Pith. sign in

REVIEW 5 major objections 5 minor 38 references

GROW: Group-Relative Advantage-Weighted On-Policy Reinforcement Learning of Autoregressive-Diffusion Text-to-Speech model

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read For strong pretrained flow-matching TTS models, reweighting the standard regression objective with a signed group-relative advantage improves intelligibility and speaker similarity while cutting training cost relative to trajectory-level po

desk verdict A strong empirical paper on advantage-weighted flow-matching TTS whose theoretical framing is undercut by an unbounded objective. read the letter →

arxiv 2608.03215 v1 pith:W4KZWVU3 submitted 2026-08-04 eess.AS cs.AIcs.CL

Guanrou Yang , Tian Tan , Qian Chen , Ziyang Ma , Yakun Song , Zhikang Niu , Qi Chen , Wenming Tu
show 3 more authors
Haitao Li Shan Yang Xie Chen
This is my paper
classification eess.AScs.AIcs.CL
keywords text-to-speechflowmatchingreinforcementlearningadvantageweightingautoregressivediffusionzero-shotspeechsynthesisreward-weightedregressionWassersteinregularization
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

GROW is a recipe for post-training a strong pretrained text-to-speech model when the objective is to improve intelligibility and speaker similarity. Existing trajectory-level reinforcement learning methods for flow-matching TTS are heavy: they turn the deterministic sampling ODE into a stochastic SDE and evaluate per-step likelihood ratios. GROW instead reweights the ordinary flow-matching regression loss with a signed, group-normalized advantage, computed from a small group of on-policy utterances per prompt, and anchors the velocity to the frozen pretrained model with a Wasserstein-2 penalty. The paper's core claim is that this simple change outperforms trajectory-level policy gradients: average WER drops from 2.016 to 1.558, speaker similarity rises from 0.676 to 0.715, perceptual quality is preserved, and cheaper 10-step rollouts train 2.9x faster than the 32-step baseline. A sympathetic reader would care because it suggests the expensive policy-gradient machinery is not necessary for RL fine-tuning of flow-matching speech generators.

What carries the argument

The load-bearing object is the signed, group-normalized advantage weight $$ $w^{{(k)}}$=\sum_{r\in\{\mathrm{sim},\mathrm{wer}\}}\lambda_r\frac{$R_r^{{(k)}}$-\mu_r}{\sigma_r+\epsilon}, $$ applied as a per-sample multiplier on the standard flow-matching regression loss, $$ \sum_k $w^{{(k)}}$\|v_\$\theta$($x_t^{{(k)}}$,t)-v^\star\|^2+\beta_{W2}\sum_k\|v_\$\theta$-v_{\theta_{\mathrm{ref}}}\|^2. $$ The first term is the pretraining objective, reweighted; the second is a Wasserstein-2 anchor to the frozen reference velocity. The key identity is the decomposition $L_{aw}=\bar{w}\sum_k\ell^{(k)}+\sum_k\delta^{(k)}\ell^{(k)}$: a strictly positive weight such as $\exp(\tau A)$ leaves a large self-imitation term, while the

What would settle it

Take a set of real prompt groups, compute the GROW update, and compare its direction with the gradient of the KL-regularized reward objective under the current policy; if a substantial fraction of groups show non-positive alignment, the signed linear weight is not acting as a policy-improvement step. A complementary experiment: give the exponential weight a per-group temperature calibrated so that $\tau\sigma$ matches the observed advantage scale; if it then matches or beats $A/\sigma$, the linearization rather than the centering is what carries the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that for a strong pretrained autoregressive-diffusion TTS model, the effective way to run on-policy reinforcement learning is not to convert the deterministic flow ODE into a stochastic SDE and track per-step likelihoods, but to keep the standard flow-matching regression objective and reweight each sample by a signed, group-normalized advantage. It identifies a failure mode of positive exponential reward weighting: when the pretrained policy already scores high and rewards are tightly clustered, the exponential weight splits into a dominant self-imitation term and a weak reward-contrast term. GROW removes the self-imitation component by using the zero-mean linear

Load-bearing premise

The method only works as advertised if replacing the exponential reward tilt with the signed linear advantage (justified by a first-order Taylor expansion for small advantages) keeps the update heading toward the reward-optimal policy, and if one gradient update approximately realizes the weighted objective; neither is proven for the actual model.

Editorial extensions

If this is right

  • RL fine-tuning of flow-matching TTS can stay fully deterministic: no ODE-to-SDE conversion or per-step likelihood evaluation is required to move reward-relevant metrics.
  • For strong pretrained generators with concentrated rewards, any strictly positive weight (e.g., $\exp(\tau r)$ or $\exp(\tau A)$) is dominated by self-imitation; zero-mean signed advantage weights preserve within-group credit assignment.
  • Reducing rollout NFE from 32 to 10 costs little final quality while cutting wall-clock training time by a factor of 2.9, so RL post-training can be run at larger scale or with more prompts.
  • The regression loss should target the conditional velocity, not the CFG-guided velocity; training on the guided velocity nearly removes the benefit.
  • A light Wasserstein-2 anchor is load-bearing: with no anchor the signed weights drift and speaker similarity does not improve, while a strong anchor pins the policy and erases most gains.

Reading between the lines

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

  • If the self-imitation/reward-contrast decomposition is the real mechanism, the same zero-mean standardized advantage should also beat exponential weighting for other saturated-reward generative models, but the paper only tests TTS, so this is an extrapolation.
  • The linearization is a first-order approximation; a natural boundary condition is that as within-group reward variance grows, exact exponential weighting may recover relevance. An adaptive switch or a second-order correction could be tested against $A/\sigma$ on deliberately high-variance reward groups.
  • The asymmetric reward coupling—speaker-similarity rewards also reduce WER, but intelligibility rewards barely move similarity—suggests shared latent structure between timbre adherence and recognition that the paper does not investigate.
  • One practical extension is that group-relative standardization removes the need to tune $\tau$ against an unknown, drifting reward scale, which may make multi-reward tuning easier in other domains; the paper does not claim this.
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

5 major / 5 minor

Summary. The paper proposes GROW, a post-training method for the DiTAR autoregressive-diffusion TTS model. For each prompt, GROW samples a group of G on-policy utterances, scores them with intelligibility (1-WER) and speaker-similarity rewards, standardizes each reward within the group, combines them into a signed zero-mean advantage, and uses this advantage to reweight the standard flow-matching regression loss. A Wasserstein-2 velocity term anchors the model to a frozen pretrained reference. The paper reports that across LibriSpeech-PC and Seed-TTS EN/ZH, GROW reduces average WER from 2.016 to 1.558 and raises speaker similarity from 0.676 to 0.715 while preserving UTMOS, and that with 10-NFE rollouts it trains 2.9x faster than the 32-NFE DiTAR-GRPO baseline at comparable final quality. The paper positions GROW as a simpler and more efficient alternative to trajectory-level policy-gradient RL for flow-matching TTS.

Significance. If the empirical results hold, GROW is a practically valuable recipe: direct advantage-weighted flow-matching regression appears to outperform trajectory-level GRPO on a strong AR-diffusion TTS model while reducing training cost considerably. The paper includes informative ablations of the weighting function, rollout CFG, reward composition, and anchor strength, and it promises open-source code and checkpoints. However, the theoretical grounding is currently not established, and there are methodological concerns—test-set hyperparameter selection, lack of error bars, and reward/evaluation metric overlap—that materially reduce confidence in the headline numbers.

major comments (5)
  1. [Advantage-Weighted Flow Matching with Reference Anchoring, Eq. (11)] The objective is unbounded below. For a single rollout, the loss as a function of the network output a = v_theta(x_t,t) is a quadratic with curvature 2(w^(k)+beta_W2). Because w is zero-mean and standardized per group (Eq. 9), with G=8 there will typically be some k with w^(k) < -beta_W2 = -0.025; for example, a z-score around -1 per stream gives w ≈ -2. That term is then concave and can be driven to -infinity. Summing over k does not remove this, since individual negative blocks remain even though the sum of curvatures is G*beta_W2 > 0. The statement that the reference penalty 'mitigates this tendency' is not formally correct: a small positive beta only shifts the stationary point and does not convexify the per-sample term. This breaks the RWR/tilted-optimum derivation (Eqs. 3-7), because the algorithm is not minimizing a well-defined objective with a finite optimum. Please either (a) p
  2. [Experiments, 'Learning rate and W2-anchor strength', Table 4] Hyperparameters beta_W2 and eta, and also the rollout CFG gamma in Table 3, are selected by evaluating directly on the three test sets used for the headline claims. Table 4 reports the sweep, and the default beta_W2=0.025, eta=2e-6 is the configuration that performs best on those test sets. This is selection on the test data; the reported average improvements (WER -0.458, SIM +0.039) are therefore optimistic and are not unbiased estimates of generalization. Use a held-out validation split for model selection and report test-set numbers only for a fixed configuration, or apply an appropriate multiple-testing correction.
  3. [Experimental Results, Tables 1-5] No variance estimates are provided. All tables report single runs, and differences as small as 0.019 WER or 0.001 SIM are used to compare configurations (e.g., Table 1 GROW NFE=10 vs. NFE=32; Table 3 gamma=1.5 vs. gamma=0). Given the stochasticity of rollouts, training, and the Whisper/WavLM metrics, these differences are likely within plausible noise. Please report mean +/- std or confidence intervals over at least three training seeds for the main comparison (Table 1) and for the key ablations.
  4. [From Exponential to Linear Group-Relative Advantage, Eq. (6)] The theoretical justification relies on (i) the first-order Taylor expansion exp(tau A) = 1 + tau A + O((tau A)^2) being accurate for within-group advantages, and (ii) the RWR assumption that the model exactly fits the weighted objective each round. Neither is validated. The final weight in Eq. (9) divides by the group standard deviation, so the weight magnitude is scale-free even when tau A is not small; no empirical distribution of tau A during training is reported. Moreover, a single AdamW step cannot achieve the perfect-fit RWR optimum. Without these conditions, the link from exponential tilting (Eq. 4) to the signed linear weight (Eq. 8) is not established. Please supply empirical diagnostics (e.g., histograms of advantages/weights, or a comparison of update magnitudes against the Taylor remainder) or soften the theoretical claims.
  5. [Evaluation Metrics and Algorithm 1] The evaluation metrics are computed with Whisper-large-v3 and WavLM-large speaker embeddings, which are the same models that provide the RL rewards. Gains in WER and SIM are therefore partially optimized for these specific evaluators; they do not by themselves establish improvement in general intelligibility or speaker fidelity. Add at least one held-out ASR and one held-out speaker encoder (e.g., a different ASR model and a different speaker-embedding model) to the evaluation and report those numbers.
minor comments (5)
  1. [Advantage-Weighted Flow Matching with Reference Anchoring] The claim that the squared velocity discrepancy in Eq. (11) is a 'tractable upper bound' on Wasserstein-2 distance is unsupported. For arbitrary velocity fields, the pointwise squared L2 difference at the same noised input is not generally an upper bound on W2 between the generated distributions. Please add a proof or weaken the claim to a heuristic regularizer.
  2. [Algorithm 1, line 5] The pseudocode defines w^(k) using lambda_wer and lambda_sim but does not restate their values or the zero-variance redraw rule; both are only in the text. This is easy to fix but improves self-containedness.
  3. [Experimental Setup] The text says 'Both on-policy rollout and final evaluation use the same sampling procedure' but Table 3 varies rollout CFG gamma. Please clarify that the default setting is fixed and Table 3 is an ablation.
  4. [Table 1] The header 'NFESpeed-up' is missing a space; also, the average block could state the number of test sets used for averaging. Minor formatting issue.
  5. [Abstract and Introduction] The abstract says 'evaluated on LibriSpeech' but the experiments use LibriSpeech-PC test-clean; please be precise in the abstract or introduce the subset name consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GROW's derivation is self-contained and its reported improvements are empirical outcomes, not inputs renamed as predictions.

full rationale

The paper's derivation chain does not reduce to its inputs by construction. The signed group-relative advantage weight w = A/σ is explicitly constructed from the group-standardized rewards (Eq. 9) and justified as the leading-order Taylor term of the exponential advantage weight (Eq. 6), which is a mathematical approximation rather than a circular equivalence. The training objective (Eq. 11) is a weighted flow-matching regression plus a Wasserstein-2 reference penalty; it is not a fitted parameter disguised as a prediction. The reported WER and SIM improvements are measured on held-out test sets, not on the training rollouts, and no fitted value is relabeled as a forecast. The paper discloses that the evaluation ASR and speaker models are the same frozen models used to compute the RL rewards ('These are the same recognition and speaker models used to compute the RL reward'), which is a reward/evaluation overlap worth noting as a limitation, but it does not make the empirical claim definitionally circular: optimizing a reward and then reporting that same metric improved is expected behavior, not a hidden reduction. Self-citations to prior work by overlapping authors (e.g., F5-TTS, Semantic-VAE) are used as architectural or representational building blocks, not as load-bearing justifications for the central advantage-weighting claim. The skeptics' concern about Eq. (11) being unbounded below for negative weights is a formal correctness/stability issue, not a circularity issue. Therefore no circular step is present; the appropriate score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central result rests on several tuned hyperparameters and on two unproven theoretical bridges: the linearized advantage weight (via Taylor expansion plus the RWR perfect-fit assumption) and the W2 upper-bound claim. The method introduces no new physical entities.

free parameters (7)
  • β_W2 (W2 anchor weight) = 0.025
    Selected by sweeping β_W2 ∈ {0, 0.025, 0.05, 0.1, 1} on the three test sets (Table 4); the central reported gains depend on this value.
  • λ_wer, λ_sim (reward fusion coefficients) = 1, 1
    Set equal by hand; not swept, but the multi-objective combination's balance is a design choice affecting the reported trade-off.
  • learning rate η = 2e-6
    Tuned; Table 4 shows η=2e-6 gives the best results at β_W2=0.025; the reported gains depend on this value.
  • rollout CFG scale γ = 1.5
    Swept (Table 3): γ=1.5 was chosen because it maximizes speaker similarity; WER is slightly better at γ=0.
  • rollout NFE (number of function evaluations) = 10
    Chosen as default (Table 1) to trade a 2.9x speedup against a small WER loss; the headline numbers use NFE=10 rollouts.
  • group size G = 8
    Fixed by hand; no sweep reported. The advantage statistics are computed over the group, so results depend on G.
  • zero-variance group redraw rule = σ_sim+σ_wer = 0 groups redrawn
    A data-dependent exclusion rule adopted from DAPO; removes groups that carry no gradient, a post-hoc selection step.
assumptions (5)
  • standard math Exponential-tilt optimum of the KL-regularized reward objective (Eq. 4)
    Classical result from Peters et al. 2010; used to motivate reward-weighted regression.
  • domain assumption Reward-weighted regression exactly realizes the tilted distribution under the perfect-fit assumption each round
    Invoked from Fan et al. 2025 and Peters & Schaal 2007; the paper does not show that a single gradient step of DiTAR satisfies the perfect-fit assumption.
  • ad hoc to paper First-order Taylor expansion of exp(τA) is valid for concentrated advantages
    Eq. (6) in 'From Exponential to Linear Group-Relative Advantage'; assumes τA is small, which is asserted but not measured.
  • ad hoc to paper Squared velocity discrepancy is a tractable upper bound on Wasserstein-2 distance
    Stated after Eq. (11) in 'Advantage-Weighted Flow Matching with Reference Anchoring'; no proof is given, and evaluating both velocities at the same noised input is not an obvious valid coupling for such a bound.
  • domain assumption Flow-matching regression objective (Eq. 2) is the correct surrogate for the tilted distribution
    Standard flow-matching theory; assumed from the pretraining literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GROW: Group-Relative Advantage-Weighted On-Policy Reinforcement Learning of Autoregressive-Diffusion Text-to-Speech model." pith.science (2026). https://pith.science/paper/W4KZWVU3

@misc{pith2026260803215,
  author       = {Pith},
  title        = {Pith review of: GROW: Group-Relative Advantage-Weighted On-Policy Reinforcement Learning of Autoregressive-Diffusion Text-to-Speech model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W4KZWVU3}},
  note         = {Machine review of arXiv:2608.03215}
}
read the original abstract

Reinforcement learning for flow-matching text-to-speech is complicated by deterministic ODE sampling: trajectory-level policy-gradient methods typically convert the ODE into an SDE and track per-step likelihood ratios, introducing stochastic perturbations and substantial overhead. We propose GROW, a group-relative advantage-weighted on-policy RL method that acts directly on the standard flow-matching objective. For each prompt, GROW samples a group of on-policy utterances, separately standardizes intelligibility and speaker-similarity rewards within the group, and combines them to reweight flow-matching regression. A Wasserstein-2 velocity penalty anchors the updated model to a frozen pretrained reference. A group-mean reward baseline is introduced to convert reward weighting into advantage weighting. For strong pretrained TTS models with concentrated rewards, positive exponential weighting is dominated by reward-agnostic self-imitation, whereas a zero-mean signed advantage preserves effective within-group credit assignment. Instantiated on DiTAR and evaluated on LibriSpeech and Seed-TTS EN/ZH, GROW reduces average WER from 2.016 to 1.558 and raises speaker similarity from 0.676 to 0.715 while keeping UTMOS. With 10-NFE training rollouts and 32-NFE evaluation, GROW retains comparable performance while training 2.9x faster than 32-NFE DiTAR-GRPO. We will open-source complete GROW codes, faithful DiTAR reproduction, and all model checkpoints.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

  1. [1]

    Proc.ICLR , year=

    Online reward-weighted fine-tuning of flow matching with wasserstein regularization , author=. Proc.ICLR , year=

    work page

  2. [2]

    arXiv preprint , year=

    DiTAR: Diffusion transformer autoregressive modeling for speech generation , author=. arXiv preprint , year=

    work page

  3. [3]

    Liu, Jie and Liu, Gongye and Liang, Jiajun and Li, Yangguang and Liu, Jiaheng and Wang, Xintao and Wan, Pengfei and Zhang, Di and Ouyang, Wanli , journal=. Flow-

  4. [4]

    arXiv preprint , year=

    FlowTTS-GRPO: Online Reinforcement Learning with Multi-Objective Reward Optimization for Flow-Matching Based Text-to-Speech , author=. arXiv preprint , year=

    work page

  5. [5]

    Proc.ICASSP , year=

    Direct Preference Optimization For Speech Autoregressive Diffusion Models , author=. Proc.ICASSP , year=

    work page

  6. [6]

    VGPO: Fine-Tuning Speech Autoregressive Diffusion Models with Value Guided Policy Optimization , author=

    work page

  7. [7]

    arXiv preprint , year=

    Resonate: Reinforcing Text-to-Audio Generation via Online Feedback from Large Audio Language Models , author=. arXiv preprint , year=

    work page

  8. [8]

    Proc.NeurIPS , year=

    Dapo: An open-source llm reinforcement learning system at scale , author=. Proc.NeurIPS , year=

    work page

Show all 38 references
  1. [9]

    arXiv preprint , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint , year=

  2. [10]

    arXiv preprint , year=

    Advantage-weighted regression: Simple and scalable off-policy reinforcement learning , author=. arXiv preprint , year=

  3. [11]

    arXiv preprint , year=

    Advantage weighted matching: Aligning rl with pretraining in diffusion models , author=. arXiv preprint , year=

  4. [12]

    arXiv preprint , year=

    FlowAWR: Online Adaptive Flow Reinforcement via Advantage-Weighted Rectification , author=. arXiv preprint , year=

  5. [13]

    arXiv preprint , year=

    Diffusionnft: Online diffusion reinforcement with forward process , author=. arXiv preprint , year=

  6. [14]

    arXiv preprint , year=

    FlowSE-GRPO: Training Flow Matching Speech Enhancement via Online Reinforcement Learning , author=. arXiv preprint , year=

  7. [15]

    arXiv preprint , year=

    Differentiable Reward Optimization for LLM based TTS system , author=. arXiv preprint , year=

  8. [16]

    arXiv preprint , year=

    F5r-tts: Improving flow-matching based text-to-speech with group relative policy optimization , author=. arXiv preprint , year=

  9. [17]

    arXiv preprint , year=

    Coefficients-Preserving Sampling for Reinforcement Learning with Flow Matching , author=. arXiv preprint , year=

  10. [18]

    arXiv preprint , year=

    MixGRPO: Unlocking flow-based grpo efficiency with mixed ode-sde , author=. arXiv preprint , year=

  11. [19]

    Diffusion

    Zheng, Kaiwen and Chen, Huayu and Ye, Haotian and Wang, Haoxiang and Zhang, Qinsheng and Jiang, Kai and Su, Hang and Ermon, Stefano and Zhu, Jun and Liu, Ming-Yu , journal=. Diffusion

  12. [20]

    arXiv preprint , year=

    SwanVoice: Expressive Long-Form Zero-Shot Speech Synthesis for Both Monologue and Dialogue , author=. arXiv preprint , year=

  13. [21]

    Anastassiou, Philip and Chen, Jiawei and Chen, Jitong and Chen, Yuanzhe and Chen, Zhuo and Chen, Ziyi and Cong, Jian and Deng, Lelai and Ding, Chuang and Gao, Lu and others , journal=. Seed-

  14. [22]

    arXiv preprint , year=

    Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training , author=. arXiv preprint , year=

  15. [23]

    Cui, Jiayan and Yang, Zhihan and Li, Naihan and Tian, Jiankun and Ma, Xingyu and Zhang, Yi and Chen, Guangyu and Yang, Runxuan and Cheng, Yuqing and Zhou, Yizhi and others , journal=

  16. [24]

    arXiv preprint , year=

    Explore the Reinforcement Learning for the LLM based ASR and TTS system , author=. arXiv preprint , year=

  17. [25]

    tts Technical Report , author=

    dots. tts Technical Report , author=. arXiv preprint , year=

  18. [26]

    Proc.AAAI , year=

    Relative entropy policy search , author=. Proc.AAAI , year=

  19. [27]

    Proc.ICML , year=

    Reinforcement learning by reward-weighted regression for operational space control , author=. Proc.ICML , year=

  20. [28]

    Machine learning , volume=

    Simple statistical gradient-following algorithms for connectionist reinforcement learning , author=. Machine learning , volume=. 1992 , publisher=

    1992

  21. [29]

    Policy gradient methods for reinforcement learning with function approximation , author=

  22. [30]

    Proc.ICML , year=

    Robust speech recognition via large-scale weak supervision , author=. Proc.ICML , year=

  23. [31]

    arXiv preprint , year=

    Semantic-VAE: Semantic-Alignment Latent Representation for Better Speech Synthesis , author=. arXiv preprint , year=

  24. [32]

    arXiv preprint , year=

    Qwen3 technical report , author=. arXiv preprint , year=

  25. [33]

    arXiv preprint , year=

    Utmos: Utokyo-sarulab system for voicemos challenge 2022 , author=. arXiv preprint , year=

    2022

  26. [34]

    Proc.ICASSP , year=

    Large-scale self-supervised speech representation learning for automatic speaker verification , author=. Proc.ICASSP , year=

  27. [35]

    Proc.SLT , year=

    Emilia: An extensive, multilingual, and diverse speech dataset for large-scale speech generation , author=. Proc.SLT , year=

  28. [36]

    arXiv preprint , year=

    Libritts: A corpus derived from librispeech for text-to-speech , author=. arXiv preprint , year=

  29. [37]

    Chen, Yushen and Niu, Zhikang and Ma, Ziyang and Deng, Keqi and Wang, Chunhui and JianZhao, JianZhao and Yu, Kai and Chen, Xie , booktitle=. F5-

  30. [38]

    arXiv preprint , year=

    Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition , author=. arXiv preprint , year=

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.