Pith. sign in

REVIEW 3 major objections 6 minor 48 references

Communicating Chess Strategies in Natural Language

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Natural language can convey chess engine strategies well enough to improve both human and LLM play while staying far more compact than raw strategy trees.

desk verdict Solid methods paper: strategy verbalization + off-main-line utility eval actually works; engine-as-truth is the known soft limit, not a hidden collapse. read the letter →

arxiv 2607.11486 v1 pith:F7VMVXET submitted 2026-07-13 cs.CL

Langyuan Cui , Chun Kai Ling , Hwee Tou Ng This is my paper
classification cs.CL
keywords chessstrategyverbalizationnaturallanguageenginespartialtreesLLMevaluationpuzzlesolvingexploitabilityhuman-AIalignment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Chess engines already outplay humans, but their search trees and move lists are hard to absorb as usable plans. This paper defines chess strategy verbalization: convert an engine-derived plan for a puzzle into natural language that a solving player can follow. The authors prune the engine tree to a partial strategy that keeps only the top opponent replies, then use an LLM to describe that tree, optionally with concept keywords and self-reflection. They score descriptions not by fluency but by downstream utility: how much better a player scores when guided by the text, including under off-main-line replies chosen to hurt the solver. Results show verbalized strategies help both LLM and human solvers, pure concept lists usually fail to define executable plans, and natural language grows much more slowly than the equivalent JSON tree as more variations are kept.

What carries the argument

Partial strategy tree construction (keep only the top-kstrat engine-ranked opponent moves under engine-optimal solver play) plus LLM verbalization, evaluated by Tree-Expanded Puzzle Playing: at each opponent turn sample top-keval engine moves, roll out one solver reply, and expand the branch that is worst for the solver, scoring the final position by engine win probability.

What would settle it

On held-out puzzles, if players given the verbalized descriptions show no gain over the no-strategy baseline under Tree-Expanded Puzzle Playing as keval grows, or if humans systematically score lower with the verbalizations than with the same-size JSON trees, the claim that language is an effective compact medium would fail.

Watch

Extended reading notes

Core claim

Natural language is a workable, compact medium for communicating engine strategies: descriptions of pruned partial strategy trees raise puzzle performance for both LLM and human players above no-strategy baselines, while remaining far smaller than the full JSON strategy representation as opponent branching increases.

Load-bearing premise

Top engine-ranked opponent moves are treated as the right branches both to keep in the strategy and to stress-test during evaluation, so the reported scores only upper-bound true worst-case play if those rankings miss the lines that actually break the player.

Editorial extensions

If this is right

  • Strategy explanations should be judged by how well players perform with them, including off the main line, not only by fluency or concept coverage.
  • High-level chess concepts alone rarely define an executable plan; concrete move contingencies are needed.
  • Natural language can compress strategy trees more gracefully than structured JSON as branching factor grows.
  • LLM-generated descriptions can help LLM solvers more than human solvers (and vice versa for human-written text), so audience matters.
  • Self-reflection helps only when the verbalizer can already spot strategic omissions; weaker models can make descriptions worse.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same prune-then-verbalize pattern could apply to other perfect-information games that already have strong engines but weak human-facing explanations.
  • Coaching tools may need separate generation targets for human readers versus LLM proxies, given the preference misalignment the paper reports.
  • Limited worst-branch tree expansion is a practical stand-in for full exploitability when querying humans or large models is expensive.
  • Concept keywords may work better as compression aids after the concrete tree is known, rather than as a substitute for it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces chess strategy verbalization: converting engine-derived (partial) strategies into natural-language descriptions that help a solving player recover and execute the strategy on chess puzzles. It proposes (i) a pipeline that builds a kstrat-pruned partial strategy tree from Stockfish and verbalizes it with LLMs under strategy/concept guidance and optional self-reflection, and (ii) Tree-Expanded Puzzle Playing, which scores descriptions by downstream engine win-probability after probing up to keval opponent replies and expanding the worst branch for the solver. Experiments on 100 Lichess puzzles (LLM solvers o3 and gpt-oss-120b), ablations on kstrat/keval, a Xiangqi transfer study, and a 30-person human study with balanced assignment show that verbalized strategies improve over no-strategy baselines for both LLMs and humans, that concept-only guidance is weak, that NL is more compact than JSON as branching grows, and that LLM vs human preferences for description style can misalign—especially off the main line.

Significance. If the results hold, the work cleanly carves out a pedagogically motivated task distinct from post-hoc move commentary and abstract concept extraction, and supplies a utility-based evaluation that is more principled than surface metrics (BLEU, G-Eval). Strengths include controlled baselines (No Strategy, C, S, R, Optimal), multi-run reporting with standard deviations, explicit ablations on branching factors, a human study with Elo-stratified participants and round-robin strategy assignment, and an honest limitations section on player confounding and LLM–human misalignment. The compactness result (Fig. 9) and the demonstration that pure concepts do not yield executable strategies are useful contributions for interpretable game AI and for NLG evaluation via downstream task performance.

major comments (3)
  1. [§5, Tree-Expanded Puzzle Playing; Tables 2–3; Fig. 8] Stockfish rankings define both the partial strategy tree (top-kstrat opponent moves, §4.1) and the off-main-line probe (top-keval then worst branch, §5). The authors correctly note that Tree-Expanded Puzzle Playing only upper-bounds exploitability and assumes near-optimal solver play when ranking branches. This is load-bearing for insight (a) and for the claim of a “robust” evaluation framework. The manuscript should more explicitly quantify how method rankings change between keval=1 and keval=3 (Tables 2–3 already hint at a flip: LLM descriptions look stronger on the main line; human-written ones better for humans off-main-line) and discuss the risk that engine-suboptimal but human-plausible opponent moves are systematically under-sampled. A short sensitivity experiment (e.g., sampling some non-top engine moves, or reporting rank correlation of methods across keval) would make the centr
  2. [§6.3; Appendix D; Tables 2–3] Human evaluation (§6.3, Appendix D) terminates early when the participant’s move on the expanded branch is not the engine-optimal move, so reported scores are upper bounds and off-main-line expansion is incomplete. The authors apply the same modified procedure to the LLM player for fairness, which is appropriate, but this weakens the strength of the claim that human-written descriptions better communicate off-main-line variations (Table 2, keval=3). The paper should state more clearly what fraction of puzzles terminate early under each condition and whether the Human vs LLM gap survives when restricted to fully expanded trajectories, or treat the human off-main-line comparison as suggestive rather than conclusive.
  3. [§6.2, Table 1] Table 1 reports means ± one std over five runs but no significance tests or confidence intervals on ΔBase / pairwise method differences. Several key contrasts have overlapping intervals (e.g., o3 as solver: V-S 0.813±0.037 vs S 0.851±0.009; RV-S vs S). Given that the headline is that verbalization is “beneficial but lossy” relative to JSON and better than no strategy, the paper should either run paired tests (or bootstrap over puzzles) or temper language where differences are not clearly separated. This is fixable without new experiments beyond re-analysis of existing runs.
minor comments (6)
  1. [Abstract; Figure 1; §5] Figure 1 and the abstract promise “objective evaluation,” but evaluation still depends on the solving player class (§8). A one-sentence clarification in the abstract or intro that utility is always relative to a player class would avoid overclaim.
  2. [§4.2; §6.2] Concept keywords are taken as gold from Lichess player votes (§4.2, §6.1). A brief note on noise/coverage of those tags would help readers interpret the weak V-C / C results.
  3. [Appendix B; §6.4] Xiangqi results (Appendix B, Table 4) are interesting but under-discussed in the main text; one sentence in §6.4 or the conclusion on what the larger S–Optimal gap implies for “JSON as instructions vs hints” would strengthen the generalization claim.
  4. [Appendix C; Table 8] Qualitative examples (Appendix C) are valuable; labeling illegal or hallucinated moves more systematically (as in the d5 knight/pawn case in Table 8) would make the failure modes easier to scan.
  5. [Abstract; §1; §4–6] Typo/consistency: “chessstrategy verbalization” appears without a space in the abstract and §1; standardize to “chess strategy verbalization.” Also “kstrat” / “keval” formatting is inconsistent in places (subscripts vs inline).
  6. [§2; Appendix A] Related work on human–AI alignment in chess (Maia / Maia-2) is cited; a short explicit contrast—verbalizing engine strategy vs modeling human style—would situate the pedagogical goal more sharply.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: verbalization quality is scored by independent play outcomes, not by re-reading the description or by construction from fitted inputs.

full rationale

This is an empirical systems paper, not a first-principles derivation. The load-bearing chain is: (i) build a pruned partial engine strategy tree (top-kstrat opponent moves), (ii) verbalize it with an LLM under optional concept/self-reflection guidance, (iii) score the description by how well a solving player (LLM or human) plays under Tree-Expanded Puzzle Playing, converting final Stockfish centipawns to win probability via the fixed Lichess map and reporting ΔBase / ΔOpt. None of these steps is equivalent to its inputs by construction. Verbalization is explicitly lossy (NL typically underperforms raw JSON S; concept-only V-C is weak; self-reflection is mixed). Evaluation is not self-definitional: success is not “does the text match the tree” or “does an LLM judge the text fluent,” but whether a separate player, given only the text, reaches better engine-evaluated terminal states than a no-strategy baseline, including off-main-line branches. Shared use of Stockfish for tree construction and terminal scoring is a disclosed ground-truth assumption (§3, §5), not a fitted parameter renamed as a prediction, and not a self-citation uniqueness theorem. There is no self-citation load-bearing chain, no uniqueness imported from the authors’ prior work, and no ansatz smuggled in via citation. Comparative claims (verbalized > no-strategy; NL more compact than JSON as kstrat grows; human–LLM misalignment) remain independently falsifiable under the stated protocol. Score 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The central claim rests on treating a strong engine as strategy ground truth, on hand-chosen pruning/evaluation widths, on LLM verbalizers/players as competent enough agents, and on puzzle win-probability as pedagogical utility—not on new physical entities. Free parameters are experimental knobs; axioms are domain modeling choices; invented entities are the task and evaluation procedure themselves.

free parameters (5)
  • kstrat (partial strategy tree opponent branching)
    Default set to 3 (varied 1–5 in ablations); directly controls how much of the engine strategy is available to verbalize and how large JSON grows.
  • keval (Tree-Expanded Puzzle Playing opponent samples)
    Default 3 (varied 1/3/5); controls how harshly off-main-line play is tested and thus reported player scores.
  • self-reflection rounds
    Fixed at three critic–verbalizer iterations for RV-* methods; affects description length and sometimes quality.
  • Lichess puzzle evaluation subset size
    100 puzzles chosen from Lichess; human study uses 30 book puzzles—selection and size are experimental choices that bound statistical strength.
  • Xiangqi fixed puzzle length
    Set to 5 plies for all positions because the source dataset lacks lengths; affects transfer results in Table 4.
assumptions (5)
  • domain assumption Stockfish (and Fairy-Stockfish) best moves and centipawn/win-probability evaluations may be treated as ground-truth strategy and position quality for puzzles.
    Stated in §3 (“sufficiently accurate that we can treat their move recommendations and board state evaluations as the ground truth”) and used throughout construction and scoring.
  • domain assumption Top engine-ranked opponent moves are a reasonable heuristic for strategically important and exploitable branches.
    Used for kstrat pruning (§4.1) and keval sampling (§5); authors note it assumes optimal solver play and only upper-bounds exploitability.
  • domain assumption LLMs have enough chess knowledge to verbalize strategies and act as scalable solving-player proxies.
    Design premise in §4.2 and §5; partially stress-tested by human study and Xiangqi transfer where the premise weakens.
  • domain assumption Lichess formula score = (1 + e^{-0.00368·cp})^{-1} is an appropriate utility metric for puzzle outcomes.
    Adopted in §6.1 Metrics without independent derivation; maps engine cp to reported player scores.
  • standard math Standard game-tree and partial-strategy formalisms (mapping states to moves; ply-limited puzzles).
    Background definitions in §3 used to define complete vs partial strategies and puzzle length.
invented entities (3)
  • Chess strategy verbalization (task)
    purpose: Name and scope the problem of faithfully compressing engine strategies into natural language for pedagogical play, distinct from post-hoc move commentary.
    Introduced in abstract/§1/§3 as the paper’s primary task framing.
  • Tree-Expanded Puzzle Playing
    purpose: Approximate worst-case evaluation of a description-guided player by iteratively expanding the opponent branch that most hurts the solver.
    Defined in §5 and Figure 5 as the core evaluation procedure; not a prior standard chess metric under this name.
  • Partial strategy tree with kstrat opponent pruning
    purpose: Make engine strategies compact enough to verbalize while retaining key contingencies.
    Construction method in §4.1 / Figure 3; engineering object rather than a physical entity, but central to the pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Communicating Chess Strategies in Natural Language." pith.science (2026). https://pith.science/paper/F7VMVXET

@misc{pith2026260711486,
  author       = {Pith},
  title        = {Pith review of: Communicating Chess Strategies in Natural Language},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7VMVXET}},
  note         = {Machine review of arXiv:2607.11486}
}
read the original abstract

Chess engines have long achieved superhuman playing strength. However, the underlying strategy behind their move suggestions is difficult for human players, even skilled ones, to comprehend. Motivated by this, we propose the task of chess strategy verbalization, which is to describe chess strategies in natural language. We design (i) a pipeline for verbalizing strategies and (ii) an evaluation framework for objective evaluation of generated strategy descriptions. Our experiments show that natural language is a promising and interpretable medium for communicating strategic information to both human and LLM players. We glean additional interesting insights, including (a) the importance of evaluating strategies beyond the main line, (b) the limitations of pure concept-based descriptions, and (c) the limitations of relying on LLMs rather than humans for evaluation.

Figures

Figures reproduced from arXiv: 2607.11486 by the authors.

Figure 1
Figure 1. Overview of our proposed frameworks. We selectively extract strategies from chess engines and verbalize them with LLMs, and robustly evaluate descriptions through Tree-Expanded Puzzle Playing ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Chess strategy verbalization. The verbalizer outputs a natural language strategy (top) that a player utilizes to perform well in the puzzle (bottom). not formally optimal, they are sufficiently accurate that we can treat their move recommendations and board state evaluations as the ground truth. The task of strategy verbalization in chess is to convert an ES into a natural language descrip￾tion that enables a player… view at source ↗
Figure 3
Figure 3. Left: Game tree for a hypothetical 3-ply puzzle rooted at the initial board state. Nodes (resp. ) represent board states where it is the solving player’s (resp. opponent’s) turn to move, while leaves contain engine evaluation scores. Edges represent moves, ordered from best to worst for the player-to-move from left to right. For simplicity, both players are assumed to have three legal moves. Center: Reachable subtre… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustration of when a strategy may appear sound against an optimal opponent, but is easily exploitable in off-main-line play. The cross player is following a strategy that prioritizes moves in the following order: (1) Win if possible, (2) Prevent opponent wins, (3) Ce…
Figure 5
Figure 5. Figure 5: Tree-Expanded Puzzle Playing. (1) The solving player makes a first move. (2) At opponent node , we sample keval legal moves. (3) For each sam￾pled move, we query the solving player for a response and evaluate the resulting state with a chess engine. (4) We continue wit…
Figure 6
Figure 6. Figure 6: Performance of solving players as kstrat increases. 0 250 500 750 1000 1250 1500 1750 2000 ELO 0 1 2 3 4 Count Participant Elo Rating 800 800 1300 > 1300 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 9
Figure 9. Figure 9: Token Count of Verbalized vs JSON Strat￾egy as kstrat increases. Token count measured using o3 tokenizer, taken over puzzle instances in one run for each kstrat. show the generated descriptions in Appendix C.3. Generalization to Low-Resource Games. We also study strate…
Figure 8
Figure 8. Figure 8: Performance as keval increases. V-S by o3. keval consistently degrades the performance of all methods, illustrating how our proposed evaluation framework exposes flaws in the players’ strategy. Growth in Size of Strategy Representation. While JSON effectively conveys s…
Figure 10
Figure 10. Figure 10: A chess puzzle from the Lichess Puzzle Dataset. Black to move. C.1 Concept-Guided vs. Strategy-Guided Descriptions Verbalized strategies under the V-S/V-C/V-S+C set￾tings are shown in [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 12
Figure 12. Figure 12: Human evaluation user interface for play￾ing puzzles in the No Strategy setting. Participants submit one move in each board state by dragging the move on the interactive chess board. that can be reached under Tree-Expanded Puzzle Playing with keval ≤ 3, assuming the p…
Figure 13
Figure 13. Figure 13: Human evaluation user interface for play￾ing puzzles in the Human/LLM setting. The strategy description is displayed below the corresponding chess board throughout all board states in the puzzle. Additional Details We recruited a total of 30 university students. Durin…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 4 linked inside Pith

  1. [1]

    Science , volume =

    David Silver and Thomas Hubert and Julian Schrittwieser and Ioannis Antonoglou and Matthew Lai and Arthur Guez and Marc Lanctot and Laurent Sifre and Dharshan Kumaran and Thore Graepel and Timothy Lillicrap and Karen Simonyan and Demis Hassabis , title =. Science , volume =

  2. [2]

    Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Bridging the gap between expert and language models: Concept-guided chess commentary generation and evaluation , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2025

  3. [3]

    Learning to Generate Move-by-Move Commentary for Chess Games from Large-Scale Social Forum Data

    Jhamtani, Harsh and Gangal, Varun and Hovy, Eduard and Neubig, Graham and Berg-Kirkpatrick, Taylor. Learning to Generate Move-by-Move Commentary for Chess Games from Large-Scale Social Forum Data. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics. 2018

    2018

  4. [4]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    Automated chess commentator powered by neural chess engine , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

  5. [5]

    2002 , booktitle =

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing , title =. 2002 , booktitle =

    2002

  6. [6]

    Feng, Xidong and Luo, Yicheng and Wang, Ziyan and Tang, Hongrui and Yang, Mengyue and Shao, Kun and Mguni, David and Du, Yali and Wang, Jun , journal=. Chess

  7. [7]

    Complete chess games enable

    Zhang, Yinqi and Han, Xintian and Li, Haolong and Chen, Kedi and Lin, Shaohui , booktitle=. Complete chess games enable

  8. [8]

    Kolasani, Sai and Saplin, Maxim and Crispino, Nicholas and Montgomery, Kyle and Davis, Jared Quincy and Zaharia, Matei and Wang, Chi and Wang, Chenguang , journal=

Show all 48 references
  1. [9]

    arXiv preprint arXiv:2412.12119 , year=

    Mastering board games by external and internal planning with language models , author=. arXiv preprint arXiv:2412.12119 , year=

    arXiv

  2. [10]

    arXiv preprint arXiv:2212.08195 , year=

    Improving chess commentaries by combining language models with symbolic reasoning engines , author=. arXiv preprint arXiv:2212.08195 , year=

    arXiv

  3. [11]

    2008 , publisher=

    The Rating of Chessplayers: Past and Present , author=. 2008 , publisher=

    2008

  4. [12]

    Campbell, Murray and Hoane Jr, A Joseph and Hsu, Feng-hsiung , journal=. Deep. 2002 , publisher=

    2002

  5. [13]

    Stockfish , author =

  6. [14]

    Luo, Qinyu and Ye, Yining and Liang, Shihao and Zhang, Zhong and Qin, Yujia and Lu, Yaxi and Wu, Yesai and Cong, Xin and Lin, Yankai and Zhang, Yingli and others , booktitle=. Repo

  7. [15]

    Yang, Dayu and Simoulin, Antoine and Qian, Xin and Liu, Xiaoyi and Cao, Yuwei and Teng, Zhaopu and Yang, Grey , booktitle=. Doc

  8. [16]

    Summarizing Source Code using a Neural Attention Model

    Iyer, Srinivasan and Konstas, Ioannis and Cheung, Alvin and Zettlemoyer, Luke. Summarizing Source Code using a Neural Attention Model. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics. 2016

    2016

  9. [17]

    2018 , booktitle =

    Hu, Xing and Li, Ge and Xia, Xin and Lo, David and Jin, Zhi , title =. 2018 , booktitle =

    2018

  10. [18]

    Feng, Zhangyin and Guo, Daya and Tang, Duyu and Duan, Nan and Feng, Xiaocheng and Gong, Ming and Shou, Linjun and Qin, Bing and Liu, Ting and Jiang, Daxin and others , booktitle=. Code

  11. [19]

    Madaan, Aman and Tandon, Niket and Gupta, Prakhar and Hallinan, Skyler and Gao, Luyu and Wiegreffe, Sarah and Alon, Uri and Dziri, Nouha and Prabhumoye, Shrimai and Yang, Yiming and others , journal=. Self-

  12. [20]

    Liu, Yang and Iter, Dan and Xu, Yichong and Wang, Shuohang and Xu, Ruochen and Zhu, Chenguang , booktitle=. G-

  13. [21]

    arXiv preprint arXiv:2508.10925 , year=

    gpt-oss-120b & gpt-oss-20b model card , author=. arXiv preprint arXiv:2508.10925 , year=

    arXiv

  14. [22]

    Gonzalez and Hao Zhang and Ion Stoica , booktitle=

    Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica , booktitle=. Efficient Memory Management for Large Language Model Serving with

  15. [23]

    2003 , publisher=

    Chess Tactics for Kids , author=. 2003 , publisher=

    2003

  16. [24]

    Xu, Jiannan and Li, Gujie and Jiang, Jane Yi , booktitle=. A

  17. [25]

    2026 , note =

    Wait, Theo and Anand, Viswanathan , title =. 2026 , note =

    2026

  18. [26]

    Proceedings of the International Joint Conference on Artificial Intelligence , pages=

    Approximate Exploitability: Learning a Best Response , author=. Proceedings of the International Joint Conference on Artificial Intelligence , pages=

  19. [27]

    Advances in neural information processing systems , volume=

    A unified game-theoretic approach to multiagent reinforcement learning , author=. Advances in neural information processing systems , volume=

  20. [28]

    Scientific American , volume=

    A grandmaster chess machine , author=. Scientific American , volume=. 1990 , publisher=

    1990

  21. [29]

    Fichter, Fabian and The Fairy-Stockfish Community , url =. Fairy-

  22. [30]

    Xiangqi-r1: Enhancing spatial strategic reasoning in

    Chen, Yuhao and Liu, Shuochen and Lyu, Yuanjie and Zhang, Chao and Shi, Jiayao and Xu, Tong , journal=. Xiangqi-r1: Enhancing spatial strategic reasoning in

  23. [31]

    Superhuman

    Brown, Noam and Sandholm, Tuomas , journal=. Superhuman

  24. [32]

    Science , volume=

    Human-level play in the game of diplomacy by combining language models with strategic reasoning , author=. Science , volume=. 2022 , publisher=

    2022

  25. [33]

    Exploring large language models for communication games: An empirical study on

    Xu, Yuzhuang and Wang, Shuo and Li, Peng and Luo, Fuwen and Wang, Xiaolong and Liu, Weidong and Liu, Yang , journal=. Exploring large language models for communication games: An empirical study on

  26. [34]

    Light, Jonathan and Cai, Min and Shen, Sheng and Hu, Ziniu , journal=. Avalon

  27. [35]

    Hidden-role games:

    Carminati, Luca and Zhang, Brian Hu and Farina, Gabriele and Gatti, Nicola and Sandholm, Tuomas , journal=. Hidden-role games:

  28. [36]

    Maia-2: A unified model for human-

    Tang, Zhenwei and Jiao, Difan and McIlroy-Young, Reid and Kleinberg, Jon and Sen, Siddhartha and Anderson, Ashton , journal=. Maia-2: A unified model for human-

  29. [37]

    Evaluation of human-

    Siu, Ho Chit and Pe. Evaluation of human-. Advances in Neural Information Processing Systems , volume=

  30. [38]

    Aligning superhuman

    McIlroy-Young, Reid and Sen, Siddhartha and Kleinberg, Jon and Anderson, Ashton , booktitle=. Aligning superhuman

  31. [39]

    On the critical role of conventions in adaptive human-

    Shih, Andy and Sawhney, Arjun and Kondic, Jovana and Ermon, Stefano and Sadigh, Dorsa , journal=. On the critical role of conventions in adaptive human-

  32. [40]

    Acquisition of chess knowledge in

    McGrath, Thomas and Kapishnikov, Andrei and Toma. Acquisition of chess knowledge in. Proceedings of the National Academy of Sciences , volume=. 2022 , publisher=

    2022

  33. [41]

    arXiv preprint arXiv:1912.12191 , year=

    Explain your move: Understanding agent actions using specific and relevant feature attribution , author=. arXiv preprint arXiv:1912.12191 , year=

    1912 arXiv

  34. [42]

    2021 IEEE Conference on Games (CoG) , pages=

    Some chess-specific improvements for perturbation-based saliency maps , author=. 2021 IEEE Conference on Games (CoG) , pages=. 2021 , organization=

    2021

  35. [43]

    Towards Piece-by-Piece Explanations for Chess Positions with

    Spinnato, Francesco , journal=. Towards Piece-by-Piece Explanations for Chess Positions with

  36. [44]

    Bridging the human--

    Schut, Lisa and Toma. Bridging the human--. Proceedings of the National Academy of Sciences , volume=. 2025 , publisher=

    2025

  37. [45]

    , author=

    Unveiling Concepts Learned by a World-Class Chess-Playing Agent. , author=. Proceedings of the International Joint Conference on Artificial Intelligence , pages=

  38. [46]

    Extracting Search Trees from

    Chen, Sixing and Li, Ji-An and Cakir, Saner and Akcali, Sinan and Lee, Kayla and Mattar, Marcelo G , journal=. Extracting Search Trees from

  39. [47]

    Proceedings of the International Joint Conference on Artificial Intelligence , pages=

    Accelerating best response calculation in large extensive games , author=. Proceedings of the International Joint Conference on Artificial Intelligence , pages=

  40. [48]

    2026 , howpublished =

    2026

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.