Papers & Publications

We introduce Moreau Arena, a contamination-resistant benchmark for LLM strategic reasoning. Agents design creature builds (animal choice plus stat allocation) for a novel stochastic combat game whose rules exist only in the prompt. Three tournaments over the same benchmark core—13 agents in T001/T002, expanded to 15 in T003—and 2,609 best-of-7 series reveal a sharp boundary condition.

Under vague rules with no feedback (T001), baselines dominate and LLMs win only 37.5% of cross-group series. With exact formulas, meta-context hints, and per-game adaptation (T002), the leaderboard reverses: LLMs win 89.75%. A targeted ablation (T003) removes only the meta-context hints while preserving exact rules and feedback. Performance does not collapse uniformly; instead the field splits. GPT-5.2-Codex and Claude Opus remain strong without hints, while three models from three providers freeze on a single identical build and never adapt.

Meanwhile, GPT-5.4—released during this study—ranks 14th of 15 despite exploring 79 builds, demonstrating that frontier capability does not automatically transfer to novel strategic domains. A prompt sensitivity analysis using a semantically equivalent paraphrase yields Kendall τ = 1.0 (perfect rank preservation), providing strong initial evidence that the key findings are not obvious artifacts of a single prompt surface form.

1. A newly authored game environment designed to reduce contamination risk and to admit a derivable dominant strategy under a frozen configuration snapshot.
2. Two 13-agent round-robin tournaments (8 frontier LLMs, 5 baselines), totaling 1,559 best-of-7 series under different information scaffolding regimes.
3. A third no-meta ablation tournament (T003) with 15 agents and 1,050 series, revealing that some models remain strategically robust without hints while others collapse to rigid default builds.
4. A negative-transfer result: GPT-5.4, a state-of-the-art frontier model, ranks 14th of 15 despite active exploration, showing that general capability does not imply strategic competence in novel domains.
5. A prompt sensitivity validation (Kendall τ = 1.0) providing strong initial evidence that the T003 findings are not obvious artifacts of a single prompt surface form.
6. Evidence that the T001 → T002 performance reversal implicates comprehension and feedback rather than an absence of strategic capability, with model-specific meta-initialization dependence as the key moderator.

The paper covers the frozen 6-animal benchmark (T001–T003). In parallel, Moreau Arena operates as a living competitive layer with seasonal roster expansions. Season 1 introduces 14 animals, WIL regen mechanics, and new cross-animal interactions (anti-heal, proc suppression, charge ignoring dodge). A separate tournament with 9 LLMs and 5 baselines across 91 all-pairs series produces dramatically reshuffled rankings—notably, GPT-5.4 rises from 14th (T003) to 2nd (S1), suggesting that the expanded strategy space unlocks capabilities the narrow benchmark could not measure.

Season 1 results are available on the S1 Leaderboard. Season data is not part of the paper’s frozen benchmark but demonstrates the auto-balancer and roster expansion design described in the paper.

Additional publications by Victor Stasiuc. Moreau Arena v3 and Moreau Island will move from placeholder status to DOI-linked records after Zenodo upload.

The full benchmark—simulator, tournament data, prompts, analysis scripts, and all 89 invariant tests—is open source and available for reproducible evaluation.

@article{stasiuc2026moreau,
  title   = {Moreau Arena: Not All LLMs Need Hints to Reason
             Strategically},
  author  = {Stasiuc, Victor},
  year    = {2026},
  note    = {Config hash: b7ec5885, 89 invariant tests,
             15 agents across 3 tournament protocols,
             2609 series, PSI tau = 1.0}
}

1. Clone & Install

The full benchmark, including simulator, tournament data, and all 89 invariant tests, is available from the repository. No GPU required.

git clone https://github.com/supervitek/moreau-arena-paper.git
cd moreau-arena-paper
pip install -r requirements.txt

2. Verify Config Integrity

The core simulator config is frozen and hash-verified. Any modification will cause invariant tests to fail.

# Verify config hash (must match paper)
python scripts/verify_config_hash.py

# Expected: b7ec588583135ad640eba438f29ce45c10307a88dc426decd31126371bb60534

3. Run Invariant Tests

python -m pytest tests/test_invariants.py -v
# Expected: 89 passed

4. Recompute Rankings

# Recompute BT rankings for each tournament
python -m analysis.bt_ranking data/tournament_001/results.jsonl
python -m analysis.bt_ranking data/tournament_002/results.jsonl
python -m analysis.bt_ranking data/tournament_003/results.jsonl