Pith. sign in

REVIEW 3 major objections 4 minor 28 references

Otter: A Time-Aware, History-Conditioned Human Chess AI

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Otter predicts human chess moves 55.23% of the time by conditioning on the last 20 moves and the clock, surpassing a larger position-only model.

desk verdict Otter's internal ablations are clean and its history/time conditioning is genuinely new, but the Maia2 comparison is uncontrolled; the paper deserves review with a shared evaluation required. read the letter →

arxiv 2608.05206 v1 pith:GSLBTIZX submitted 2026-08-05 cs.AI cs.LG

Tarun Kumar S This is my paper · ORCID
classification cs.AIcs.LG
keywords humanmovepredictionchessAIhistoryconditioningclockpressureconditionedattentionablationstudyposition-onlybaselinerapid
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 the best way to predict which move a human will actually play is to treat the game as a sequential, time-pressured process rather than as a series of independent positions. Its model, Otter, feeds the board plus the last 20 moves and the current clock state into a 15.3-million-parameter network and reports 55.23% top-1 and 90.95% top-5 accuracy on a balanced 1.1-million-position test set, with an ablation showing +5.24 points from move history and +2.38 points from clock conditioning. The paper further reports that this surpasses the strongest position-only model, which uses more parameters and more training data. A sympathetic reader should care if this holds, because predicting real human choices, not just optimal engine moves, is what coaching tools, human-like opponents, and cheat detection actually need.

What carries the argument

Two conditioning streams carry the argument: a move-history encoder that embeds the last 20 moves with a small Transformer and produces both token-level representations and a pooled summary, and a time module that turns remaining clock fraction and increment into features through a two-layer MLP. These are fused with Elo buckets into one 640-dimensional conditioning vector that is projected into the queries of every attention block, so board tokens can cross-attend to history tokens with player identity and time pressure shaping what they look for. This conditioned-query attention is the mechanism that lets a small model exploit path dependence and clock state.

What would settle it

Run the prior model and Otter on the same 1.1-million-position February 2025 test set with identical legal-move masking and Elo bucketing; if the prior model's top-1 accuracy reaches or exceeds 55.23%, the paper's headline comparison fails even though its internal ablations would remain valid.

Watch

Extended reading notes

Core claim

The core discovery claim is that the position-only Markov assumption is the main bottleneck in human chess move prediction. Adding a Transformer that encodes the last 20 moves and feeding clock pressure into every attention layer lifts top-1 accuracy from 47.61% to 55.23% on identical test positions, with every one of 11 Elo rating brackets gaining at least +7.22 points. The history contribution saturates quickly, with K=5 already capturing 84% of the total history gain, and the time module contributes uniformly even in the lowest rating bracket, where the paper argues clock management is least deliberate. On the paper's own comparison, Otter's 55.23% exceeds the prior model's reported 53.25% with 34% fewer parameters and 31% less training data.

Load-bearing premise

The load-bearing assumption is that the prior model's publicly reported accuracy, measured on its own evaluation set and training distribution, is directly comparable to Otter's accuracy on a balanced February 2025 test set selected by the authors; Section 5.2 and Table 7 explicitly say this comparison is not fully controlled for training data distribution.

Editorial extensions

If this is right

  • If the central claim holds, position-only human chess models are leaving a large, cheaply accessible signal unused, and future human move models should include move history and clock context.
  • The window-size curve implies most of the benefit, 84 percent, is already available with just the last 5 moves, so near-full accuracy can be reached with less compute at inference time.
  • The uniform gain across every rating bracket, including the lowest, says time pressure shapes move choice generally rather than only for strong players.
  • In the paper's framing, stronger players become more predictable once behavioral context is provided, which suggests skill-conditioned models gain most from history when they also know the player's level.
  • Concrete applications such as personalized coaching, human-like AI opponents, and cheat detection follow directly if human move prediction at this accuracy is reproducible.

Reading between the lines

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

  • Editorial inference: the paper trains only on rapid games, so the time-conditioning story is untested for bullet and blitz; the same normalized-clock features would plausibly carry over, but that is not in the paper.
  • Editorial inference: because the code and models are released, an independent researcher could settle the headline comparison by running the prior model on the identical 1.1-million-position test set; the paper's own note that the comparison is not fully controlled for training distribution is a warning the two numbers may not be apples to apples.
  • Editorial inference: the acknowledged under-prediction of blunders points to a separate research direction, namely using a tail-weighted or uncertainty-aware loss to predict rare high-cost errors, rather than a defect the history and time context were designed to fix.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces Otter, a 15.3M-parameter neural network for predicting human moves in chess, trained on 117M Lichess rapid games. Otter augments a position-only convolutional board encoder with two conditioning signals: a Transformer encoder over the last 20 moves, and a time-control/clock module encoding game format and remaining time. The paper reports 55.23% top-1 accuracy on a balanced February 2025 test set of 1.1M positions, claims to surpass Maia2 (53.25%), and presents ablations isolating +5.24pp from history and +2.38pp from time, with gains uniform across Elo brackets and game phases. Code, models, and training logs are released.

Significance. If the internal ablations are reliable, the paper provides a clear demonstration that move-history and time-pressure conditioning improve human move prediction over a position-only baseline of the same architecture, with a relatively compact model. The release of code, models, and training logs is a strength. However, the headline SOTA claim rests on a comparison with Maia2 using published aggregate numbers from a different evaluation setup; the paper itself acknowledges this in Section 5.2. The central contribution is therefore the ablation study and the architecture, not the cross-dataset superiority claim as currently stated.

major comments (3)
  1. [§5.2, Table 7] The claim that Otter surpasses Maia2 is not supported by the presented evidence because the two accuracy numbers are measured on different evaluation sets with different rating-bracket composition and time periods. Maia2's 53.25% is a macro-average over three Skilled/Advanced/Master cohorts from Tang et al.; Otter's 55.23% is a macro-average over 11 equally weighted Elo buckets from February 2025 Lichess games selected by the authors. Section 5.2 itself acknowledges the comparison is 'not fully controlled for training data distribution,' yet the abstract and Section 1 present the SOTA claim as a headline result. Since the paper's own Base model scores 47.61%, far below Maia2's 53.25%, the attribution of the entire margin to history/time conditioning requires commensurability that is not established. The authors should either run Maia2 (or a Maia2 reimplementation) on the same February 2025 test set, or at minimum report Otter's accuracy on Maia2's evaluation cohorts and test distribution, and adjust the claims accordingly.
  2. [§5.6] The history-window sensitivity analysis does not test the design choice of K=20. Truncating the context of a model trained with K=20 (by zeroing the oldest tokens) at inference time measures the model's robustness to missing context, not the effect of training with a shorter window. A model trained with K=5 might behave differently, for example by allocating capacity differently. The conclusion that 'K=5 already retains 84% of the total history advantage' is therefore not a valid justification for the chosen architecture. Please train separate models with K=5, 10, and 20, or rephrase the result as an inference-time truncation analysis.
  3. [§5.1–§5.2] No uncertainty quantification is provided for any of the reported accuracies. At n=1.1M, the sampling error on 55.23% is about ±0.05 percentage points (95% CI), so the internal ablation gaps of 5–8pp are clearly significant under a binomial model. However, for the cross-dataset Maia2 comparison, the uncertainty in Maia2's published number is unknown and the comparison is confounded by evaluation-set differences. The authors should report confidence intervals (e.g., bootstrap) for the main numbers, and ideally train each ablation variant with multiple seeds to rule out seed-dependent conclusions.
minor comments (4)
  1. [Throughout] The paper uses 'Maia2' and 'Maia 2' inconsistently; pick one form and use it consistently across the abstract, body, and tables.
  2. [§4.3] The text states that the value loss 'does not decrease significantly' yet the value head is retained as a regularizer; a sentence on whether the regularization actually improves policy accuracy (e.g., an ablation without the value head) would strengthen this claim.
  3. [Figure 4] The x-axis labels in Figure 4 are cramped (e.g., '1100 1199' appears compressed); consider rotating the labels or adjusting the layout for legibility.
  4. [§5.2] The sentence 'confirming that architectural differences alone do not explain the improvement' is odd because the Base model is 5.64pp below Maia2, which suggests the base is much weaker; the intended meaning is that the full model's gain comes from the added conditioning, not from the architecture. Please rephrase for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported accuracies are measured on held-out positions, ablations share the same evaluation, and the Maia 2 comparison uses an externally published benchmark.

full rationale

The paper's central derivation chain is empirical rather than circular. Otter is trained by supervised learning on human moves, and its top-1 and top-5 accuracies are measured on a held-out February 2025 test set that is disjoint from the 2024 training period. The ablations compare three variants (base, history-only, full) on identical test positions, so the reported +5.24 pp from history and +2.38 pp from time are measured effects of adding those input signals, not quantities reconstructed from fitted constants or redefinitions of the inputs. The Maia 2 comparison cites the publicly reported 53.25% figure from Tang et al., an external paper, and the paper itself explicitly notes in Section 5.2 that the comparison is not fully controlled for training data distribution; that is an evaluation-validity caveat, not a circularity. No self-citations are load-bearing, no uniqueness theorem is imported, and no known result is merely renamed as a new coordinate system. The paper does not recycle any fitted parameter as a prediction, and no equation in the paper equates an output to an input by construction. The history and time features are conditional inputs to the model, and the accuracy gain is the empirical outcome of training, so the derivation chain is self-contained with respect to circularity concerns.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central empirical claims rest on domain assumptions rather than new mathematics. The most important are the sufficiency of a 20-move window, the comparability of the external Maia2 benchmark, and the assumption that clock fractions normalized by base time capture time pressure. No invented entities or exotic axioms appear. Hyperparameters such as the history window, loss weights, and bracket discretization are chosen by hand, although the history window choice is backed by a sensitivity analysis.

free parameters (4)
  • History window K = 20
    Default window length in Section 3.3; sensitivity analysis shows K=5 retains 84% of the history benefit and K=10 to K=20 adds only +0.28pp, so the exact value is not highly load-bearing.
  • Loss weights (policy/value/aux) = 1.0 / 0.25 / 0.5
    Hand-chosen weighting in Equation 4; the value loss is explicitly reported as flat and noisy, so the 0.25 weight has little measured effect.
  • Elo bucket boundaries = 11 buckets: <1100, 100-point bins, >=2000
    Discretization of skill used in the conditioning vector; standard for the Maia line but arbitrary.
  • Time control buckets = 5 buckets: bullet, blitz, rapid, classical, other
    Categorical encoding of game format from Table 1; the training set contains only rapid games, so most categories are never observed and the embedding is largely degenerate.
assumptions (6)
  • domain assumption The last 20 moves, padded and canonicalized, capture sufficient behavioral context; longer histories would not materially change predictions.
    Used for the history encoder design and K=20 default in Section 3.3; sensitivity analysis tests K=5 and K=10 but not longer windows.
  • domain assumption Maia2's published aggregate accuracy is directly comparable to Otter's balanced February 2025 test set despite different evaluation protocols.
    Underpins the SOTA claim in Section 5.2; the paper flags the comparison as not fully controlled.
  • domain assumption The clock features f1 and f2, normalized by base time, capture the relevant time-pressure signal across formats.
    Equation 1 defines the normalization; assumes relative clock fraction and increment fraction are sufficient pressure signals.
  • domain assumption The time-control bucket provides useful game-format conditioning even though the training set contains only rapid games.
    Section 4.1 restricts training to rapid formats; Table 1 defines five time-control categories, so most categories are never observed in training.
  • domain assumption Supervised cross-entropy on human moves is the correct objective for human move prediction.
    Standard Maia-style objective; the paper uses it without derivation beyond prior literature.
  • standard math Standard transformer and convolution operations behave as documented in public frameworks.
    Background assumption for PyTorch and transformer layers; no new mathematics is proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Otter: A Time-Aware, History-Conditioned Human Chess AI." pith.science (2026). https://pith.science/paper/GSLBTIZX

@misc{pith2026260805206,
  author       = {Pith},
  title        = {Pith review of: Otter: A Time-Aware, History-Conditioned Human Chess AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GSLBTIZX}},
  note         = {Machine review of arXiv:2608.05206}
}
read the original abstract

Otter is a 15.3M-parameter human chess AI that predicts human move selection by modeling play as a time-aware, sequential process rather than treating each position in isolation. It combines two conditioning signals: (1) a move history encoder that conditions predictions on the last 20 moves, capturing opening preferences, positional drift, and intra-game behavioral tendencies; and (2) a time control module that modulates predictions based on clock pressure. Otter is trained on 6.1 billion positions from 117 million Lichess rapid games over 30 days on a single T4 GPU. Otter achieves 55.23% top-1 and 90.95% top-5 move-prediction accuracy, surpassing the prior state-of-the-art human chess model, Maia 2, with far fewer parameters and less training data. Across 11 Elo brackets (<1100 to >=2000), accuracy peaks at 57.38% in the 1900-1999 bracket. These results show that modeling chess as a time-aware, sequential activity yields more human-accurate move prediction than position-only approaches, using a smaller model. Code, trained models, and complete training logs are publicly released.

Figures

Figures reproduced from arXiv: 2608.05206 by the authors.

Figure 1
Figure 1. Overview of the Otter design. Our complete architecture encompasses four main stages: dual-stream [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Conditioned attention mechanism. The 640-dimensional conditioning vector is projected via a learned linear layer WC and added to the board token queries to produce conditioned queries Q∗ = Q+cond· WC . The conditioned queries attend to history tokens via cross-attention, followed by a feedforward network, with residual connections and pre-layer normalization throughout. The same mechanism is reused in the four subse… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Per-bracket accuracy across ablation vari￾ants. Top-1 accuracy (top) for all three variants and top-5 accuracy (bottom) for Base and Full model across all 11 Elo brackets. Peak top-1 of 57.38% and top-5 of 92.85% are both achieved in the 1900–1999 bracket [PITH_FULL_I…
Figure 5
Figure 5. Figure 5: Move prediction accuracy by game phase. Comparison of the position-only Base model and the Full model across openings (ply 0–29), middlegames (ply 30–79), and endgames (ply 80+). Absolute percent￾age point improvements are annotated above each pair. The sample size n r…
Figure 6
Figure 6. Figure 6: History window size sensitivity. Top-1 accu￾racy as a function of the history window size K. K = 0 represents the no-history Base model baseline (47.61%, plotted as a dashed line). Performance rises steeply at K = 5 and flattens out towards the chosen configuration of …
Figure 7
Figure 7. Figure 7: Training dynamics and convergence curves across 3.0M optimization steps. Panel 1 plots the policy, auxiliary, and value training losses (solid lines) alongside the validation total loss (dashed line). Panel 2 compares the training top-1 move prediction accuracy with va…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [1]

    Aligning superhuman ai with human behavior: Chess as a model system

    Reid McIlroy-Young, Siddhartha Sen, Jon Kleinberg, and Ashton Anderson. Aligning superhuman ai with human behavior: Chess as a model system. InPro- ceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), KDD ’20, pages 1677–1687. ACM, August 2020

    work page 2020

  2. [2]

    Maia-2: A unified model for human-ai alignment in chess

    Zhenwei Tang, Difan Jiao, Reid McIlroy-Young, Jon Kleinberg, Siddhartha Sen, and Ashton Anderson. Maia-2: A unified model for human-ai alignment in chess. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  3. [3]

    Stockfish: A strong open source chess engine

    The Stockfish developers. Stockfish: A strong open source chess engine. https://stockfishchess. org, 2024

    work page 2024

  4. [4]

    Leela chess zero

    The LCZero Authors. Leela chess zero. https: //lczero.org/, 2024

    work page 2024

  5. [5]

    Michael L. Littman. Markov games as a frame- work for multi-agent reinforcement learning. In William W. Cohen and Haym Hirsh, editors,Ma- chine Learning Proceedings 1994, pages 157–163. Morgan Kaufmann, San Francisco (CA), 1994

    work page 1994

  6. [6]

    Detecting indi- vidual decision-making style: Exploring behavioral stylometry in chess

    Reid McIlroy-Young, Russell Wang, Siddhartha Sen, Jon Kleinberg, and Ashton Anderson. Detecting indi- vidual decision-making style: Exploring behavioral stylometry in chess. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P. S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pages 24482–24497. Curran As- so...

    work page 2021

  7. [7]

    Speed, quality, and the optimal timing of com- plex decisions: Field evidence.arXiv preprint arXiv:2201.10808, 2022

    Uwe Sunde, Dainis Zegners, and Anthony Strittmat- ter. Speed, quality, and the optimal timing of com- plex decisions: Field evidence.arXiv preprint arXiv:2201.10808, 2022

    arXiv 2022

  8. [8]

    Johannes Carow and Niklas M. Witzig. Time pressure and strategic risk-taking in professional chess.Journal of Economic Behavior & Organi- zation, 238:107218, 2025

    work page 2025

Show all 28 references
  1. [9]

    A general reinforcement learn- ing algorithm that masters chess, shogi, and go through self-play.Science, 362(6419):1140–1144, 2018

    David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, Timothy Lillicrap, Karen Simonyan, and Demis Hassabis. A general reinforcement learn- ing algorithm that masters chess...

    2018

  2. [10]

    Learn- ing models of individual behavior in chess

    Reid McIlroy-Young, Russell Wang, Siddhartha Sen, Jon Kleinberg, and Ashton Anderson. Learn- ing models of individual behavior in chess. InPro- ceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’22, pages 1253–1263. ACM, August 2022

    2022

  3. [11]

    How time pressure modulates individual differ- ences in the functional connectivity of chunk memory in chess games.Neuroscience, 552:39–46, 2024

    Chantat Leong, Yuwen Lin, Juan Zhang, and Zhen Yuan. How time pressure modulates individual differ- ences in the functional connectivity of chunk memory in chess games.Neuroscience, 552:39–46, 2024

    2024

  4. [12]

    Chase and Herbert A

    William G. Chase and Herbert A. Simon. Percep- tion in chess.Cognitive Psychology, 4(1):55–81, 1973

    1973

  5. [13]

    Learning to play the chess variant crazyhouse above world cham- pion level with deep neural networks and human data

    Johannes Czech, Moritz Willig, Alena Beyer, Kris- tian Kersting, and Johannes Fürnkranz. Learning to play the chess variant crazyhouse above world cham- pion level with deep neural networks and human data. Frontiers in Artificial Intelligence, 3, April 2020

    2020

  6. [14]

    Chalmers

    Daniel Monroe and Philip A. Chalmers. Master- ing chess with a transformer model.arXiv preprint arXiv:2409.12272, 2024

    2024 arXiv

  7. [15]

    Chess as a testbed for language model state tracking

    Shubham Toshniwal, Sam Wiseman, Karen Livescu, and Kevin Gimpel. Chess as a testbed for language model state tracking. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11385–11393, 2022

    2022

  8. [16]

    Lewis, Joel Veness, and Tim Genewein

    Anian Ruoss, Grégoire Delétang, Sourabh Meda- pati, Jordi Grau-Moya, Li Kevin Wenliang, Elliot Catt, John Reid, Cannada A. Lewis, Joel Veness, and Tim Genewein. Amortized planning with large-scale transformers: A case study on chess. In A. Glober- son, L. Mackey, D. Belgrave, ...

    2024

  9. [17]

    Predicting human chess moves: An ai assisted analysis of chess games using skill-group specific n-gram language models.arXiv preprint arXiv:2512.01880, 2025

    Daren Zhong, Dingcheng Huang, and Clayton Greenberg. Predicting human chess moves: An ai assisted analysis of chess games using skill-group specific n-gram language models.arXiv preprint arXiv:2512.01880, 2025

    2025

  10. [18]

    Batch nor- malization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. Batch nor- malization: Accelerating deep network training by reducing internal covariate shift. In Francis Bach and David Blei, editors,Proceedings of the 32nd Interna- tional Conference on Machine Learning, volume 37 ofProceedings of Machin...

    2015

  11. [19]

    Deep residual learning for image recog- nition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

    2016

  12. [20]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems, 2017

    2017

  13. [21]

    Scalable diffu- sion models with transformers

    William Peebles and Saining Xie. Scalable diffu- sion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  14. [22]

    Gaus- sian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016

    Dan Hendrycks and Kevin Gimpel. Gaus- sian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016

    2016 arXiv

  15. [23]

    Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958, 2014

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhut- dinov. Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958, 2014

    1929

  16. [24]

    Lichess open database

    Lichess.org. Lichess open database. https:// database.lichess.org, 2024

    2024

  17. [25]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Confer- ence on Learning Representations, 2019

    2019

  18. [26]

    SGDR: Stochas- tic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: Stochas- tic gradient descent with warm restarts. InInter- national Conference on Learning Representations, 2017

    2017

  19. [27]

    Mixed precision training

    Paulius Micikevicius, Sharan Narang, Jonah Al- ben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed precision training. InInternational Conference on Learning Representations, 2018

    2018

  20. [28]

    Pytorch: An im- perative style, high-performance deep learning li- brary

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Te- jani, Sasank Chilamkurthy, Benoit Steiner, L...

    2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.