Complete Chess Games Enable LLM Become A Chess Master

Haolong Li; Kedi Chen; Shaohui Lin; Xintian Han; Yinqi Zhang

REVIEW 5 major objections 6 minor 34 references

Complete Chess Games Enable LLM Become A Chess Master

T0 review · 5 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A 3-billion-parameter LLM, fine-tuned only on chess positions and Stockfish's best moves, reaches an Elo of 1788.

desk verdict A genuine full-game LLM chess result is buried under an Elo computation that contradicts the paper's own match scores. read the letter →

arxiv 2501.17186 v2 pith:5PH3PKSW submitted 2025-01-26 cs.AI cs.CLcs.LG

Yinqi Zhang , Xintian Han , Haolong Li , Kedi Chen , Shaohui Lin This is my paper · ORCID

classification cs.AIcs.CLcs.LG

keywords largelanguagemodelschesssupervisedfine-tuningForsyth-EdwardsNotationStockfishEloratinggameplayingimitationlearning

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 claims that a large language model can become a strong chess player through plain supervised fine-tuning, with no search, no reinforcement learning, and no hand-coded chess knowledge. It encodes each board position as a Forsyth-Edwards Notation string paired with Stockfish's best move, and trains a 3-billion-parameter model on over 20 billion tokens of such pairs drawn from complete games. The resulting model, ChessLLM, can finish full games against Stockfish and is reported to reach an Elo of 1788 when allowed up to ten sampled legal moves. The authors also argue that data quality, not just quantity, matters: supervision on long-round endgame data improves Elo by about 350 points over short-round data alone. If these results hold, they would suggest that imitating a strong engine's move choices can give an LLM genuine strategic competence in an abstract, rule-governed game.

What carries the argument

The load-bearing mechanism is the FEN-best-move pair as a text-format training instance. Forsyth-Edwards Notation (FEN) gives a fixed-length, one-line encoding of a board position, so unlike PGN move histories it does not grow with game length; the model can be trained to predict the best move from the current position alone. The dataset is built with Stockfish: short-round positions from Chessdb with search depths of 12 to 50, and long-round endgames from Stockfish self-play with depths of 50 to 200. The argument is that a constant-size FEN state plus a strong move label is enough for an autoregressive LLM to internalize both move legality and strategic choice, which is what lets ChessLLM finish whole games.

What would settle it

The rating claim can be tested by calibrating Stockfish skill level 2 against players with known human ratings and then playing ChessLLM against the same players; if the resulting win rates do not match the expected-score curve for an 1788 player, the headline Elo is not reliable.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that complete chess games can be turned into a next-token-prediction task and solved by supervised fine-tuning. ChessLLM starts from a pretrained 3-billion-parameter autoregressive model and is fine-tuned on FEN-to-best-move examples generated by Stockfish searches of depth 12 to 200. In 100-game matches, it wins 61% of games at Stockfish skill level 0, 56% at level 1, and 30% at level 2, which the paper converts, via Stockfish's skill-level-to-Elo mapping, into ChessLLM Elo ratings of 1632, 1753, and 1788 respectively. The 1788 figure is the headline: a professional-level amateur rating achieved purely by imitation. The paper further shows that adding long-round endgame data produced a roughly 350-Elo improvement over short-round data, and that legal-move accuracy above 99.8% is accompanied by a sharp jump in Elo.

Load-bearing premise

The whole Elo claim rests on the assumption that the conversion from Stockfish's internal difficulty levels to Elo ratings is accurate and that 100 games are enough to measure the model's true strength.

Editorial extensions

If this is right

A 3-billion-parameter LLM can finish full chess games from textual input alone, so full-game competence does not require a chess-specific architecture or explicit search tree.
Dataset composition is a lever comparable to scale: long-round data supervision adds about 350 Elo over short-round data, so collecting deeper, endgame-heavy games is a direct route to stronger play.
Legal-move accuracy is a gating factor: the model's Elo climbs sharply once legal-move accuracy passes about 99.8%, implying that eliminating illegal moves is where the largest immediate gains come from.
Head-to-head play against other general-purpose language models shows large win-rate margins, indicating that chess skill in this setup comes from the chess-specific fine-tuning rather than from generic language ability.
Complete-game evaluation against Stockfish, with win rate and Elo as metrics, offers a template for assessing LLM performance in interactive, long-horizon tasks beyond static benchmarks.

Reading between the lines

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

Beyond the paper: the 1788 Elo is reported under a 10-sample legal-move selection budget; at pass@1 the model is weaker, so the headline number is a sampling-budgeted ceiling rather than the model's unaided strength.
Beyond the paper: the same FEN-plus-best-move recipe should transfer to other rule-based games with compact textual state encodings, but the paper does not test that transfer.
Beyond the paper: the skill-level-to-Elo polynomial is an engine-relative calibration; an independent rating against human-listed opponents would be needed to know whether 1788 is a true human-comparable Elo.
Beyond the paper: because the model imitates Stockfish's choices, its style and errors should resemble the training engine's, so a natural next test is whether ChessLLM exhibits the same tactical blind spots as Stockfish at shallow search depths.

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.

Desk Editor's Note

read the letter

Two things you should know. The headline result, a 1788 Elo, does not survive contact with the paper's own Table 1. For skill level 2, ChessLLM scores 30.5/100 against Stockfish rated 1570-1720. A 1788-rated player would expect about 69% against the midpoint 1645, more than double the observed 30.5%. Solving the same formula backward gives an implied rating near 1503, about 285 points lower. The other rows show the same pattern. That is not a calibration wobble; the Elo column cannot have been computed from the reported matches.

The second thing is that there is a real core under the mess. The paper demonstrates that a 3B LLM, supervised on FEN-best-move pairs, can complete whole chess games with high move legality and can beat ChessGPT and other text LMs in head-to-head games. That is new relative to the cited ChessGPT work, and the long-round vs. short-round data comparison (350 Elo points) is a plausible finding, even if the absolute numbers are suspect.

What the paper does well: the FEN-based formatting is clean, the dataset scale is large (20B tokens), the head-to-head LM win rates are a useful benchmark, and the move legality analysis is properly reported.

The soft spots are serious. The skill-to-Elo polynomial in Eqs. (3)-(4) is undeclared, uncited, and likely mis-transcribed (the e3/e2 terms are unexplained). The evaluation protocol allows up to 10 samples per move, which inflates strength and is different from the one-sample protocol used for the pass@1 numbers. No confidence intervals or seeds are reported for the 100-game matches, and the paper says "winning 61%" when Table 1 shows 61 wins, 10 draws, 29 losses (66%). The "only ones" claim is an overreach as stated.

The core idea is fine; the empirical claim of amateur-master level is not established. A serious referee could ask for a corrected Elo analysis and a clearer statement of the sampling protocol, and the paper might be resubmitted as a modest but useful demonstration. As it stands, I would not cite the headline number, and I would tell the authors to recompute.

For peer review: yes, send it out—the setup is legitimate and the full-game capability is worth refereeing. But the referee should be told to check the arithmetic in Table 1.

Referee Report

5 major / 6 minor

Summary. The paper introduces ChessLLM, a 3B-parameter GPT-like model fine-tuned on FEN-best-move pairs generated by Stockfish, with the goal of playing complete chess games. It describes a large dataset (claimed to exceed 20B tokens), evaluates legal-move and best-move accuracy on static positions, plays full games against Stockfish at skill levels 0-2 and against several other language models, and reports that ChessLLM reaches an Elo of 1788, with a 350-Elo improvement from adding long-round data. The core claim is that simple supervised fine-tuning on text-formatted chess positions is sufficient to reach a professional-level rating.

Significance. If the Elo claim were supported, this would be a notable demonstration that an LLM can acquire full-game chess skill through imitation learning on text-formatted positions, and the proposed full-game evaluation protocol would be a useful complement to static evaluation sets. The direct comparisons with RedPajama, LLaMA, and ChessGPT are a useful empirical contribution. However, the central quantitative claim is internally inconsistent with the paper's own match results, the data-generation protocol is described contradictorily, and no code, data, or model checkpoints are provided, so the result is currently not reproducible.

major comments (5)

[Table 1, §5.1] The reported Elo ratings are not consistent with the win/draw/loss totals. For skill level 2, the midpoint of the stated Stockfish range is 1645 and ChessLLM scores 30.5/100, so the implied rating from the standard performance-rating calculation is about 1503, not 1788; a 1788-rated player would be expected to score roughly 0.69. The same pattern appears for skill level 0 (implied about 1510 versus the reported 1632) and skill level 1 (implied about 1572 versus 1753). The paper does not report the initial rating, K-factor schedule, or aggregation method that would let the reader reproduce the table, so the headline Elo of 1788 is unsupported.
[§3 vs §5.2] Section 3 says best moves were generated with search depths of 12-50 for short-round data and 50-200 for long-round data, limiting each search to two seconds, while Section 5.2 says the training data were generated with depth=1 and timelimited=0.1. These statements describe different data-generation procedures, and the paper never reconciles them. Because the 350-Elo data-quality claim depends on the distinction between short- and long-round supervision, the actual protocol must be stated unambiguously.
[§5.1, Eqs. (3)-(4)] The text says the method converts between Stockfish skill level and Elo rating, but Eq. (3) expresses skill as a function of Elo, and the inverse mapping is never given. The polynomial is also presented without a source, so the Stockfish Elo ranges in Table 1 (e.g., 1350-1440 for skill level 0) cannot be checked independently.
[§6.2] The Elo matches use up to 10 sampling iterations to obtain a legal move, whereas the pass@1 and other win-rate evaluations use a single sample per position. The headline rating therefore describes a top-10 sampling policy, not the single-sample policy used elsewhere in the paper. The paper should report single-sample Elo as well, disclose the sampling budget in the abstract, and describe how the uncertainty intervals in Table 1 were computed.
[§1, §5.1] The claim that ChessLLM reaches a 'professional-level' Elo of 1788 also assumes that the Stockfish skill-level-to-Elo conversion is equivalent to human rating scales. This external-validity assumption should be justified with an independent calibration, for example by comparing Stockfish skill levels against human databases or an independent engine; otherwise the rating is only a chess-engine skill-level label.

minor comments (6)

[§5.2] The sentence 'These two datasets are from different domains' is ambiguous: the preceding sentence says the eval set was produced with the same parameters as the train set, so the reader cannot tell which two datasets are being contrasted.
[§1] The statement that the authors are 'the only ones using a large language model for chess that can complete full games' is not supported by a survey or by the cited literature and should be removed or substantiated.
[Throughout] Equations (1)-(2) contain raw LaTeX artifacts (for example 'Elo N = Elo O') and should be typeset cleanly, with all symbols (K, RA, RE, EloS, EloM) defined in the text.
[Table 1] The table layout is confusing: 'ChessLLM' appears as a standalone row/column header and the ranges are not explicitly linked to Eqs. (3)-(4); please reformat and specify how the ranges were obtained.
[§8] The limitations section is a list of future work rather than a discussion of the evaluation's limitations; it should address the sampling protocol, the dependence on Stockfish-generated labels, and the external validity of the Elo estimate.
[§3, §6] The 20B-token dataset and the fine-tuned model are not released, which prevents other researchers from reproducing the data-quality comparisons or the Elo measurements.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular step reduces ChessLLM's headline Elo to its training inputs; the Elo-versus-score mismatch in Table 1 is an internal-consistency concern, not a circular derivation.

full rationale

ChessLLM's reported capability is an empirical fine-tuning result: the training set is Stockfish-generated FEN-best-move pairs, and the evaluation is a match record against Stockfish with Elo computed by the explicit formulas in Eqs. (1)-(4). The evaluation shares its teacher with the training labels, so the benchmark is not fully independent of the training signal, but no claimed result is defined in terms of the target answer or fitted to the headline number. The Elo numbers in Table 1 are, however, not reproducible from the table's own win/loss/draw counts via Eq. (2): at skill level 2 (Stockfish midpoint approximately 1645), a 1788-rated player would be expected to score about 69%, while ChessLLM scored 30.5/100; the score-implied rating is about 1503. Similar gaps occur at skill levels 0 and 1. This is a serious numerical contradiction, but it is a correctness or soundness defect, not a circularity: the derivation chain is formally checkable and fails, rather than being vacuously satisfied by construction. The paper's self-citations (Li et al. 2024; Chen et al. 2024) appear only in generic remarks about scaling and everyday LLM use and are not load-bearing for the chess claims. No uniqueness theorem is invoked, and no fitted parameter is renamed as a prediction. Under the required standard of exhibiting a specific reduction, no circular step is established.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's headline Elo is not an externally measured quantity; it is an arithmetic consequence of an unvalidated skill-level conversion and 100 games. All experimental conclusions depend on Stockfish both generating the labels and serving as the opponent, so the benchmark is not independent of the training signal. The main free parameters are evaluation hyperparameters and the arbitrary depth and time limits for label generation.

free parameters (3)

Skill-to-Elo polynomial coefficients = a=37.247, b=40.852, c=22.294, d=0.311, with e=(Elo-1320)/1870
Load-bearing for all Elo scores in Table 1; presented without derivation or citation in Section 5.1, Eqs. (3)-(4).
Sampling budget for Elo games = temperature=0.7, top_k=50, up to 10 samples
Used in the actual-game Elo evaluation in Section 6.2; the reported strength depends on this sampling allowance.
Stockfish search depth limits for data generation = 12-50 (short), 50-200 (long), time limit 2 seconds
Chosen to build best-move labels in Section 3; the quality of the dataset and the reported improvements depend on this arbitrary choice.

assumptions (4)

domain assumption Stockfish's selected move is the ground-truth best move for each FEN position.
Used throughout to construct labels and to score best-move accuracy; Section 3.
domain assumption A policy conditioned only on the current FEN, with no move history, is sufficient for strong full-game play.
The dataset is FEN-move pairs; Section 3 argues that human players focus on the current position rather than past moves.
ad hoc to paper The Stockfish skill level-to-Elo mapping given by Eq. (3) is valid.
Introduced without source or derivation; the entire Elo claim in Table 1 depends on this mapping.
domain assumption Next-token supervised fine-tuning on best-move labels transfers to full-game performance without search.
Core training setup in Section 4; no reinforcement learning or search is used.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Complete Chess Games Enable LLM Become A Chess Master." pith.science (2026). https://pith.science/paper/5PH3PKSW

@misc{pith2026250117186,
  author       = {Pith},
  title        = {Pith review of: Complete Chess Games Enable LLM Become A Chess Master},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PH3PKSW}},
  note         = {Machine review of arXiv:2501.17186}
}

read the original abstract

Large language models (LLM) have shown remarkable abilities in text generation, question answering, language translation, reasoning and many other tasks. It continues to advance rapidly and is becoming increasingly influential in various fields, from technology and business to education and entertainment. Despite LLM's success in multiple areas, its ability to play abstract games, such as chess, is underexplored. Chess-playing requires the language models to output legal and reasonable moves from textual inputs. Here, we propose the Large language model ChessLLM to play full chess games. We transform the game into a textual format with the best move represented in the Forsyth-Edwards Notation. We show that by simply supervised fine-tuning, our model has achieved a professional-level Elo rating of 1788 in matches against the standard Elo-rated Stockfish when permitted to sample 10 times. We further show that data quality is important. Long-round data supervision enjoys a 350 Elo rating improvement over short-round data.

Figures

Figures reproduced from arXiv: 2501.17186 by the authors.

**Figure 1.** Left: pass@1 increases with the number of tokens. After introducing long-round data, pass@1 further increases. Right: The Elo Rating of ChessLLM with the number of training tokens. Skill level indicates the level of Stockfish. 2 Related work 2.1 Large Language Model The emergence of Large language models (LLMs) GPT-4 (Achiam et al., 2023), stands as a noteworthy testament to the significant advancements in natural … view at source ↗

**Figure 2.** One example of training data. 5 Evaluation Methods Chess requires a dynamic evaluation method beyond a fixed set typical of NLP tasks. We propose 3 http://chessdb.sourceforge.net supplementing the evaluation set with actual games to better assess the model’s capabilities. 5.1 Actual Games Playing against Stockfish, a top chess engine, offers a strategic challenge. Stockfish uses advanced algorithms to determine op… view at source ↗

**Figure 3.** Left: Best Move Accuracy of ChessLLM training with short round data. The accuracy of the best move increases with the number of training tokens. Right: Legal Move Accuracy of ChessLLM training with short round data. The accuracy of the legal move increases with the number of training tokens [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Left: Correlation between ChessLLM’s best move accuracy and its Elo rating. Right: Correlation between ChessLLM’s legal move accuracy and its Elo rating. 5.2 Evaluation Set While games against Stockfish provide a robust performance assessment, their length introduces s…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 9 canonical work pages

[1]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

arXiv 2023
[2]

Yuvanesh Anand, Zach Nussbaum, Brandon Duderstadt, Benjamin Schmidt, and Andriy Mulyar. 2023. Gpt4all: Training an assistant-style chatbot with large scale data distillation from gpt-3.5-turbo. https://github.com/nomic-ai/gpt4all

2023
[3]

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966

arXiv 2023
[4]

Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. 2023. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397--2430. PMLR

2023
[5]

Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, et al. 2022. Gpt-neox-20b: An open-source autoregressive language model. arXiv preprint arXiv:2204.06745

arXiv 2022
[6]

Murray Campbell, A Joseph Hoane Jr, and Feng-hsiung Hsu. 2002. Deep blue. Artificial intelligence, 134(1-2):57--83

work page 2002
[7]

Kedi Chen, Qin Chen, Jie Zhou, Yishen He, and Liang He. 2024. https://arxiv.org/abs/2403.00896 Diahalu: A dialogue-level hallucination evaluation benchmark for large language models . Preprint, arXiv:2403.00896

work page Pith review arXiv 2024
[8]

Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90\ See https://vicuna. lmsys. org (accessed 14 April 2023)

2023

Show all 34 references

[9]

Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30

2017
[10]

Together Computer. 2023. https://github.com/togethercomputer/RedPajama-Data Redpajama: an open dataset for training large language models

2023
[11]

David, Nathan S

Omid E. David, Nathan S. Netanyahu, and Lior Wolf. 2016. https://doi.org/10.1007/978-3-319-44781-0_11 DeepChess: End-to-End Deep Neural Network for Automatic Learning in Chess , page 88–96. Springer International Publishing

2016 doi
[12]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://arxiv.org/abs/1810.04805 Bert: Pre-training of deep bidirectional transformers for language understanding . Preprint, arXiv:1810.04805

2019 arXiv
[13]

Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2021. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360

2021 arXiv
[14]

Xidong Feng, Yicheng Luo, Ziyan Wang, Hongrui Tang, Mengyue Yang, Kun Shao, David Mguni, Yali Du, and Jun Wang. 2023. Chessgpt: Bridging policy learning and language modeling. arXiv preprint arXiv:2306.09200

2023 arXiv
[15]

Xinyang Geng and Hao Liu. 2023. https://github.com/openlm-research/open_llama Openllama: An open reproduction of llama

2023
[16]

Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR)

2021
[17]

Jeremy Howard and Sebastian Ruder. 2018. https://arxiv.org/abs/1801.06146 Universal language model fine-tuning for text classification . Preprint, arXiv:1801.06146

2018 arXiv
[18]

o pf, Yannic Kilcher, Dimitri von R \

Andreas K \"o pf, Yannic Kilcher, Dimitri von R \"u tte, Sotiris Anagnostidis, Zhi-Rui Tam, Keith Stevens, Abdullah Barhoum, Nguyen Minh Duc, Oliver Stanley, Rich \'a rd Nagyfi, et al. 2023. Openassistant conversations--democratizing large language model alignment. arXiv prepr...

2023 arXiv
[19]

Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, Minguk Kang, Taesung Park, Jure Leskovec, Jun-Yan Zhu, Li Fei-Fei, Jiajun Wu, Stefano Ermon, and Percy Liang. 2023. https...

2023 arXiv
[20]

Haolong Li, Yu Ma, Yinqi Zhang, Chen Ye, and Jie Chen. 2024. https://doi.org/10.18653/v1/2024.findings-acl.55 Exploring mathematical extrapolation of large language models with synthetic data . In Findings of the Association for Computational Linguistics: ACL 2024, pages 936--...

2024 doi
[21]

Thomas McGrath, Andrei Kapishnikov, Nenad Toma s ev, Adam Pearce, Martin Wattenberg, Demis Hassabis, Been Kim, Ulrich Paquet, and Vladimir Kramnik. 2022. Acquisition of chess knowledge in alphazero. Proceedings of the National Academy of Sciences, 119(47):e2206625119

2022
[22]

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 3...

2022
[23]

Javier Rando and Florian Tram \`e r. 2023. Universal jailbreak backdoors from poisoned human feedback. arXiv preprint arXiv:2311.14455

2023 arXiv
[24]

David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. 2018. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science, ...

2018
[25]

Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023 a . Stanford alpaca: An instruction-following llama model

2023
[26]

Hashimoto

Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023 b . Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca

2023
[27]

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023 a . Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

2023 arXiv
[28]

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, W...

2023 arXiv
[29]

Alan M Turing. 1953. Digital computers applied to games. Faster than thought

1953
[30]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

2017
[31]

Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. 2022. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682

2022 arXiv
[32]

P Xing, Hao Zhang, Joseph E

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. https://arxiv.org/abs/2306.05685 Judging llm-as-a-judge with mt-bench and chatbot arena . P...

2023 arXiv
[33]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[34]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

arXiv 2023

[2] [2]

Yuvanesh Anand, Zach Nussbaum, Brandon Duderstadt, Benjamin Schmidt, and Andriy Mulyar. 2023. Gpt4all: Training an assistant-style chatbot with large scale data distillation from gpt-3.5-turbo. https://github.com/nomic-ai/gpt4all

2023

[3] [3]

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966

arXiv 2023

[4] [4]

Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. 2023. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397--2430. PMLR

2023

[5] [5]

Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, et al. 2022. Gpt-neox-20b: An open-source autoregressive language model. arXiv preprint arXiv:2204.06745

arXiv 2022

[6] [6]

Murray Campbell, A Joseph Hoane Jr, and Feng-hsiung Hsu. 2002. Deep blue. Artificial intelligence, 134(1-2):57--83

work page 2002

[7] [7]

Kedi Chen, Qin Chen, Jie Zhou, Yishen He, and Liang He. 2024. https://arxiv.org/abs/2403.00896 Diahalu: A dialogue-level hallucination evaluation benchmark for large language models . Preprint, arXiv:2403.00896

work page Pith review arXiv 2024

[8] [8]

Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90\ See https://vicuna. lmsys. org (accessed 14 April 2023)

2023