pith. sign in

arxiv: 2605.27767 · v1 · pith:QPUXG7EInew · submitted 2026-05-26 · 💻 cs.CL · cs.AI· cs.LG

UniMaia: Steering Chess Policies with Language for Human-like Play

Sherman Siu (1) , Lesley Istead (1 , 2) ((1) University of Waterloo , (2) Carleton University) This is my paper

Pith reviewed 2026-06-29 17:45 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.LG
keywords chess policy networkslanguage-conditioned controlparameter-efficient adaptationprompt-based steeringhuman-like playfrozen model conditioningmove prediction benchmarks
0
0 comments X

The pith

A frozen chess policy network can be steered by natural language prompts using a lightweight text encoder and conditioning layer.

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

The paper establishes that a specialized chess policy network trained only on game data can be adapted to follow natural language instructions for choosing openings or adjusting playing strength. It does so by keeping the core policy frozen and adding a small text-processing component plus a conditioning mechanism, avoiding the need for full retraining on combined language and chess data. This setup produces competitive accuracy on tasks that test whether the model follows the given prompt while also performing well on standard move-prediction benchmarks. The results indicate that semantic control over domain-specific policies is achievable with modest overhead.

Core claim

Attaching a parameter-efficient text encoder to a frozen chess policy network and applying a ControlNet-style conditioning mechanism allows natural language prompts to modulate gameplay decisions such as opening selection and strength level. The resulting system reaches state-of-the-art expected accuracy on prompt-conditioned benchmarks and remains competitive with metadata-conditioned baselines on human move prediction tasks, while an auxiliary variant further improves behavioral modeling at a small cost to top-move accuracy.

What carries the argument

Parameter-efficient text encoder plus ControlNet-style conditioning mechanism that modulates activations inside the frozen policy network according to text input.

If this is right

  • Language prompts can select specific chess openings without changing the underlying policy weights.
  • Playing strength can be adjusted through text instructions while the model retains its core move-generation capability.
  • The same conditioning approach yields competitive results on both prompt-following and general move-prediction benchmarks.
  • Auxiliary temporal and behavioral objectives further improve prompt adherence and human-move modeling.
  • Trade-offs appear between the degree of controllability gained and raw predictive performance on unprompted positions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method could extend to other board games or structured decision tasks where a strong policy network already exists.
  • For applications that prioritize maximum playing strength over language control, pure metadata conditioning may still be preferable.
  • The prompt-generation pipeline and metadata-augmented dataset open the door to systematic testing of controllability across many instruction types.
  • If the conditioning mechanism generalizes, similar lightweight adapters might allow language interfaces for other specialized AI systems without joint retraining.

Load-bearing premise

The original chess policy's internal representations remain intact and useful once the text conditioning layer is added on top.

What would settle it

A large drop in move-prediction accuracy on standard chess positions when the conditioning layer is active, even without any prompt, would show the approach fails to preserve domain grounding.

Figures

Figures reproduced from arXiv: 2605.27767 by 2) ((1) University of Waterloo, (2) Carleton University), Lesley Istead (1, Sherman Siu (1).

Figure 1
Figure 1. Figure 1: UniMaia architecture. A frozen Lc0-CF back [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Benchmark performance across monthly checkpoints during training. Most benchmarks continue improving throughout training, although gains diminish over time. Vertical lines indicate year boundaries. behavioral prediction tasks. The largest gains oc￾cur for resignation prediction, suggesting improved modeling of human gameplay behavior beyond next-move prediction. 6.2 Training Dynamics [PITH_FULL_IMAGE:figu… view at source ↗
Figure 4
Figure 4. Figure 4: ABB accuracy as a function of White and Black Elo. Left: top-move accuracy. Right: expected accuracy. old on Lichess (Duplessis, 2019; Fiekas, 2023). We further analyze rating-dependent behavior using the metadata-conditioned ABB benchmark [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the data processing pipeline. Stage 1 performs lightweight string-based processing of [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of training losses between the [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Simple moving averages of the training losses [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Simple moving averages of the training losses [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Simple moving averages of the training losses [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Simple moving averages of the training losses for the learning rate schedule ablations. The WSD-S schedules converge faster initially due to their shorter warmup, but the custom schedule catches up during the decay phase. F.12.1 Template Mixture Ablations We evaluate different sampling mixtures be￾tween LICHESSTEMPLATES-PRETRAIN and LICHESSTEMPLATES-INSTRUCT: 1. 50/50 mixture: sample from both template fa… view at source ↗
Figure 12
Figure 12. Figure 12: Training loss (left) and entropy (right) dur [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Training losses for two local and two server [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Left: raw training loss. Right: simple mov [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 17
Figure 17. Figure 17: Top-move accuracy and expected accuracy across monthly checkpoints. Vertical lines indicate year boundaries. 0 2 4 6 8 10 12 Month index 0.46 0.48 0.50 0.52 0.54 0.56 Top Move Accuracy 0 2 4 6 8 10 12 Month index 0.34 0.36 0.38 0.40 0.42 0.44 Expected Accuracy Benchmark LIF-Aux LGB-Aux LIF-D LIF M1-S Benchmark scores [PITH_FULL_IMAGE:figures/full_fig_p030_17.png] view at source ↗
Figure 15
Figure 15. Figure 15: LOB-P templates partitioned by final top [PITH_FULL_IMAGE:figures/full_fig_p030_15.png] view at source ↗
Figure 19
Figure 19. Figure 19: Top-move accuracy and expected accuracy for LIF-Aux and LGB-Aux during fine-tuning. Unlike UniMaia, UniMaia-Aux often continues improving expected accuracy even when top-move accuracy plateaus, suggesting that auxiliary super￾vision primarily improves policy calibration rather than only top predictions. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Top-move accuracy and expected accuracy for LIF-Aux and LGB-Aux with opening plies excluded. Performance on temporally augmented bench￾marks generally improves throughout fine-tuning, especially when the first 10 plies are omitted. In contrast, performance on the original benchmarks slightly declines later in training, suggesting par￾tial overfitting to the auxiliary supervision setup or prompt format. Th… view at source ↗
Figure 21
Figure 21. Figure 21: Top-5 move probabilities predicted by Uni [PITH_FULL_IMAGE:figures/full_fig_p038_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Quantitative policy continuity diagnostics [PITH_FULL_IMAGE:figures/full_fig_p040_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Quantitative policy continuity diagnostics [PITH_FULL_IMAGE:figures/full_fig_p040_23.png] view at source ↗
Figure 25
Figure 25. Figure 25: Quantitative policy continuity diagnostics [PITH_FULL_IMAGE:figures/full_fig_p042_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Distribution of mean and maximum absolute [PITH_FULL_IMAGE:figures/full_fig_p043_26.png] view at source ↗
Figure 28
Figure 28. Figure 28: Residual concentration ratio for each encoder [PITH_FULL_IMAGE:figures/full_fig_p043_28.png] view at source ↗
Figure 27
Figure 27. Figure 27: Mean and maximum absolute residual up￾dates as a function of ply index for encoder layer 14 and the controllable policy head. L.2.2 Cross-attention Sinks To analyze the interaction between prompts and board representations, we examine cross-attention behavior in the ControlNet branch. Following prior work on attention sinks (Xiao et al., 2023), we compute the proportion of exam￾ples for which the largest … view at source ↗
Figure 30
Figure 30. Figure 30: Mean absolute cross-attention activations [PITH_FULL_IMAGE:figures/full_fig_p044_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: ABB accuracy as a function of White and Black Elo. Left: top-move accuracy. Right: expected accuracy. [PITH_FULL_IMAGE:figures/full_fig_p046_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Number of ABB entries in each rounded White and Black Elo bucket. board games, additional online platforms, curated engine-analysis datasets, and human evaluations would enable more robust assessment of behavioral realism and prompt controllability in real-world settings. Finally, future interpretability work should move beyond descriptive analysis toward causal interven￾tion. Techniques such as activatio… view at source ↗
Figure 33
Figure 33. Figure 33: LIF accuracy as a function of White and Black Elo. Left: top-move accuracy. Right: expected accuracy. [PITH_FULL_IMAGE:figures/full_fig_p047_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Number of LIF entries in each rounded White [PITH_FULL_IMAGE:figures/full_fig_p047_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: LIF-D accuracy as a function of White and Black Elo. Left: top-move accuracy. Right: expected [PITH_FULL_IMAGE:figures/full_fig_p048_35.png] view at source ↗
Figure 36
Figure 36. Figure 36: Number of LIF-D entries in each rounded White and Black Elo bucket. 48 [PITH_FULL_IMAGE:figures/full_fig_p048_36.png] view at source ↗
Figure 37
Figure 37. Figure 37: LIF-T10 accuracy as a function of White and Black Elo. Left: top move accuracy. Right: expected [PITH_FULL_IMAGE:figures/full_fig_p049_37.png] view at source ↗
Figure 38
Figure 38. Figure 38: Number of LIF-T10 entries in each rounded [PITH_FULL_IMAGE:figures/full_fig_p049_38.png] view at source ↗
Figure 39
Figure 39. Figure 39: LGB accuracy as a function of White and Black Elo. Left: top move accuracy. Right: expected accuracy [PITH_FULL_IMAGE:figures/full_fig_p050_39.png] view at source ↗
Figure 40
Figure 40. Figure 40: Number of LGB entries in each rounded White and Black Elo bucket. 50 [PITH_FULL_IMAGE:figures/full_fig_p050_40.png] view at source ↗
Figure 41
Figure 41. Figure 41: M1-S accuracy as a function of White and Black Elo. Left: top move accuracy. Right: expected accuracy [PITH_FULL_IMAGE:figures/full_fig_p051_41.png] view at source ↗
Figure 42
Figure 42. Figure 42: Number of M1-S entries in each rounded White and Black Elo bucket. 51 [PITH_FULL_IMAGE:figures/full_fig_p051_42.png] view at source ↗
Figure 43
Figure 43. Figure 43: M2R accuracy as a function of White and Black Elo. Left: top move accuracy. Right: expected accuracy [PITH_FULL_IMAGE:figures/full_fig_p052_43.png] view at source ↗
Figure 44
Figure 44. Figure 44: Number of M2R entries in each rounded White and Black Elo bucket. 52 [PITH_FULL_IMAGE:figures/full_fig_p052_44.png] view at source ↗
read the original abstract

Recent advances in large language models have enabled natural language to serve as a flexible interface for controlling complex systems, but often at the cost of large-scale multimodal training or weakened domain-specific inductive biases. In structured decision-making domains such as chess, specialized policy networks achieve strong performance but lack semantic controllability, while prompt-conditioned language models are more flexible yet typically exhibit weaker domain grounding. We propose $\textbf{UniMaia}$, a framework for prompt-conditioned policy modulation that adapts a frozen Lc0-based chess policy network using a parameter-efficient text encoder and a ControlNet-style conditioning mechanism. UniMaia enables semantic control over gameplay, including opening selection and player strength, while preserving the pretrained policy representations. We further introduce $\textbf{UniMaia-Aux}$, which incorporates auxiliary temporal conditioning and behavioral prediction objectives. To support this work, we construct a large-scale metadata-augmented Lichess dataset, develop a semi-automated prompt-generation pipeline, and introduce benchmarks spanning both prompt-conditioned and metadata-conditioned settings. UniMaia achieves state-of-the-art expected accuracy on several prompt-conditioned benchmarks and competitive top-move accuracy on general instruction-following tasks, while remaining competitive with dedicated metadata-conditioned approaches on human move prediction benchmarks. UniMaia-Aux further improves expected accuracy and behavioral modeling across several evaluation settings, with modest trade-offs in top-move accuracy. Overall, our results demonstrate that prompt-conditioned control of domain-specific policy networks is feasible without end-to-end multimodal training, while highlighting trade-offs between controllability and predictive performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes UniMaia, a framework for prompt-conditioned policy modulation that adapts a frozen Lc0 chess policy network via a parameter-efficient text encoder and ControlNet-style conditioning mechanism. This enables semantic control (e.g., opening selection, player strength) without end-to-end multimodal training. It also introduces UniMaia-Aux with auxiliary temporal conditioning and behavioral prediction objectives, constructs a metadata-augmented Lichess dataset, develops a semi-automated prompt-generation pipeline, and defines new benchmarks for prompt- and metadata-conditioned settings. Results claim SOTA expected accuracy on several prompt-conditioned benchmarks, competitive top-move accuracy on instruction-following tasks, and competitiveness with metadata-conditioned baselines on human move prediction.

Significance. If the results hold, the work demonstrates a viable path to adding natural-language controllability to strong domain-specific policy networks while preserving their pretrained representations and avoiding large-scale multimodal retraining. The construction of the new dataset, prompt pipeline, and benchmarks constitutes a concrete positive contribution that can support future research on controllable agents in structured domains. The explicit discussion of controllability-performance trade-offs is also useful.

major comments (2)
  1. [§4] §4 (Experimental evaluation): The abstract and results claim SOTA expected accuracy and competitive top-move accuracy, but the provided text supplies no details on dataset splits, error bars, number of runs, or statistical significance testing. This information is load-bearing for verifying the central feasibility claim.
  2. [§3.2] §3.2 (Conditioning mechanism): The claim that frozen Lc0 representations remain intact and useful after parameter-efficient modulation is central to the approach but is not supported by explicit checks such as policy divergence metrics or unconditioned move-prediction accuracy before versus after adding the text-conditioning module.
minor comments (2)
  1. The description of the prompt-generation pipeline would benefit from one or two concrete examples of input metadata and resulting prompts.
  2. [Table 2] Table 2 (or equivalent results table): Ensure all compared methods list their trainable parameter counts and whether they use the same frozen Lc0 backbone for fair comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's potential contribution. We address each major comment below and commit to revisions that strengthen the manuscript.

read point-by-point responses

  1. Referee: [§4] §4 (Experimental evaluation): The abstract and results claim SOTA expected accuracy and competitive top-move accuracy, but the provided text supplies no details on dataset splits, error bars, number of runs, or statistical significance testing. This information is load-bearing for verifying the central feasibility claim.

    Authors: We agree that these experimental details are essential for verifying the claims. In the revised manuscript we will explicitly describe the train/validation/test splits of the metadata-augmented Lichess dataset, report error bars (standard deviation across runs), state the number of independent runs performed for each result, and include statistical significance testing (e.g., paired t-tests) for the reported improvements in expected accuracy. revision: yes

  2. Referee: [§3.2] §3.2 (Conditioning mechanism): The claim that frozen Lc0 representations remain intact and useful after parameter-efficient modulation is central to the approach but is not supported by explicit checks such as policy divergence metrics or unconditioned move-prediction accuracy before versus after adding the text-conditioning module.

    Authors: This observation is correct; the original submission relies on the design of the parameter-efficient adapter but does not provide direct empirical verification. We will add the requested checks in the revision: KL-divergence (or equivalent policy divergence) between the original Lc0 policy and the conditioned policy, plus a direct comparison of unconditioned top-1 move accuracy on held-out positions before versus after the text-conditioning module is attached. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces an architectural framework (frozen Lc0 policy + parameter-efficient text encoder + ControlNet-style modulation) and evaluates it on newly constructed datasets and benchmarks for prompt-conditioned chess move prediction. No equations, first-principles derivations, or fitted-parameter predictions appear in the provided text. Claims of feasibility and trade-offs rest on empirical results from external benchmarks rather than any self-referential reduction, self-citation chain, or ansatz smuggled via prior work. The work is self-contained against its stated evaluation metrics.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated assumption that ControlNet-style conditioning can be applied to chess policies without domain degradation.

pith-pipeline@v0.9.1-grok · 5823 in / 1046 out tokens · 32475 ms · 2026-06-29T17:45:09.172492+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 10 canonical work pages

  1. [1]

    Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Sia- mak Shakeri, Dara Bahri, Tal Schuster, and 1 others

    Maia-2: A unified model for human-ai align- ment in chess.Advances in Neural Information Pro- cessing Systems, 37:20919–20944. Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Sia- mak Shakeri, Dara Bahri, Tal Schuster, and 1 others

  2. [2]

    Tran, Dani Yogatama, and Donald Metzler

    Ul2: Unifying language learning paradigms. arXiv preprint arXiv:2205.05131. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etie...

    work page arXiv 2024

  3. [3]

    Generate natural-language prompts condi- tioned on game metadata. 15

  4. [4]

    Convert metadata spans into parameterized Jinja2 fields

  5. [5]

    Instantiate templates across many games

  6. [6]

    A major challenge is maintaining correctness under iterative modifications

    Manually review outputs and refine formatting rules. A major challenge is maintaining correctness under iterative modifications. Because templates combine independently generated metadata snip- pets, small changes can introduce grammatical or formatting errors across many instantiations.2 Consequently, some prompts contain minor arti- facts such as duplic...

    2048

  7. [7]

    AdamW:AdamW (Loshchilov and Hutter, 2017), used in our initial experiments

    2017

  8. [8]

    Following prior work (Jor- dan et al., 2024; Liu et al., 2025), we use AdamW for all remaining parameters

    Muon:Muon (Jordan et al., 2024) orthogonal- izes updates for 2D non-embedding and non- output parameters. Following prior work (Jor- dan et al., 2024; Liu et al., 2025), we use AdamW for all remaining parameters. 20 Experiment LOB-P LIF-D M1-S Mean µloss,Dec13 Single-step cross-attention (early fusion)0.4696 0.4176 0.4315 0.4396 1.6685 Per-layer cross-att...

    work page arXiv 2024

  9. [9]

    Figure 8 shows the training losses

    NorMuon:NorMuon (Li et al., 2025) extends Muon with neuron-wise normalization using second-order momentum statistics. Figure 8 shows the training losses. Although Muon and NorMuon initially converge more slowly, both quickly outperform AdamW and maintain lower losses throughout training. 0 25000 50000 75000 100000 125000 150000 175000 200000 Positions 1.6...

    2025

  10. [10]

    Unless otherwise specified, DoRA additionally uses rsLoRA and LoRA+

    DoRA:Decomposes low-rank updates into magnitude and direction components (Liu et al., 2024). Unless otherwise specified, DoRA additionally uses rsLoRA and LoRA+

    2024

  11. [11]

    LoRA:Standard LoRA adapters (Hu et al.,

  12. [12]

    without DoRA magnitude decomposi- tion

  13. [13]

    No rsLoRA:Removes rsLoRA scaling (Kala- jdzievski, 2023)

    2023

  14. [14]

    No LoRA+:Removes LoRA+ (Hayou et al., 2024), using identical learning rates for the LoRA matrices A and B instead of the default ratio ηB ηA = 2

    2024

  15. [15]

    Figure 9 shows that all LoRA-based variants con- verge similarly

    Frozen text encoder:Fully freezes the text encoder while retaining the remaining Con- trolNet conditioning architecture. Figure 9 shows that all LoRA-based variants con- verge similarly. Although the frozen text encoder reaches a comparable loss, it consistently under- performs on downstream benchmarks, indicating that adapting the language representation...

    work page arXiv 2024

  16. [16]

    t3:A 0.10B-parameter model initialized from t3-512x15x16h-swa-2815000

  17. [17]

    BT3:A 0.12B-parameter model initialized fromBT3-768x15x24h-swa-2790000

  18. [18]

    Increasing the backbone width generally im- proves performance

    BT4:A 0.21B-parameter model initialized fromBT4-1024x15x32h-swa-5000000. Increasing the backbone width generally im- proves performance. While BT3 achieves the high- est mean benchmark score in this ablation, BT4 attains the lowest loss and remains competitive across all benchmarks. We note that BT3 also outperforms BT4 in the standalone Lc0 evaluation (s...

    work page arXiv 2023

  19. [19]

    Group-level balancing: whether to partition games into groups (e.g., by opening or time control) and allocate separate budgets to each group using Unimax

  20. [20]

    In all settings, plies are sampled uniformly within each game after budgets are assigned

    Per-game budget allocation: whether to al- locate plies proportionally to game length or using Unimax, which caps the contribution of long games. In all settings, plies are sampled uniformly within each game after budgets are assigned. Un- less otherwise specified, the baseline uses Unimax grouped by opening. Table 19 shows that balancing across time con-...

    2024

  21. [21]

    WSD-S (monthly):warmup and decay are applied independently each month

  22. [22]

    WSD-S (annual):a single warmup–stable– decay cycle is applied across all of 2013. Because the original schedule spends nearly all of January 2013 in warmup, we shorten the warmup phase to 512 steps and allocate an additional 512 decay steps, leaving1,024stable-training steps. As shown in Figure 11, the WSD-S schedules reduce loss more rapidly early in tra...

    work page arXiv 2013

  23. [23]

    50/50 mixture: sample from both template families with equal probability

  24. [24]

    Pretrain only: sample exclusively from 24 LICHESSTEMPLATES-PRETRAIN

  25. [25]

    Using only instruct templates yields the strongest average benchmark performance, while the 50/50 mixture achieves comparable downstream accuracy with lower training loss

    Instruct only: sample exclusively from LICHESSTEMPLATES-INSTRUCT. Using only instruct templates yields the strongest average benchmark performance, while the 50/50 mixture achieves comparable downstream accuracy with lower training loss. We therefore adopt the mixed-template configuration in the final model. One possible explanation is that the instruct t...

    2024

  26. [26]

    ChessGPT-play:ChessGPT-base fine- tuned directly for next-move prediction on the 2013 split using the template {prompt}<|endoftext|>{pgn}

    2013

  27. [27]

    ControlNet (ChessGPT-base):The standard ControlNet configuration using the pretrained ChessGPT-base encoder

  28. [28]

    ControlNet (ChessGPT-play):The same ControlNet configuration, but replacing the text encoder with ChessGPT-play

  29. [29]

    ChessGPT-play converges rapidly within the first ∼50 steps (Figure 12)

    ControlNet (ChessGPT-play), end-of-text PGNs:A variant using the explicit end-of-text separator {prompt}<|endoftext|>{pgn} during ControlNet training to reduce dis- tribution shift relative to ChessGPT-play pretraining. ChessGPT-play converges rapidly within the first ∼50 steps (Figure 12). Interestingly, prediction 25 Experiment LOB-P LIF-D M1-S Mean µlo...

    work page arXiv

  30. [30]

    QKV projections:Apply LoRA only to the query, key, and value projections

  31. [31]

    QKV + O projections:Additionally adapt the output projection

  32. [32]

    Table 26 shows that all configurations perform similarly

    QKV + O + FFN:Further extend LoRA to the feedforward layers. Table 26 shows that all configurations perform similarly. Targeting both QKV and O projections achieves the highest mean score, while adding FFN adapters slightly reduces downstream performance despite a marginally lower loss. The differences are small and likely within the variance induced by r...

    work page arXiv 2016

  33. [33]

    Consis- tent with observations from the Kimi Team (Team et al., 2025b), Muon also produces exploding atten- tion logits more frequently than AdamW

    for approximating polar(·) occasionally in- troduces gradient spikes in the LoRA adapters, slightly increasing loss and making NaNs more likely in some execution environments. Consis- tent with observations from the Kimi Team (Team et al., 2025b), Muon also produces exploding atten- tion logits more frequently than AdamW. DoRA adapters are generally more ...

    work page arXiv 2025

  34. [34]

    Maia KDD Testing Set

    within the layer-coupled ControlNet archi- tecture described in Section 4. The frozen Lc0 pol- icy network is conditioned through a text encoder adapted using LoRA with rsLoRA and LoRA+, using rank 16 adapters. Optimization uses NorMuon for all 2D non- embedding, non-output parameters and AdamW for the remaining parameters. We use weight de- cay 0.01, ϵ= ...

    work page arXiv 2013

  35. [35]

    Prompt: The prompt followed by the PGN move list without headers: {prompt}\n\n{pgn}

  36. [36]

    To ensure reliable move extraction, we use outlines (Willard and Louf, 2023) to constrain the generated output with a predefined regular ex- pression

    Elo and time control PGN header4: A PGN header describing both players’ Elos and the time control (Appendix J.4), followed by two newlines and the PGN move list. To ensure reliable move extraction, we use outlines (Willard and Louf, 2023) to constrain the generated output with a predefined regular ex- pression. For ChessGPT-base, the move number may optio...

    2023

  37. [37]

    Ap- plying the chat template twice—once without the generation prompt and once with it—improves performance on the Lichess openings benchmark

    Simple, duplicate: We prompt ChessGPT-chat with {prompt}\n\nPGN: {pgn}\n\nJust write your move. Ap- plying the chat template twice—once without the generation prompt and once with it—improves performance on the Lichess openings benchmark

  38. [38]

    Simple, original: The same prompt as above, except the chat template is applied only once with the generation prompt

  39. [39]

    As with ChessGPT-base, we use outlines (Willard and Louf, 2023) to con- strain generation

    General policy: A prompt adapted from Feng et al., consisting of a prompt to play chess in a given position, a PGN header with certain fields masked by “??”, the PGN move list, and a final instruction to produce the next move. As with ChessGPT-base, we use outlines (Willard and Louf, 2023) to con- strain generation. For ChessGPT-chat, the output must matc...

    2023

  40. [40]

    GM PGN header + prompt: A natural- language prompt inserted between the header and the move list

  41. [41]

    Original PGN header: The original PGN header, followed by the move list

  42. [42]

    time_control

    Elo and time control PGN header: A PGN header specifying player ratings and time con- trol (Appendix J.4), followed by the move list. For autoregressive language models, generated moves are constrained to valid chess formats where possible. Illegal moves are replaced with uniformly sampled legal moves during evaluation. I.5 Model Comparison The evaluated ...

    work page arXiv 2023

  43. [43]

    Play”, while none begin with “Open

    of instruction templates, begin with the verb “Play”, while none begin with “Open”. Many templates instead begin with contextual pre- fixes such as player names or titles. Consequently, prompts that more closely match the training distri- bution tend to produce more stable behavior than semantically equivalent but distributionally differ- ent phrasing. Th...

  44. [44]

    King’s Pawn Opening: You are an anonymous white player rated {elo}, playing against an anonymous black player rated {elo}, using the King’s Pawn Opening

  45. [45]

    Queen’s Pawn Opening: You are an anonymous white player rated {elo}, playing against an anonymous black player rated {elo}, using the Queen’s Pawn Opening

  46. [46]

    English Opening: You are an anonymous white player rated {elo}, playing against an anonymous black player rated {elo}, using the English Opening

  47. [47]

    Both top-move and expected accuracy generally increase with player Elo, reflecting the greater con- sistency of higher-rated play

    Zukertort Opening: You are an anonymous white player rated {elo}, playing against an anonymous black player rated {elo}, using the Zukertort Opening N Performance by Elo Range Figures 31 and 33 show benchmark performance as a function of White and Black Elo. Both top-move and expected accuracy generally increase with player Elo, reflecting the greater con...

  48. [48]

    Qb3 Qc7 27

    Bxd5 exd5 26. Qb3 Qc7 27. Rxd5 Rfd8

  49. [49]

    h3 Rd2 30

    Rxd8+ Rxd8 29. h3 Rd2 30. Qe3 Qd8

  50. [50]

    Rxd1 Qxd1+ 33

    Qxa7 Rd1 32. Rxd1 Qxd1+ 33. Kh2 Qd6+ 34. g3 Qf6 35. Qe3 Qd6 36. Qc5 Qd2

  51. [51]

    300+0." Playing with the white pieces was a player using the username

    a5 Kg7 38. Kg2 Kf6 39. Qb6+ Kg7 40. Qxb7 h5 41. a6 The game in question was a rated blitz chess match hosted on the online chess platform Lichess.org on January 1, 2023. The match featured a rapid time control of five minutes with no additional time per move, denoted as "300+0." Playing with the white pieces was a player using the username "Talca" who hel...

    2023