSON-encoded strings (parse with json.loads; keys are round numbers as strings). Missing data in these columns is encoded as the JSON string "null", never as a Parquet NULL, so json.loads succeeds on every row.

Table	Column	Type
tournaments	tournament_id	string
tournaments	experiment_kind	string
tournaments	model_slug	string
tournaments	model_display_name	string
tournaments	ablation_condition	string
tournaments	condition	string
tournaments	game	string
tournaments	target_hash	string
tournaments	target_display	string
tournaments	seed	int64
tournaments	n_rounds	int64
tournaments	distances	string (JSON: round → double, null if evaluation failed)
tournaments	best_distance	double
tournaments	best_round	int64
tournaments	evaluation_failed	string (JSON: round → bool)
tournaments	exit_status	string (JSON: round → string)
tournaments	per_round_failures	string (JSON: round → {eval_fail, no_submit})
tournaments	per_round_probes	string (JSON: round → int)
tournaments	per_round_probe_failures	string (JSON: round → int)
tournaments	total_probe_count	int64
tournaments	total_probe_attempts	int64
tournaments	per_round_usage	string (JSON: round → {prompt_tokens, completion_tokens, reasoning_tokens, answer_tokens, api_calls, cost})
tournaments	config_id	string
tournaments	dataset_version_added_in	string
tournaments	total_prompt_tokens	int64
tournaments	total_completion_tokens	int64
tournaments	total_reasoning_tokens	int64
tournaments	total_answer_tokens	int64
tournaments	total_api_calls	int64
tournaments	total_cost	double
messages	tournament_id	string
messages	round	int64
messages	turn_index	int64
messages	role	string
messages	content	string
messages	thinking	string
messages	tool_calls	string
messages	extra_keys	list<item: string>
messages	timestamp	double
simulations	tournament_id	string
simulations	round	int64
simulations	sim_index	int64
simulations	kind	string
simulations	opponent_target_hash	string
simulations	winner	string
experiment_configs	config_id	string
experiment_configs	model_name	string
experiment_configs	model_provider	string
experiment_configs	system_prompt	string
experiment_configs	instance_template	string
experiment_configs	agent_kwargs	struct<step_limit: int64, cost_limit: double, max_context_chars: int64, keep_recent_observations: int64, max_past_output_chars: int64, max_output_chars: int64>
experiment_configs	arena_kind	string
experiment_configs	output_license_notice	string
experiment_configs	notes	string

Quick start

import json

import pyarrow.parquet as pq

tournaments = pq.read_table("tournaments.parquet").to_pandas()
messages = pq.read_table("messages.parquet").to_pandas()

# JSON-encoded columns parse on every row (missing data decodes to None).
tournaments["distances"] = tournaments["distances"].map(json.loads)

# Example: messages JOIN tournaments to filter by model and game.
joined = messages.merge(
    tournaments[["tournament_id", "model_slug", "game"]],
    on="tournament_id",
    how="inner",
)
print(joined.head())

Known gaps

• messages.thinking is null for every row; provider reasoning content was not retained.
• System messages of the v1.0 runs have empty content; the system prompt is stored once per configuration in experiment_configs.system_prompt.
• Grok-4.1-fast (model_slug grok-4-1-fast) has no rows in messages and zero token, call and cost counters; its tournaments and simulations rows are complete.
• simulations holds per-game outcomes (20 opponents × 6 rounds per tournament) for the v1.1 and v1.2 runs. For the v1.0 runs it only contains 375 placeholder rows with null opponent_target_hash and winner; their per-round results are in tournaments.distances.
• 37 tournaments have null best_distance / best_round (no round produced an evaluable hypothesis) and 15 ended before round 5. Failed rounds are kept and flagged in evaluation_failed, exit_status and per_round_failures.
• experiment_configs.model_name of the gpt-5-4-mini configuration (c_084b09c572dfa8a1) wrongly reads openai/gpt-5-mini; those runs used GPT-5.4-mini (openai/gpt-5.4-mini).
• experiment_configs.arena_kind is BattleSnake for every row; a configuration applies to all five arenas, so use tournaments.game.
• v1.0 message text contains <REDACTED-...> placeholders from the anonymized review release (see Provenance).

Versions

Each row in tournaments carries a dataset_version_added_in tag. To restrict messages or simulations to a specific dataset version, join through tournaments.dataset_version_added_in rather than filtering the message/simulation tables directly.

• v1.0 — initial release.
• v1.1 — adds GPT-5, GPT-5.4-mini, GPT-5.5 (low and medium reasoning effort), GPT-oss-120b, and Grok-4.1-fast.
• v1.2 — adds the GPT-5.5 Codex harness ablations (model_slug gpt-5-5-codex-low / gpt-5-5-codex-high, ablation_condition codex-low / codex-high): the same model driven by the Codex CLI agent instead of the default harness, at two reasoning-effort settings, across all five arenas. Drops the never-populated per_round_sim_distances, per_round_sim_distance_stds, and simulations_absent_rounds columns from tournaments.
• v1.2.1 — metadata only, no data changes: adds croissant.json (Croissant 1.1 with Responsible AI and PROV-O provenance fields) and provenance.json, and updates this card for the NeurIPS 2026 camera-ready version.

Codex ablation notes (v1.2)

The Codex harness logs differ from the default harness in a few ways that show up in the data:

• messages.role is assistant (agent commentary) or tool (command executions, MCP tool calls, file changes); the original codex event type is kept in messages.extra_keys, and command/tool payloads are JSON in messages.tool_calls.
• messages.thinking and messages.timestamp are always null.
• Cost fields (total_cost, per_round_usage[*].cost) are 0.0 — codex exec does not report cost.
• Prompt-token counts include cached input tokens: codex resumes the session each round, replaying the conversation, so per-round prompt_tokens grow with round number.

Provenance

• Source. All rows are extracted from logs of the RevengeBench harness (bethgelab/revenge-bench, MIT), run by the authors between April and July 2026 against public model APIs. The game arenas and the strategy corpus from which targets and opponents were drawn come from CodeClash (MIT). The learner harness is mini-SWE-agent, or the OpenAI Codex CLI for the v1.2 ablation. No human annotation is involved: all labels (action distances, failures, winners) are computed by the harness.
• Review release. The v1.0 runs were first published anonymized for double-blind review as RevengeBench/revengebench-traces (2026-05-07). This dataset is its de-anonymized successor and supersedes it: the 1,125 v1.0 tournaments carry the same tournament_ids, configurations, distances and per-run message counts as the review release. Free text in the v1.0 rows was scrubbed of author names, institutions, cluster names, e-mail addresses and home-directory paths for review, which is why it contains <REDACTED-...> placeholders. Rows added in v1.1 and v1.2 were not scrubbed.
• Versions. Each tournaments row records the version it was added in (dataset_version_added_in).
• Machine-readable record. provenance.json lists the authors, upstream sources, release history with Hub commits, SHA-256 and row count of every file, per-run row counts and collection windows, and the v1.0 consistency check against the review release. The same information is in the prov:* fields of croissant.json.

Citation

@inproceedings{rahmani2026revengebench,
  title={RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments},
  author={Rahmani, Babak and Dziadzio, Sebastian and Str{\"u}ber, Joschka and Hern{\'a}ndez-Guti{\'e}rrez, Sergio and Bethge, Matthias},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year={2026},
  eprint={2606.26094},
  archivePrefix={arXiv}
}