Pith. sign in

REVIEW 3 major objections 6 minor 28 references

Learning to Plan via Supervised Contrastive Learning and Strategic Interpolation: A Chess Case Study

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A chess model reaches 2593 Elo using only six-ply search.

desk verdict Original approach to planning in chess via contrastively learned embeddings and a direction-of-advantage search, but the 2593 Elo headline rests on unvalidated Stockfish skill caps. read the letter →

arxiv 2506.04892 v1 pith:RTPIVQFP submitted 2025-06-05 cs.CV

Andrew Hamara , Greg Hamerly , Pablo Rivas , Andrew C. Freeman This is my paper
classification cs.CV
keywords contrastivelearningrepresentationchessAIlatentplanningtransformermodelsevaluation-basedsearch
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

The paper attempts to show that strong chess play does not require deep tree search if the model has a well-structured representation of position quality. It trains a transformer to place board states so that distance reflects how similarly Stockfish evaluates them, then chooses moves by advancing in the direction of increasing advantage. With only a six-ply beam search, the system is estimated at 2593 Elo. If the result holds, it suggests that what humans call chess intuition can be modeled as movement through a learned evaluation geometry.

What carries the argument

The central object is the advantage axis $\vec{a} = \mu_{\text{white}} - \mu_{\text{black}}$, the difference between the mean embeddings of forced-mate-for-White and forced-mate-for-Black positions. Move selection scores each legal child position's embedding $z'$ by $\cos(z', \vec{a})$ and keeps the top-$k$; repeating to a fixed depth $S$ yields an embedding-guided beam search. The supervised contrastive loss with evaluation margin $\delta = 0.05$ is what makes distance in this space track evaluative similarity, so the axis points toward winning regions.

What would settle it

Play the trained model against a pool of human-rated players or a second engine with independent Elo calibration; if the Bayesian estimate deviates from 2593 by more than the match error bars allow, the headline rating is an artifact of the capped-Stockfish calibration. A direct behavioral check is whether cosine similarity to the advantage axis ranks legal moves by their actual effect on win probability on a held-out set—if it does not, the embedding is not what is doing the planning.

Watch

Extended reading notes

Core claim

The paper claims that move selection can be carried out entirely in a learned embedding space: a transformer trained with supervised contrastive learning on five million Stockfish-evaluated positions organizes boards so that cosine similarity to an advantage vector ranks moves by their promise, and a greedy beam search of width 3 and depth 6 reaches an estimated 2593 Elo against capped Stockfish opponents. Performance improves with both model size and embedding dimensionality, and visualized game trajectories move smoothly along evaluative trends.

Load-bearing premise

The headline Elo rests on the assumption that the capped-strength settings of Stockfish used as opponents are accurate measures of true playing strength, since the 2593 is computed from match results against those capped opponents.

Editorial extensions

If this is right

  • The base model's estimated Elo rises monotonically with search depth, from 2115 at depth 2 to 2593 at depth 6, with gains tapering by depth 6.
  • Larger model size and higher embedding dimensionality both improve estimated playing strength, so the approach scales with capacity.
  • At a fixed beam width of 3, the greedy beam search can miss strong continuations once an early commitment is made.
  • Because the representation is aligned to learnable win-probability evaluations, the same embedding-guided search generalizes to other perfect-information games such as Go and Shogi.
  • Latent trajectories of real games move smoothly along evaluative trends, giving the learned space an interpretable structure.

Reading between the lines

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

  • The paper leaves implicit that the advantage axis is a form of embedding arithmetic: linear interpolation between forced-mate states acts as a planning direction, suggesting other game-theoretic quantities could be encoded as vector directions in the same space.
  • If the 2593 Elo estimate transfers to matches outside the capped-strength setting, it would support a broader claim that shallow search plus a well-structured representation can replace deep search in other decision-making domains with learnable scalar utilities.
  • A testable consequence is that training the same architecture on human move preferences or on self-play evaluations would produce a different advantage axis and a different playing style, effectively separating the geometry of intuition from the evaluation teacher.
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. This paper proposes SOLIS, a chess engine that replaces deep search with a contrastively trained embedding space. A transformer encoder is trained with supervised contrastive learning on ChessBench positions labeled with Stockfish win probabilities, so that embedding distance reflects evaluative similarity. At inference, the model computes an advantage axis as the difference between mean embeddings of positions with win probability 1.0 and 0.0, and uses a beam search (width 3, depths 2–6) that scores candidate moves by cosine similarity to this axis. In matches against Stockfish configured with UCI_LimitStrength Elo caps, the model's BayesElo rating increases monotonically with search depth, reaching 2593 for the Base model at depth 6. The paper also presents UMAP visualizations and game trajectories.

Significance. If the results hold, the paper makes a useful contribution: it demonstrates that a contrastively learned state representation can support strong, shallow-search move selection, with a simple and interpretable mechanism. The monotonic depth ablation, the public code release, and the use of a standard benchmark dataset are strengths. However, the headline Elo number is only as trustworthy as the UCI_LimitStrength calibration and the BayesElo conversion, issues that the paper currently leaves unexamined; and several claims (scaling with embedding dimensionality, the value of the learned axis over raw evaluation) are not yet backed by the reported experiments.

major comments (3)
  1. [Section 4.1] The headline estimate of 2593 Elo rests on the assumption that Stockfish's UCI_LimitStrength caps correspond to true Elo values and that BayesElo's conversion is valid. UCI_LimitStrength is a heuristic strength-limiting mechanism that is known to be only an approximation, with deviations that vary by version, hardware, and rating range; the paper reports no validation against a calibrated rating list, no confidence intervals, and no per-opponent game counts ('at least 600 games' per depth is insufficient). Without a sensitivity analysis or an independent rating anchor, this central quantitative claim is not established.
  2. [Section 3.4] The advantage axis is computed as the difference of mean embeddings of terminal positions using the same encoder that was trained to organize positions by win probability; the cosine score is therefore, by construction, a linear probe of the training label. The paper should compare this score against a simple baseline that uses the Stockfish win probability directly (or a linear probe trained on those probabilities) to quantify what the learned embedding adds to shallow search, and should discuss the extent to which the 'planning' result is a property of the embedding rather than of the evaluation signal.
  3. [Section 4.1 and Table 1] The abstract and conclusion claim that performance scales with embedding dimensionality, but the only comparison (Small vs. Base) changes H, D, MLP size, and parameter count simultaneously. An isolated embedding-dimensionality ablation is needed to support that claim, or the claim should be softened.
minor comments (6)
  1. [Title and Section 3.4] The method is called 'strategic interpolation,' but no interpolation of embeddings is actually performed during move selection; the term appears only in the title/abstract and in descriptions of trajectory visualization. Please clarify whether interpolation is used in planning or only as an exposition device.
  2. [Section 4.1] Please state the exact number of games played at each Elo cap and the number of caps used; BayesElo's error bars depend on both.
  3. [Figure 1] The caption refers to the advantage axis ®a but does not explain how it is computed; please add a pointer to Section 3.4.
  4. [Abstract] 'Regressive evaluation' is an unusual term; 'regression-based evaluation' would be clearer.
  5. [Section 3.3] The text says 'randomly sampled positions from the ChessBench dataset' but does not state whether the test positions used for evaluation are disjoint from the training set; please clarify the data split.
  6. [Table 2] Since BayesElo provides error bars, please include them in the table or in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training labels and the external Elo benchmark are distinct, and the advantage-axis scoring is a learned evaluation rather than a restatement of the inputs.

full rationale

The paper trains a transformer with supervised contrastive learning on Stockfish win probabilities, then scores moves by cosine similarity to an advantage axis computed from the model's own terminal-position embeddings. This is a supervised distillation of an evaluation function, not a circular derivation: the contrastive objective only encourages embedding closeness for positions with similar win probability, and the advantage axis is a function of the learned embeddings rather than a direct copy of the label. The central quantitative claim, an estimated Elo of 2593, is measured in independent games against Stockfish configured with UCI_LimitStrength, an external benchmark that is not fitted from the model's training data and is not produced by the model's own equations. No parameter is fitted to the match outcomes and then reported as a prediction. The citations to prior work are used for data, tokenization, and baseline components, not to import an unverified uniqueness claim or to define the method's core result. The concern about UCI_LimitStrength calibration is a substantive correctness risk regarding the absolute Elo scale, but it is not a circularity because the rating is obtained from external game outcomes rather than from the training objective. The paper is therefore self-contained against an external benchmark and exhibits no circular step that reduces a prediction to its inputs by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central result depends on a small set of hyperparameters that were chosen without published sensitivity analyses (δ, τ, k) and on an advantage axis computed from the model's own embeddings, which together define the move-selection mechanism. The method further assumes that Stockfish win probabilities are a reliable training target, that the learned static-position embedding transfers to game trajectories, and that Stockfish's UCI_LimitStrength Elo caps are accurate for BayesElo calibration. No new physical or conceptual entities are introduced.

free parameters (4)
  • evaluation margin δ = 0.05
    Used to define positive pairs in SupCon; chosen without sensitivity analysis and directly shapes the embedding's locality.
  • temperature τ = 0.07
    Standard SupCon temperature from prior contrastive papers; no ablation reported.
  • beam width k = 3
    Greedy beam search width; selected without ablation, affects search behavior and Elo.
  • advantage axis a = vector μ_white - μ_black
    Derived from the model's own embeddings of forced-checkmate positions; the move-selection score is cosine similarity to this axis, so the central mechanism depends on this fitted direction.
assumptions (4)
  • domain assumption Stockfish win probabilities are a valid ground-truth evaluation for training.
    The entire training signal is Stockfish's p in [0,1]; see Sec. 3.3.
  • domain assumption The learned embedding generalizes from random ChessBench positions to game-playing trajectories.
    Training uses random positions, but inference follows game trajectories; no explicit verification of distribution shift.
  • domain assumption UCI_LimitStrength Elo caps in Stockfish produce accurate absolute ratings for BayesElo comparison.
    Elo estimates in Sec. 4.1 assume the capped Stockfish strengths are reliable.
  • standard math FEN tokenization from Ruoss et al. adequately represents chess positions.
    Adopted from prior work [21] without modification; reasonable but an input assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Plan via Supervised Contrastive Learning and Strategic Interpolation: A Chess Case Study." pith.science (2026). https://pith.science/paper/RTPIVQFP

@misc{pith2026250604892,
  author       = {Pith},
  title        = {Pith review of: Learning to Plan via Supervised Contrastive Learning and Strategic Interpolation: A Chess Case Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTPIVQFP}},
  note         = {Machine review of arXiv:2506.04892}
}
read the original abstract

Modern chess engines achieve superhuman performance through deep tree search and regressive evaluation, while human players rely on intuition to select candidate moves followed by a shallow search to validate them. To model this intuition-driven planning process, we train a transformer encoder using supervised contrastive learning to embed board states into a latent space structured by positional evaluation. In this space, distance reflects evaluative similarity, and visualized trajectories display interpretable transitions between game states. We demonstrate that move selection can occur entirely within this embedding space by advancing toward favorable regions, without relying on deep search. Despite using only a 6-ply beam search, our model achieves an estimated Elo rating of 2593. Performance improves with both model size and embedding dimensionality, suggesting that latent planning may offer a viable alternative to traditional search. Although we focus on chess, the proposed embedding-based planning method can be generalized to other perfect-information games where state evaluations are learnable. All source code is available at https://github.com/andrewhamara/SOLIS.

Figures

Figures reproduced from arXiv: 2506.04892 by the authors.

Figure 1
Figure 1. UMAP projection of the learned embedding space [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. System overview of the embedding-guided beam search. Candidate moves are embedded, scored via similarity to an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Latent trajectory visualizations of three games em [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 11 canonical work pages

  1. [1]

    The LCZero Authors. 2018. LeelaChessZero. https://lczero.org Accessed: 2024- 04-25

    work page 2018

  2. [2]

    Cyril Burt. 1955. Faster than Thought: A Symposium on Digital Computing Machines. Edited by BV Bowden. Learning to Plan via Supervised Contrastive Learning and Strategic Interpolation: A Chess Case Study KDD-UMC ’25, August 03–07, 2025, Toronto, Canada

    work page 1955

  3. [3]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A Simple Framework for Contrastive Learning of Visual Representations. arXiv:2002.05709 [cs.LG] https://arxiv.org/abs/2002.05709

    arXiv 2020

  4. [4]

    Shaif Chowdhury, Greg Hamerly, and Monica McGarrity. 2024. Active Learning Strategy Using Contrastive Learning and K-means for Aquatic Invasive Species Recognition. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (W ACVW). 848–858. doi:10.1109/WACVW60836.2024.00097

    arXiv 2024

  5. [5]

    Rémi Coulom. 2008. Whole-history rating: A Bayesian rating system for players of time-varying strength. In Computers and Games (LNCS, Vol. 5131) . Springer, 113–124. doi:10.1007/978-3-540-87608-3_10

    work page doi:10.1007/978-3-540-87608-3_10 2008

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs.CL] https://arxiv.org/abs/1810.04805

    arXiv 2019

  7. [7]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929 [cs.CV] https://arxiv.org/abs/2010.11929

    arXiv 2021

  8. [8]

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. 2023. ImageBind: One Embedding Space To Bind Them All. arXiv:2305.05665 [cs.CV] https://arxiv.org/abs/2305.05665

    arXiv 2023

Show all 28 references
  1. [9]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep Residual Learning for Image Recognition. arXiv:1512.03385 [cs.CV] https://arxiv.org/abs/ 1512.03385

    2015 arXiv

  2. [10]

    Dan Hendrycks and Kevin Gimpel. 2023. Gaussian Error Linear Units (GELUs). arXiv:1606.08415 [cs.LG] https://arxiv.org/abs/1606.08415

    2023 arXiv

  3. [11]

    Feng-Hsiung Hsu. 2022. Behind Deep Blue: Building the computer that defeated the world chess champion . Princeton University Press

    2022

  4. [12]

    Feng-hsiung Hsu, Murray S Campbell, and A Joseph Hoane Jr. 1995. Deep Blue system overview. In Proceedings of the 9th international conference on Supercom- puting. 240–244

    1995

  5. [13]

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2021. Supervised Contrastive Learning. arXiv:2004.11362 [cs.LG] https://arxiv.org/abs/2004.11362

    2021 arXiv

  6. [14]

    Matthew Lai. 2015. Giraffe: Using Deep Reinforcement Learning to Play Chess. arXiv:1509.01549 [cs.AI] https://arxiv.org/abs/1509.01549

    2015 arXiv

  7. [15]

    Reid McIlroy-Young, Siddhartha Sen, Jon Kleinberg, and Ashton Anderson. 2020. Aligning Superhuman AI with Human Behavior: Chess as a Model System. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’20) . ACM, 1677–1687. ...

    2020 doi

  8. [16]

    Leland McInnes, John Healy, and James Melville. 2020. UMAP: Uni- form Manifold Approximation and Projection for Dimension Reduction. arXiv:1802.03426 [stat.ML] https://arxiv.org/abs/1802.03426

    2020 arXiv

  9. [17]

    Yu Nasu. 2018. Efficiently Updatable Neural-Network-Based Evaluation Functions for Computer Shogi. Accessed: 2024-04-25

    2018

  10. [18]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

    2019 arXiv

  11. [19]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.000...

    2021 arXiv

  12. [20]

    Tord Romstad, Marco Costalba, Joona Kiiski, Gary Linscott, Yu Nasu, Motohiro Isozaki, and Hisayori Noda. 2008. Stockfish. https://stockfishchess.org Accessed: 2025-04-22

    2008

  13. [21]

    Lewis, Joel Veness, and Tim Ge- newein

    Anian Ruoss, Grégoire Delétang, Sourabh Medapati, Jordi Grau-Moya, Li Kevin Wenliang, Elliot Catt, John Reid, Cannada A. Lewis, Joel Veness, and Tim Ge- newein. 2024. Amortized Planning with Large-Scale Transformers: A Case Study on Chess. arXiv:2402.04494 [cs.LG] https://arxi...

    2024 arXiv

  14. [22]

    Claude E Shannon. 1950. XXII. Programming a computer for playing chess. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science 41, 314 (1950), 256–275

    1950

  15. [23]

    David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Grae- pel, Timothy Lillicrap, Karen Simonyan, and Demis Hassabis. 2017. Mastering Chess and Shogi by Self-Play with a General R...

    2017 arXiv

  16. [24]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1 (Jan. 2014), 1929–1958

    2014

  17. [25]

    Zhenwei Tang, Difan Jiao, Reid McIlroy-Young, Jon Kleinberg, Siddhartha Sen, and Ashton Anderson. 2024. Maia-2: A Unified Model for Human-AI Alignment in Chess. arXiv:2409.20553 [cs.AI] https://arxiv.org/abs/2409.20553

    2024 arXiv

  18. [26]

    J v. Neumann. 1928. Zur theorie der gesellschaftsspiele. Mathematische annalen 100, 1 (1928), 295–320

    1928

  19. [27]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2019. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 [cs.LG] https://arxiv.org/ abs/1807.03748

    2019 arXiv

  20. [28]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 [cs.CL] https://arxiv.org/abs/1706.03762 Received 15 May 2025

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.