CHSM8: A Chess Transformer that Follows Natural-Language Style Prompts

Ilia Badanin

cs.LGOct 18, 2026

CHSM8: A Chess Transformer that Follows Natural-Language Style Prompts

Authors: Ilia Badanin

Abstract

We present CHSM8, a 210M-parameter chess transformer that plays moves conditioned on a free-text description of a playing style, e.g. "go for the Sicilian Defense", "solid, safe, risk-free chess" or "You are Magnus Carlsen in his prime". The model is pretrained without search on 1.07B filtered Lichess games, using a factorized move representation and legal-move masking. It reaches 1837 Elo against a strength-limited Stockfish and, in 1,000-game matches, beat every other public model we tested that was likewise trained without preference tuning, rewards or search, including one with 13 times as many parameters. We then fine-tune it to follow text on a synthetic dataset: Gemma-4-31B describes the playing style of 3M game sides, a 1B Gemma-3 model distilled from the 2.4M of those descriptions that contradict nothing on the board labels the rest, and every one of 30M selected game sides gets five descriptions. Supervised fine-tuning attaches a frozen text encoder through gated cross-attention, and direct preference optimization (DPO) uses the same moves under a mismatched description as the rejected sample. After DPO, style adherence rises on every prompt we test (for example, queens traded early 53% -> 67%, pawn moves 29% -> 32% of all moves, queenside castling 97.3% -> 99.7%), and strength without a prompt stays within noise of SFT (1792 vs. 1797 Elo). Design choices at every stage were made by fixed-budget automated search loops on top of standard architectures and training recipes. We document the full pipeline and the decisions behind it, report the ablations that failed, and evaluate the models across many style prompts and kinds of games, with 1,000 games per setting. The work is fully open source and reproducible: weights, datasets with all training labels, prompts, configurations, inference code and playable demos are public and free to build on.

Figures & tables

tier	conversion	output	process time	wall-clock time
1	download, 162 monthly .pgn.zst	2.51 TB	6.5 h	6.5 h
2a	PGN $\rightarrow$ 16-bit move ids + metadata (cheap filters)	784 GB	170 h	22 h (8 months at a time, 16 workers each); largest month 1.9 h
2b	strength, time forfeits, clock cut	156 GB	4.7 h	4.7 h (one month at a time); largest month 7 min
3	move ids $\rightarrow$ factorized tokens + legal masks	1.39 TB	76 h	about 1.2 h (96 workers); largest month 1.2 h
3p	greedy packing into 512-position rows	1.4 TB	—	12–13 min (162 workers)

Table 1: Data processing tiers, from the raw archive to training rows, as measured in our build. Process time is the sum over all monthly processes; wall-clock time is the real time of the build at the parallelism shown. Each month is an independent process, so at full parallelism the largest month sets the wall-clock time. The released conversion scripts, rewritten for the public dataset, rebuild training rows from its Parquet files in about 2 h on one node (dataset card).

stage	data	order
pretraining chsm8-pt	1.07B games, 78.32B positions, one pass chsm8-pre-training	5% uniform, then 2/3 shuffled mixed with 1/3 newest-to-oldest; from 60% to the end globally reshuffled
games-only chsm8-player-base	26.0M games of the post-training selection (train split), no text, one pass chsm8-post-training	uniformly shuffled
player fine-tunes chsm8-player-*	one player’s own games, own moves only, 2 passes chsm8-top-players	uniformly shuffled
SFT with text chsm8-sft	27.08M described sides, all five bands, one pass; top-player games with “play like { player } ” prompts in about 6% of batches (3 passes over that corpus) chsm8-post-training , chsm8-top-players	uniformly shuffled across time; length-bucketed batches; 10% of descriptions dropped
DPO chsm8	the same 27.08M sides as label-swap pairs, one random band per game; the main model saw 0.56M, the full-pass model all chsm8-post-training	uniformly shuffled; length-bucketed batches

Table 2: Data and data order for each training stage, with the released model under each stage and the released dataset under its data. The DPO model is released as chsm8 ; chsm8-dpo in this report is the same model.

Figure 2: Pretraining loss over the pass for every branch (moving average). The shaded band is the 60–70% bump caused by the 2021-09 to 2022-05 shards; the late rise from about 85% appears in both restarted branches.

round 1 (30 min, 47.7M model)		round 2 (30 min, final architecture)
setting	val loss	soft-penalty weight	val loss
no legality term	3.3389	0.5	3.3038
hard mask	6.4273	1.0	3.3128
soft penalty 0.1	3.3379	2.0	3.3258
soft penalty 0.5	3.3352	4.0	3.3333

Table 3: Illegal-move handling during pretraining (validation loss is always computed without masking; single seed each). The two rounds use different models, so losses are comparable only within a round.

configuration (55 min each)	params	trained	positions seen	throughput	val loss
custom factorized decoder
4L $\times$ 256	3.4M	3.4M	586M	178k	4.959
8L $\times$ 512	25.6M	25.6M	219M	66k	4.417
8L $\times$ 512, lr $6{\times}10^{-4}$	25.6M	25.6M	244M	74k	4.065
+ no dropout	25.6M	25.6M	259M	78k	4.026
+ lr $10^{-3}$	25.6M	25.6M	409M	124k	3.682

Table 4: Selected runs from the first pretraining search round (39 runs, 55 minutes each on one node). “Trained” counts the parameters updated; throughput is positions seen per second of the 55-minute budget; frozen bodies train only our tokenization, embedding and output layers. Learning rate and dropout mattered more than size, and larger models saw far fewer positions in the same time. Figure 4 shows all three search rounds.

architecture	params (total / trainable)	val loss	positions/s	MFU	full pass
20L $\times$ 768, SwiGLU 2560, QK-norm	212.7M / 165.5M	3.0485	305.5k	9.8%	71 h
12L $\times$ 960, SwiGLU 3520, QK-norm	210.5M / 166.2M	3.0598	541.8k	17.1%	40 h
20L $\times$ 768, no QK-norm	212.7M / 165.5M	3.0660	392.3k	12.5%	55 h
12L $\times$ 960, no QK-norm	210.5M / 166.2M	3.0710	501.4k	15.8%	43 h
20L $\times$ 768, plain FF 3840	213.3M / 166.0M	3.1094	463.9k	14.8%	47 h
12L $\times$ 960, plain FF 4800	200.1M / 155.8M	3.1142	633.1k	19.0%	34 h

Table 5: Fixed-data architecture benchmark: 862M positions each, 4 GH200, identical data and validation batches. “Total” includes the frozen cross-attention scaffold. “Full pass” is the expected time for our 78.32B-position pretraining pass on the same node at the measured throughput. The full run of the chosen model held this throughput for the whole pass: about 530k positions/s (5th–95th percentile 505k–546k across the long job segments), since the data was packed in advance and I/O stayed constant.

Figure 5: Playing strength during pretraining (200 games per checkpoint,

\pm

50 Elo; final point 999 games, 95% CI). Strength rises quickly to about 1700 within 5% of the pass and then improves slowly.

opponent	type	parameters	W / D / L	score	Elo advantage (95% CI)
Maia-2 blitz, rating input 2000 ( Tang et al., 2024 )	board policy	23M	102 / 52 / 46	0.640	+100 (+59, +144)
Chess-GPT, 16 layers ( Karvonen, 2024 )	char-level PGN	51M	123 / 43 / 34	0.723	+166 (+122, +216)
chess-gpt2-uci-12x12x768	UCI-text GPT-2	86M	151 / 37 / 12	0.848	+298 (+249, +359)
ChessGPT-base-v1 ( Feng et al., 2023 )	LM on chess text	2.8B	158 / 32 / 10	0.870	+330 (+278, +397)
ChessFormer (Transcendence) ( Zhang et al., 2024 ) , $\leq$ 1500 data	char-level PGN	51M	163 / 30 / 7	0.890	+363 (+309, +434)

Table 6: chsm8-pt head to head, 200 games per opponent, greedy legal moves on both sides, no search.

Figure 6: Games-only continuation search in launch order, in stages of 10, 30 and 60 minutes. Blue: scaffold dropped (its constant output becomes a per-layer bias). Orange: a gated second self-attention in its place, copied or random; slightly lower loss at 30 minutes but 35–40% slower and no better at 60, so the full pass drops the scaffold.

teacher	fitness	distinctiveness	retrieval R@1	parsed
MiniMax-M2.5	0.598	0.515	0.67	100%
Gemma-4-31B	0.579	0.387	0.53	100%
Grok-4.5	0.568	0.335	0.48	100%
GPT-5.2	0.546	0.301	0.45	100%
GPT-5-mini	0.533	0.311	0.46	92%
Apertus-70B-Instruct-2509 (v1)	0.441	0.357	—	100%

Table 7: Teacher screening with the same prompt on the same held-out games. Fitness combines distinctiveness, fidelity and grounding; R@1 is how often a judge retrieves the described game. The Apertus models were screened in a later round, without the retrieval test. Bold names are the two production teachers. The scores rank the models, but read side by side the labels of the stronger models looked very similar, so the final choice was made by reading them.

model	labels contradicted	claim grounding	distillation	vLLM games/s
Gemma-4-31B teacher	4.7%	88.3%	—	—
Gemma-4-31B, second sample (ceiling)	—	87.9%	100%	—
Gemma-3-1B, full fine-tune	4.7%	88.2%	94%	154
Gemma-3-270M, full fine-tune	5.4%	88.3%	90%	193
T5Gemma 2B-2B, LoRA	5.5%	87.4%	94%	23 (HF)
Qwen3-8B, LoRA	4.0%	87.9%	92%	40

Table 8: Labeler students after 65 minutes of fine-tuning on Gemma-4-31B labels (384 test games, single GH200).

Figure 7: Labeler search: distillation score against generation speed for the students of Table 8 . Speeds are vLLM on one GH200, except T5Gemma (Hugging Face generation) and Qwen3-8B and Gemma-4-E4B (vLLM speed proxy with base weights).

trained on $\downarrow$ / tested on $\rightarrow$	Gemma-4-31B	Apertus-v1.5-70B	distilled Gemma-3-1B
Gemma-4-31B	$-0.061$	$-0.015$	$-0.068$
Apertus-v1.5-70B	$-0.014$	$-0.013$	$-0.015$
distilled Gemma-3-1B	$-0.062$	$-0.017$	$\mathbf{-0.072}$

Table 9: Right-minus-wrong description cross-entropy (nats per move; more negative means the text is used more), 1.75M training games per source.

Figure 8: Distinct label strings per band against the number of label sets sampled. Dashed: Gemma-4-31B teacher (up to 1M of its 2.4M sets); solid: distilled Gemma-3-1B; circles: power-law projection, an upper bound. Labels give the projected share at 30M sets, distilled / teacher.

Panel of Figure 9 — Figure 9: The style model: the top six decoder layers attend to the frozen encoder’s token states through gated cross-attention. Blue: frozen; orange: trained in SFT.

model, prompt	CE	move match
chsm8-sft , right name	2.850	56.2%
chsm8-sft , another player’s name	2.962	55.0%
chsm8-sft , no prompt	2.984	54.7%
chsm8-dpo , right name	2.879	55.5%
chsm8-dpo , another player’s name	3.047	53.9%
chsm8-dpo , no prompt	2.984	54.6%

Table 10: Player prompts on held-out games of all 99 top players (2,768 games): cross-entropy (nats per move) and top-1 legal move match on the player’s own moves. Per-game standard deviation of move match is about 9 points, so the standard error is about 0.2 points.

Figure 10: Final SFT model against Stockfish under each evaluation prompt (200 games per style prompt, 300 per request for strong or weak play). Gray band: 95% interval with no prompt (1,000 games).

Figure 11: SFT recipe search: validation cross-entropy on the LLM descriptions (the selection metric) for every run, on the same 4,000 validation games, in launch order: first 10–40-minute runs on one GPU, then 10–120-minute runs on four GPUs (right of the dashed line). Each round was planned from the previous ones; the line is the best so far.

Figure 12: Label-band variants, 1 hour each on the same 4M games, compared on their last periodic validation (steps 6,000–7,000). Right is better (lower cross-entropy with the right description), up is better (larger wrong minus right cross-entropy, so the text is used more); blue is the recipe we kept. The variants drop one band ( no terse , no long ), train on a single band ( terse only , short only , medium only , long only ), sample long bands more often ( long-heavy ), drop the four metadata labels ( no metadata ), order training from short to long bands ( length curriculum ), or add 10–20% label dropout ( + drop ).

prompt	behavior	no prompt
“leave your king in the center and go for an open fight”	castles in 45% of games	94%
“castle queenside and storm the enemy king with your pawns”	39% of castling queenside	10%
“trade queens as early as possible”	queens off by ply 40 in 40%	25%
“push your pawns aggressively up the board and grab space”	0.289 pawn moves per move	0.248
strong-play prompts (8) / weak-play prompts (8)	1765–1836 / 1778–1839 Elo	1828 Elo

Table 11: The full pooled-conditioning run in play (200–300 games per prompt, 1,000 without a prompt).

Figure 13: Wrong minus right cross-entropy on the LLM bands for pooled and per-token conditioning (same 500k games and schedule).

encoder and conditioning	LLM-label CE	mean gap	opening-template gap	relative time
Qwen3-Emb-0.6B, pooled (early recipe)	2.9137	$-0.050$	$-0.006$	1.00
Qwen3-Emb-0.6B, pooled + instruction	2.9109–2.9125	$-0.052$	$-0.006$	1.01
EmbeddingGemma-300m, pooled	2.9114	$-0.054$	$-0.007$	1.10
Qwen3-Emb-4B, pooled + instruction	2.9116	$-0.055$	$-0.003$	1.06
Qwen3-Emb-0.6B, per-token	2.8904	$-0.097$	$-0.070$	1.65
Qwen3-Emb-0.6B, per-token, 3 conditioned layers	2.8983	$-0.086$	$-0.055$	1.57

Table 12: Encoder comparison on identical data (500k games, 1,000 validation games). Gap = right minus wrong label CE.

encoder, pooling	bare	mixed
Qwen3-Embedding-0.6B, last token (early recipe)	98%	53%
Qwen3-Embedding-0.6B, mean	99%	4.5%
Qwen3-Embedding-0.6B, last token, instruction	100%	99%
Qwen3-Embedding-8B, last token	98%	57%
EmbeddingGemma-300m, mean	93%	79%
Qwen3-0.6B language model, mean	85%	6%

Table 13: Player identification from one pooled text vector: share of prompts whose nearest neighbor, written with a different template, names the same player (99 players, chance 1%). No training involved.

Figure 14: DPO search: unweighted

z

score of each run at screening, runs on the horizontal axis in launch order, with the best so far; early rounds used 100 games per prompt, later ones 300 for no prompt and the attack prompt and 150 for the others. The

z

scores are relative to the DPO runs themselves; any DPO run scores far above SFT (about

-20

, Figure 16 ), so the score ranks DPO recipes, not DPO against SFT.

Figure 15: Style adherence against the attack gap for SFT and the 20 DPO runs re-evaluated at 1,000 games per prompt (blue). Red: the main 20-minute run; green: the full pass.

Figure 16: Selection of the main model: summed

z

score against no-prompt Elo for the 20 short DPO runs, the full pass and SFT (star), all at 1,000 games per prompt, each scored twice: blue with weight 0.2 on the saturated queenside metric and 1 elsewhere, gray with the decorrelated weights of Section 9.3 . Red circles: the selected 20-minute run; green circles: the full pass.

Figure 17: Full-pass DPO with pairs from the long band only (solid) vs. a random band of five (dashed), one panel per metric; 1,000 games per point; the dotted line is SFT.

bands	granularity	queens (%)	pawns	queenside (%)	Elo, none	attack gap
SFT	—	51.3	.292	97.1	1789	$-41$
one random	game	65.5	.317	99.1	1788	$-157$
long only	game	67.4	.321	99.1	1811	$-150$
all five	game	65.2	.315	99.2	1793	$-160$
all five, label + game pairs	game	65.3	.315	99.4	1806	$-168$
one random	move	56.9	.298	99.1	1792	$-74$

Table 14: Pair construction, 15-minute runs, 1,000 games per prompt.

run	GPUs $\times$ games	games seen	queens (%)	pawns	queenside (%)	Elo, none	attack gap	$z$	$z$ , decorr.
main	1 $\times$ 128	0.56M	66.7	.318	99.7	1817	$-167$	$\mathbf{+4.9}$	$\mathbf{+4.8}$
C	4 $\times$ 256	2.53M	61.6	.318	99.4	1809	$-157$	$+1.1$	$+0.4$
A, new seed	1 $\times$ 128	0.59M	64.8	.315	98.1	1810	$-139$	$+0.3$	$-0.1$
D, learning rate $\times$ 2.8	4 $\times$ 256	2.50M	61.0	.317	99.8	1796	$-152$	$-0.6$	$-1.1$
B	4 $\times$ 32	0.59M	63.1	.316	99.5	1795	$-140$	$-0.7$	$-0.8$
full pass	8 $\times$ 256	27.08M	57.1	.319	99.6	1807	$-159$	$-0.8$	$-2.0$

Table 15: Batch-size test: identical recipe and 20-minute budget except the full pass; 1,000 games per prompt for every row (the first evaluation of the main run and of the full pass), sorted by

z

z

: summed

z

score against the 20 short DPO runs, weight 0.2 on queenside and 1 elsewhere; decorr.: the decorrelated weights of Section 9.3 .

	style adherence			strength
model	queens (%)	pawns	queenside (%)	Elo, none	attack gap	$z$
SFT (start)	53.0	.290	97.3	1792	$-40$	$-$ 21.0
DPO, 20-min run (main)	66.5	.319	99.7	1797	$-150$	+2.8
DPO, full pass	58.7	.317	99.6	1812	$-155$	$-$ 1.4
DPO, long band only, 25%	64.2	.318	99.1	1790	$-107$	$-$ 1.1
DPO, long band only, 100%	60.9	.308	99.6	1798	$-113$	$-$ 5.6

Table 16: Main results, 2,000 games per prompt. Bold marks the best value per column. 95% precision is about

\pm

1–2 points for shares and

\pm

21 Elo.

model	right	wrong	none	pair accuracy
SFT	2.806	3.094	2.928	0.803
DPO, 20-min run (main)	2.846	3.600	2.932	0.893
DPO, full pass	2.851	4.077	2.933	0.925

Table 17: Validation cross-entropy (nats per move, described side) with the right, a wrong or no description. Pair accuracy is the share of games whose moves are more likely under their own description than under another’s.

Figure 18: Behavior of the model’s own side under the four main prompts, SFT against the main DPO model (bold), 2,000 games per cell. Cell text is the value; color is the change from the same model with no prompt, in units of the column’s spread (red: higher, blue: lower). Under each prompt, DPO moves the targeted behavior further than SFT: more early queen trades, more pawn moves, more queenside castling, and a much larger loss of strength under the attack prompt.

prompt $\rightarrow$ metric	model	no prompt	right prompt	other prompts
“trade queens as early as possible” $\rightarrow$ queens traded by ply 40	SFT	$0.21\pm 0.41$	$0.53\pm 0.50$	$0.21\pm 0.41$
	DPO	$0.23\pm 0.42$	$\mathbf{0.66\pm 0.47}$	$0.14\pm 0.35$
“push your pawns aggressively up the board and grab space” $\rightarrow$ pawn moves per own move	SFT	$0.251\pm 0.072$	$0.290\pm 0.076$	$0.259\pm 0.075$
	DPO	$0.251\pm 0.074$	$\mathbf{0.319\pm 0.074}$	$0.265\pm 0.077$
“castle queenside and storm the enemy king with your pawns” $\rightarrow$ queenside castling, all games	SFT	$0.13\pm 0.34$	$0.89\pm 0.31$	$0.13\pm 0.33$
	DPO	$0.13\pm 0.33$	$\mathbf{0.94\pm 0.25}$	$0.13\pm 0.34$

Table 18: Targeted behavior under the matching prompt, under no prompt and under the other three main prompts: mean

\pm

standard deviation over games, all against Stockfish at UCI_Elo 2000.

n=2{,}000

games for no prompt and the right prompt,

6{,}000

for the other prompts. Figure 19 repeats this in self-play with random descriptions from the dataset as the baseline.

Figure 19: Self-play style evaluation ( Section 3 ): the targeted behavior of the prompted side for SFT and DPO with no prompt, the right prompt and a random test-split description (a different one per game), 1,000 games per bar against chsm8-player-base ; bars: 95% intervals over games.

Figure 20: Player prompts in chsm8-dpo , paired on about 1,000 held-out games per player: mean probability, over legal moves, of the player’s actual move under another top player’s name (gray) and its increase with the right name (red). The vertical axis starts at 0.44.

opponent	params	Elo (95% CI)	ms/move	chsm8-pt	player-base	player-carlsen
public models trained by supervised prediction
Maia-2 blitz	23M	1657 (1626–1684)	3	65.3%	71.8%	71.7%
ChessLFM	231M	1687 (1659–1713)	24	91.5%	93.3%	92.8%
ChessGPT-base	2.8B	1731 (1708–1752)	67	88.0%	93.7%	94.3%
ChessFormer (Transcendence) 1 1 1 A dataset repository holding the checkpoints; we used 50M-High-Elo-1500-No-Elo-Conditioning/2024-04-30---03-07-10_50M-High-Elo-1500-No-Elo-Conditioning/ckpt_1030000.pt .	51M	1733 (1710–1754)	39	86.5%	90.0%	93.9%
chess-gpt2-uci	86M	1784 (1763–1803)	24	82.2%	88.8%	91.0%

Table 19: 1,000-game matches: score of each CHSM8 model against each public model (wins plus half the draws; bold: best of ours) and its Elo against Stockfish at UCI_Elo 2000 (ours: 1837, 1890, 1903). Names link to the exact weights we ran; params of those weights (ours: 210.5M, 166.2M); ms/move at batch size 1 on one GH200 (ours: 13–14).

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

	pretraining	SFT	DPO (main)	DPO (full pass)
initialization	random	chsm8-pt	chsm8-sft	chsm8-sft
text encoder	none	EmbeddingGemma-300m (frozen)	EmbeddingGemma-300m (frozen)	EmbeddingGemma-300m (frozen)
conditioned layers	—	top 6 of 12	top 6 of 12	top 6 of 12
learning rate, decoder	$5{\times}10^{-4}$	$5{\times}10^{-7}$	$2{\times}10^{-7}$	$2{\times}10^{-7}$
learning rate, text cross-attention	—	$8{\times}10^{-4}$	$10^{-4}$	$10^{-4}$
warmup, then cosine to	5%, 10%	5%, 0	5%, 0	5%, 0

Appendix

Table 20: Main hyperparameters of the text-conditioned pipeline.

	games-only continuation ( chsm8-player-base )	player fine-tunes (14 models)
initialization	chsm8-pt , cross-attention replaced by a per-layer bias	chsm8-player-base
trained parameters	all 166M	all 166M
learning rate	$5{\times}10^{-5}$	$3{\times}10^{-5}$
warmup, then cosine to	5%, 0	5%, 0
batch per step	128 rows of 512 positions (32 per GPU, 4 GPUs)	32 rows of 512 positions (1 GPU)
steps	29,706	by player, 2 passes

Appendix

Table 21: Hyperparameters of the games-only branch ( Section 5.4 ).

					move match (%)			CE (nats per move)			Elo
player	FIDE rank	rating	games	positions	base	FT	$\Delta$	base	FT	$\Delta$ %	FT	$\Delta$
Magnus Carlsen	1	2823	17,561	1.74M	55.4	56.9	+1.5	2.969	2.915	$-$ 1.8	1903	+13
Hikaru Nakamura	2	2792	19,997	1.97M	54.0	57.2	+3.2	3.039	2.862	$-$ 5.8	1850	$-$ 40
Fabiano Caruana	3	2784	5,513	584K	54.5	55.6	+1.1	2.960	2.938	$-$ 0.7	1902	+12
Javokhir Sindarov	4	2782	1,371	135K	55.8	56.7	+0.9	2.959	2.898	$-$ 2.1	1897	+7
Nodirbek Abdusattorov	5	2780	5,356	579K	54.4	54.9	+0.5	2.967	2.995	+0.9	1906	+16

Appendix

Table 22: Player fine-tunes (FT) on each player’s held-out test games (own moves only). Bold: the biggest change in each group.

Figure 21: Top-1 move match on the 1,398 held-out games of the 10 current players (ordered by number of games): each player’s own fine-tune (blue; black mark: chsm8-player-base before fine-tuning) and chsm8-dpo prompted with “play like { player } ” (red; gray mark: chsm8-dpo with no prompt).

band	Gemma-4-31B (teacher)	Gemma-3-1B (distilled, clean)	Apertus-v1.5-70B
terse	give up a pawn to blitz the king	give up a pawn for a fast attack	trade pieces and keep pushing pawns
terse	hunt the queen then clean up the board	hunt the king with everything	attack, sacrifice, don’t hold back
short	ignore early material losses to create an overwhelming attack against the exposed king.	play recklessly by offering material to gain a devastating lead in development and initiative.	sacrifice calmly and salt kingside with pawns while trading everything else
medium	play with reckless aggression by trading pawns for rapid development and using the knights to dismantle the opponent’s position quickly.	ignore the early loss of a central pawn to launch a chaotic and bloody offensive that keeps the enemy king running.	play fast and loose with material, keep queens on the board, and pile pawns on the king after sacrificing pieces
long	sacrifice an early pawn for a lead in development and then launch a chaotic offensive that captures almost everything in sight until the king is trapped in the center of the board.	be completely fearless with material and prioritize speed over safety so you can swarm the opponent with pieces while they are still struggling to coordinate their defense in a total collapse.	ignore safe positions and keep sacrificing pieces to crash the kingside with pawns, trade relentlessly, and never stop pressing forward—this blitz pace rewards aggression over caution

Appendix

Table 23: The five descriptions of the same side from each labeler. Gemma and its distilled student both describe the pawn given up for development and the attack on the king. Apertus asks for piece sacrifices and a kingside pawn storm, which the card rules out (no sacrifice, no pawn push at the enemy king).

Panel of Table 24 — Table 24: Bradley–Terry strength (BT) with its 95% interval from all 64,746 games, the games each model played, and the Elo it gets from Stockfish alone ( Table 19 ). Elo rank: position by that Elo; triangles: places gained (green) or lost (red) under BT. Ours in bold. supervised move prediction, strong recent models without search, search.

Chess

Chess research develops and evaluates engines, human-like play models, strategic reasoning systems, and tools for analyzing boards and games.

See all 31 papers

Oct 2, 2026

Language Models that Play Chess and Explain Their Moves

Modern chess engines are silent experts: they play at a superhuman level, but do not offer explanations for their play. On the other hand, language models (LMs) can generate plausible-sounding explanations, but their weak playing strength limits the utility of their explanations. We introduce Queen, a 4B-parameter chess-language model that can explain its moves and plans while playing at the level of a typical Grandmaster. Our novel framework enables domain-specific reasoning through complementary components: an encoder-decoder architecture and an iterative distillation algorithm. This architecture integrates a silent expert chess encoder with an instruction-tuned LM through cross-attention, which we train via a question-answering curriculum to extract chess concepts from the encoder's representations. Building on this domain-adapted model, we iteratively improve its explanations with a natural-language analog of the Bellman update: the model analyzes the positions after its top candidate moves and consolidates them into an explanation of the current position, which is then distilled back into the model. Over seven iterations, our model gains over 900 Elo points (1782 to 2697), substantially surpassing all frontier models on both playing strength and puzzle accuracy, despite containing three orders of magnitude fewer parameters. Furthermore, LM-based evaluations show that our explanations are fluent and approach GPT-5.6-Sol (high) in coherence. The generality of our architecture and training procedure suggests a recipe for applying language models to domains where silent expert encoders are available, like games, robotics, and computer use.

Adithya Bhaskar, Jeffrey Cheng, Danqi Chen

Princeton Language and Intelligence, Princeton University

Chess

Ask PDF

May 18, 2026

Chessformer: A Unified Architecture for Chess Modeling

Chess has long served as a canonical testbed for artificial intelligence, but modeling approaches for its central tasks have diverged. Maximizing playing strength, predicting human play, and enabling interpretability are typically solved with disparate architectures, and these designs are often misaligned with the geometry of the domain. This raises the natural question of whether these objectives require separate modeling paradigms, or if there exists a single architecture that supports them simultaneously. We introduce Chessformer, a unified architecture that advances the state of the art on all three central goals in chess modeling. Chessformer is an encoder-only transformer that represents board squares as tokens, augments self-attention with a novel dynamic positional encoding called Geometric Attention Bias (GAB) that adapts to domain-specific geometry, and predicts actions with an attention-based source-destination policy head. We evaluate Chessformer on each front. First, we develop \maiathree, a family of models for human move prediction that reaches 57.1% move-matching accuracy, significantly surpassing the previous state of the art with fewer than a quarter of the parameters. Second, we integrate Chessformer into Leela Chess Zero, a leading open-source engine, adding over 100 Elo of playing strength and resulting in tournament victories over Stockfish in major computer chess competitions. Third, we show that Chessformer's square-token design makes attention patterns and activations directly attributable to board squares, enabling granular interpretability analyses that prior architectures do not naturally support. More broadly, our results demonstrate that aligning a model's tokenization, positional encoding, and output design with the underlying structure of a domain can yield simultaneous gains in performance, human compatibility, and interpretability.

Daniel Monroe, George Eilender, Philip Chalmers +2

University of Toronto · Williams College

Transformer Interpretability Game-Playing Agents Transformer Positional Encoding Chess

Ask PDF

May 17, 2026

Generalization or Memorization? Brittleness Testing for Chess-Trained Language Models

Recent work has fine-tuned language models on chess data and reported high benchmark scores as evidence that the resulting models can understand the rules of chess, play full chess games at a professional level, or generate human-readable explanations grounded in expert knowledge. We train KinGPT, a 25M-parameter character-level language model trained only on (position, best-move) pairs, who exceeds 3B-parameter ChessGPT on a 600-puzzle mate-in-N suite and 4B-parameter C1-4B over a 20-theme puzzle benchmark. We examine several claims made in existing literature regarding chess-trained language models and assert that their impressive benchmark performance is largely explained by pattern-matching. We also demonstrate how LLM-Modulo, a verifier-in-the-loop framework, raises RedPajama 3B's best move accuracy from 1.2% to 21.2% and move generation validity from 19.3% to 95.3% on mate-in-N chess puzzles, comparable to gains achieved from ChessGPT's fine-tuning on chess-specific web corpora at a fraction of the cost. Our results illustrate how pairing a general LLM with an external verifier offers a more flexible alternative to directly training on synthetic data for well-defined domains. We open source all training/evaluation code, datasets, puzzle samples, and KinGPT model checkpoints for reproducibility.

Ethan Tang

School of Computing and Augmented Intelligence · Arizona State University · Tempe, AZ 85281

LLM Evaluation Language Modeling Chess Language Model Robustness

Ask PDF

Abstract

Chess

Explore similar work

Language Models that Play Chess and Explain Their Moves

Chessformer: A Unified Architecture for Chess Modeling

Generalization or Memorization? Brittleness Testing for Chess-Trained Language Models