Pith. sign in

REVIEW 3 major objections 3 minor 21 references

Reinforcement learning on a latent-reasoning chess model improves resilience to disrupted thoughts, not reliance on their content; the gain lives in the weights.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-01 08:53 UTC pith:JHNYXAV7

load-bearing objection A genuinely new before/after-RL causal battery on latent thoughts, honest about its own soft spots, but the central robustness claim rests on a single OOD condition whose raw counts point the other way. the 3 major comments →

arxiv 2607.20952 v2 pith:JHNYXAV7 submitted 2026-07-23 cs.LG cs.CL

The Weight of Silence: A Causal Case for Weights Over the Scratchpad in Latent Chess Reasoning

classification cs.LG cs.CL
keywords latent reasoningcontinuous thoughtsreinforcement learningcausal interventionchessweight localizationconfabulationrobustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper trains a chess-playing language model through a staged latent-reasoning curriculum followed by reinforcement learning, and asks whether the latent 'thought' vectors are an actively consulted scratchpad. It reports that replacing or noising the thought content leaves legality unchanged, removing the thoughts costs only a little, and only exact-zero corruption causes collapse. The zero-collapse gap between the pre-RL and post-RL checkpoints (1% vs 9% retained legality) is the key evidence: reinforcement learning appears to add robustness to disruption, not reliance on thought content. The paper concludes that in this setting latent reasoning's main effect is to shape the weights during training, and it localizes the change to low-rank, concentrated weight adjustments. A sympathetic reader would care because this challenges the field's default assumption that silent thoughts function as an inference-time scratchpad, and it demonstrates a working RL gain in a domain where similar recipes have been reported to fail.

Core claim

The paper's central claim is that in a chess-playing model trained through a staged latent-thought curriculum followed by reinforcement learning, neither the pre-RL nor the post-RL checkpoint relies on the specific content of its latent thought vectors. A six-condition causal battery — replacing thoughts with a fixed vector, random noise, removing them, length-matched removal, and exact-zeroing — shows content-preserving substitutions leave legality unchanged, removal costs little, and only exact-zero vectors cause collapse. The zero-collapse gap between checkpoints (1% pre-RL vs 9% post-RL retained legality) is the key evidence: RL does not teach the model to consult its thoughts more effec

What carries the argument

The load-bearing instrument is a six-condition causal intervention suite applied identically to the same checkpoint before and after reinforcement learning, comparing each intervention's effect on legal-move rate. The latent thought vector — a hidden state fed back as the next input embedding instead of being decoded into a word — is the object intervened upon. The suite separates three questions: whether thought content matters (substitution, noise), whether the positions need to exist (ablation, length-matched ablation), and whether the model tolerates an unseen input (exact zero). A Jacobian-lens calibration and cross-board cosine similarity (~0.99) further support the claim that thoughts

Load-bearing premise

The central interpretation treats the exact-zero intervention as 'total signal loss' whose differential collapse pre/post RL proves reduced reliance on thought content; but exact-zero vectors are a never-seen input shape, so the gap could reflect different sensitivity to out-of-distribution inputs rather than different reliance.

What would settle it

Find a single board from the frozen 100-position harness where substituting the real thought vectors with the fixed average vector T* changes the legal move from legal to illegal (or vice versa) in the post-RL checkpoint; the paper reports zero such flips, so one clean counterexample would falsify content-invariance.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, latent-reasoning pipelines that feed hidden states back as inputs do not need to preserve thought content at inference time; the same behavior should be obtainable with arbitrary filler vectors.
  • The RL robustness gain is carried by the weights, so downstream analyses or steering tools that read thought-vector content would be reading a scaffold, not the computation.
  • The gated legality reward alone eliminates checkmate confabulation, suggesting that a single gate design can remove a hallucination failure mode without a dedicated reward term.
  • Weight-change localization predicts that patching dominant singular directions from the post-RL delta into the pre-RL checkpoint should transfer part of the legality gain; the paper identifies layer 19's down-projection as the prime candidate.
  • The monotone severity ordering across conditions (substitute/noise < ablate < lenmatch-ablate < zero) provides a template for testing content-invariance in other latent-reasoning models.

Where Pith is reading between the lines

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

  • Editorial inference: exact-zero vectors are an out-of-distribution input the model never saw in training; the 1%-vs-9% gap may reflect differential sensitivity to OOD inputs rather than a clean measure of content reliance, a concern the paper itself acknowledges in Section 3.3.
  • Editorial inference: the result is demonstrated on one 14B-parameter LoRA-tuned model in chess; the field-default scratchpad assumption could still hold in math/logic domains or in full fine-tuning at larger scale, and the paper's scope explicitly excludes those settings.
  • Editorial inference: a testable extension would run the same six-condition battery on a model trained with a reward that does not gate legality, to see whether the robustness gap is an artifact of the gated reward design rather than of RL per se.
  • Editorial inference: if the weight-localization result generalizes, one could attempt direct weight edits (patching singular directions) as a cheaper alternative to RL for adding robustness, provided the behavior transfers across checkpoints.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper trains a Qwen3-14B LoRA chess model through an SFT baseline, an explicit-reasoning RL checkpoint (Rung 1), a staged latent-thought curriculum (Stage-2), and latent GRPO (Rung-3). It reports a legality gain from 48% to 61%, a complete elimination of checkmate confabulation, and a six-condition causal intervention battery applied to the latent thought positions both before and after RL. Content-preserving substitution and noise leave performance essentially unchanged; ablation costs a little; exact-zero corruption causes collapse, with a larger proportional retention for Rung-3 than for Stage-2 (1% vs 9% retained legality). The paper interprets this as evidence that RL adds robustness to disruption rather than reliance on thought content, and further claims the RL gain is encoded in a low-rank weight change concentrated in MLP/gate projections. A J-lens analysis reports near-constant cross-board thought-vector similarity (~0.99), and a Gumbel-reparameterized RL control does not outperform the deterministic recipe.

Significance. If the central claim were fully established, the paper would be a valuable direct before/after test of the latent-thought-as-scratchpad assumption under RL, in a domain where RLVR over latent reasoning has previously been reported to fail. The paper has notable strengths: a fixed 100-position harness, a six-condition intervention design, explicit disclosure of the raw-count analysis that runs against the paper's preferred framing, a ten-times-larger replication for the post-RL side, a calibrated J-lens null on thought content, and an exact SVD of the LoRA delta. However, the headline mechanistic claim currently rests on one condition that the paper itself identifies as an out-of-distribution input, and the raw board-level counts move in the opposite direction for that condition. The result is therefore interesting but not yet load-bearing in its current form.

major comments (3)
  1. [4.4; Appendix A.7; 3.3] The central claim that RL adds robustness to disruption rather than reliance on thought content rests on the Zero condition, but Section 3.3 explicitly defines Zero as testing 'tolerance for an input shape the model has never seen, not content.' Appendix A.7 further discloses that under Zero, Stage-2 loses 47 legal boards and Rung-3 loses 49, a difference in the opposite direction, and a paired Wilcoxon on raw flips is not significant for any severe condition. The only significant before/after gap (+0.134 retention ratio) is therefore an artifact-sensitive comparison between two checkpoints with different baseline legal rates (48% vs 58%), using proportional normalization. Because exact-zero vectors are out-of-distribution for both checkpoints, the differential collapse could reflect different OOD sensitivity rather than reduced functional reliance on thoughts. The paper should either pr
  2. [Appendix A.8] The n=1000 replication is presented as confirming the pattern, but it only reruns Rung-3's own side of the battery. It cannot validate the paired Stage-2-versus-Rung-3 retention-gap comparison on which the central claim rests. Moreover, Lenmatch-Ablate at n=1000 drops Format compliance to 83.3%, so its legal-rate gap against Baseline is partly driven by output-format failures, and the appendix itself notes that forcing attention onto pad tokens introduces its own out-of-distribution cost. Thus the large-sample replication supports the post-RL checkpoint's internal ordering but does not provide the missing before/after evidence.
  3. [4.4; 4.7] The content-invariance results (Substitute and Noise) are convincingly null in both checkpoints and show that neither checkpoint relies on specific thought content. But this does not by itself establish that RL's benefit lives in the weights rather than in some other training-induced change, because the only significant before/after robustness difference is the OOD Zero condition. The Ablate and Lenmatch-Ablate comparisons trend in the predicted direction but are not individually significant at n=100, and the n=1000 replication does not include Stage-2. A causal patch/ablate of the dominant singular directions identified in Section 4.7 — suggested in Future work — would provide a much more direct test of the weight-localization claim and should be reported before the mechanistic conclusion is drawn at this strength.
minor comments (3)
  1. [Table 2 note] The baseline discrepancy (61% in Table 1 vs 58% in Table 2) is explained by NF4 dequantization-order variation, but it is unusual to rely on approximate dequantization for a headline number while using a different internal baseline for the causal battery. Clarify why the same evaluation script gives two different baseline values for the same adapter.
  2. [Appendix A.8] The Lenmatch-Ablate Format drop to 83.3% means the McNemar legal-status comparison for that condition is partially a formatting-failure comparison. It would be helpful to report the legal rate conditional on format compliance, or to use a format-robust legality measure.
  3. [4.7] The cross-recipe cosine similarity of +0.034 is small; the claim that it is 'a real, if modest, signal rather than noise' would be more convincing with a null distribution from randomized or permuted deltas, rather than only the argument that unrelated high-dimensional vectors concentrate near zero.

Circularity Check

0 steps flagged

No significant circularity: the causal battery, external benchmarks, and disclosed normalization choices keep the derivation self-contained.

full rationale

The paper's central claim—that post-RL robustness to thought disruption is a weight-level effect rather than inference-time reliance on thought content—is supported by a six-condition intervention battery run identically on pre- and post-RL checkpoints, with external Stockfish legality/checkmate scoring and comparison against external prior results (Ouro, Switch, Chess RLVR). No parameter is fitted to the target claim: the survival/retention analysis is a statistical comparison of frozen-harness outcomes, and the one normalization choice (proportional retention ratios in Appendix A.7) is explicitly disclosed alongside the raw-count alternative, so the +0.134 Zero gap is not constructed by hiding the opposite framing. The J-lens is calibrated on known text before being used to read thoughts, and the weight-delta localization in §4.7 is a direct arithmetic decomposition of the LoRA difference, not a fitted predictor. The acknowledged OOD nature of the exact-zero condition and the single-seed design are limitations of evidence, not circular steps; the paper itself flags them. No load-bearing self-citation or uniqueness import appears: citations are to external, partly disagreeing work (Switch, Ouro, Coconut-skeptic studies). The derivation does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The paper's central claim rests on several design choices: a specific RL reward, a specific causal-intervention interpretation, and a specific normalization for comparing retention. These are not derived from external theory, and any one of them, if wrong, changes the conclusion.

free parameters (3)
  • Rung 3 learning rate = 2e-6
    Chosen after 5e-6 caused legality collapse during RL; this hyperparameter tuning affects whether the RL gain appears at all (Section 3.1).
  • Reward quality scale denominator = 120 cp
    The exponential quality score exp(-cp_loss/120) uses 120 as a forgiveness scale; no principled derivation is given (Appendix A.1).
  • Curriculum stopping threshold = 20% legality
    The latent curriculum halts if legality falls below 20%, which determined that Stage-2 (48%) became the pre-RL checkpoint rather than a later stage (Section 3.1).
axioms (6)
  • standard math Standard statistical tools (McNemar, bootstrap, Bonferroni, SVD, QR) are correctly applied.
    The paper uses these techniques without proof; they are standard background math.
  • domain assumption A gated legality reward with exponential cp-loss quality is a valid RLVR signal for chess latent reasoning.
    The reward design is a modeling choice, not derived from external theory; results depend on it (Section 3.1, Appendix A.1).
  • domain assumption Causal interventions on frozen thought vectors (substitute/noise/ablate/zero) are valid probes of inference-time reliance on thought content.
    The paper assumes these manipulations reveal whether the model consults its thoughts; alternative interpretations (e.g., OOD effects) are possible (Section 3.3).
  • domain assumption The frozen 100-position harness (seed 3407) is representative and sufficient for the paper's claims.
    All headline and causal-battery numbers come from this fixed small set; no power analysis or seed averaging is provided (Section 3.2).
  • domain assumption J-lens calibrated on text can reveal the presence or absence of board-specific content in latent thought vectors.
    The null reading of thoughts is interpreted as absence of content; calibration on text does not guarantee sensitivity in the thought-vector representation (Section 4.5).
  • ad hoc to paper The retention-ratio normalization is the correct way to compare pre/post RL robustness.
    The paper's headline significant result uses condition legal rate divided by each checkpoint's own baseline; the raw board-drop counts show the opposite direction (Appendix A.7).

pith-pipeline@v1.3.0-alltime-deepseek · 27296 in / 13459 out tokens · 134222 ms · 2026-08-01T08:53:41.908607+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of The Weight of Silence: A Causal Case for Weights Over the Scratchpad in Latent Chess Reasoning." pith.science (2026). https://pith.science/paper/JHNYXAV7

@misc{pith2026260720952,
  author       = {Pith},
  title        = {Pith review of: The Weight of Silence: A Causal Case for Weights Over the Scratchpad in Latent Chess Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JHNYXAV7}},
  note         = {Machine review of arXiv:2607.20952}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Latent, or silent, reasoning lets language models carry out intermediate computation in continuous vector space instead of words, and is widely assumed to function as an internal scratchpad the model consults during inference. Whether that assumption survives reinforcement learning has not been tested directly: existing causal analyses of latent reasoning are confined to math and logic tasks, comparing reliance on thoughts within one checkpoint, never before and after RL. We train a chess-playing model through a staged latent-reasoning curriculum followed by reinforcement learning, and find legality climbs monotonically to 61% (from a 48% pre-RL baseline) while checkmate confabulation is eliminated entirely. To locate this gain, we run a six-condition causal intervention suite on the same model before and after RL: substituting or noising the thought vectors leaves performance unchanged, ablating them costs only mild degradation, and only exact-zero vectors cause collapse. This robustness gap is itself the finding: under exact-zero corruption, legality collapses to 1% pre-RL versus 9% post-RL, a gap that survives correction across the full battery. A 10x-larger replication of the post-RL checkpoint's own battery confirms this: removing the thoughts, with or without restoring sequence length, also reaches significance; substitution and noise remain indistinguishable from baseline. RL appears to add robustness to disruption, not reliance on thought content. These results push back against the field's default assumption that latent thoughts function as an actively consulted inference-time scratchpad, and instead indicate latent reasoning's principal effect here is shaping the model's parameters during training. We also demonstrate a working RL gain in chess, where multiple groups report the same latent-reasoning-plus-RL recipe failing to improve accuracy over SFT.

Figures

Figures reproduced from arXiv: 2607.20952 by Ishan S. Kshirsagar.

Figure 1
Figure 1. Figure 1: The six-condition causal intervention suite. Each row shows the same underlying sequence: a [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Legal-move percentage (left) and false checkmate count (right) across the four models in Table 1, [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two false-checkmate completions from unrelated boards, sampled from the stage-2 (pre-RL) [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Legal-move percentage across all six causal conditions, plotted from Table 2. Gray bars: Stage-2 [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Checkpoint-level training trajectories, from [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

21 extracted references · 13 linked inside Pith

  1. [1]

    Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769, 2024

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769, 2024

  2. [2]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

  3. [3]

    Scaling latent reasoning via looped language models.arXiv preprint arXiv:2510.25741, 2025

    Rui-Jie Zhu, Zixuan Wang, Kai Hua, Tianyu Zhang, Ziniu Li, Haoran Que, Boyi Wei, Zixin Wen, Fan Yin, He Xing, Lu Li, Jiajun Shi, Kaijing Ma, Shanda Li, Taylor Kergan, Andrew Smith, Xing- wei Qu, Mude Hui, Bohong Wu, Qiyang Min, Hongzhi Huang, Xun Zhou, Wei Ye, Jiaheng Liu, Jian Yang, Yunfeng Shi, Chenghua Lin, Enduo Zhao, Tianle Cai, Ge Zhang, Wenhao Huan...

  4. [4]

    Reinforcement learning for latent-space thinking in LLMs

    Enes ¨Ozeren and Matthias Aßenmacher. Reinforcement learning for latent-space thinking in LLMs. arXiv preprint arXiv:2512.11816, 2025

  5. [5]

    Detecting hallucinations in large language models using semantic entropy.Nature, 630:625–630, 2024

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630:625–630, 2024

  6. [6]

    Can large lan- guage models develop strategic reasoning? post-training insights from learning chess.arXiv preprint arXiv:2507.00726, 2025

    Dongyoon Hwang, Hojoon Lee, Jaegul Choo, Dongmin Park, and Jongho Park. Can large lan- guage models develop strategic reasoning? post-training insights from learning chess.arXiv preprint arXiv:2507.00726, 2025

  7. [7]

    Investigating gender bias in language models using causal mediation analysis.arXiv preprint arXiv:2004.12265, 2020

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. Investigating gender bias in language models using causal mediation analysis.arXiv preprint arXiv:2004.12265, 2020

  8. [8]

    Locating and editing factual associ- ations in GPT.arXiv preprint arXiv:2202.05262, 2022

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associ- ations in GPT.arXiv preprint arXiv:2202.05262, 2022

  9. [9]

    Towards best practices of activation patching in language models: Metrics and methods.arXiv preprint arXiv:2309.16042, 2024

    Fred Zhang and Neel Nanda. Towards best practices of activation patching in language models: Metrics and methods.arXiv preprint arXiv:2309.16042, 2024

  10. [10]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    Qwen Team. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

  11. [11]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

  12. [12]

    SofT-GRPO: Surpassing discrete-token llm reinforcement learning via gumbel-reparameterized soft-thinking policy optimization.arXiv preprint arXiv:2511.06411, 2025

    Zhi Zheng and Wee Sun Lee. SofT-GRPO: Surpassing discrete-token llm reinforcement learning via gumbel-reparameterized soft-thinking policy optimization.arXiv preprint arXiv:2511.06411, 2025

  13. [13]

    Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.arXiv preprint arXiv:1611.01144, 2016

  14. [14]

    Verbalizable representations form a global workspace in language models.Transformer Circuits Thread, 2026

    Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, and Jack Lindsey. Verbalizable representations form a global workspace in language models.Transformer Circuits Thread, 2026. 31

  15. [15]

    Interpreting gpt: The logit lens

    nostalgebraist. Interpreting gpt: The logit lens. LessWrong, 2020. URLhttps://www.lesswrong. com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens

  16. [16]

    Do latent tokens think? a causal and adversarial analysis of chain-of-continuous-thought.arXiv preprint arXiv:2512.21711, 2025

    Yuyi Zhang, Boyu Tang, Tianjie Ju, Sufeng Duan, and Gongshen Liu. Do latent tokens think? a causal and adversarial analysis of chain-of-continuous-thought.arXiv preprint arXiv:2512.21711, 2025

  17. [17]

    How do latent reasoning methods perform under weak and strong supervision? arXiv preprint arXiv:2602.22441, 2026

    Yingqian Cui, Zhenwei Dai, Bing He, Zhan Shi, Hui Liu, Rui Sun, Zhiji Liu, Yue Xing, Jiliang Tang, and Benoit Dumoulin. How do latent reasoning methods perform under weak and strong supervision? arXiv preprint arXiv:2602.22441, 2026

  18. [18]

    Unlocking the black box of latent reasoning: An interpretability-guided approach to intervention.arXiv preprint arXiv:2606.01243, 2026

    Shuochen Chang, Tong Bai, Xiaofeng Zhang, Qianli Ma, Qingyang Liu, Zhaohe Liao, Yibo Miao, and Li Niu. Unlocking the black box of latent reasoning: An interpretability-guided approach to intervention.arXiv preprint arXiv:2606.01243, 2026

  19. [19]

    Scratchpad think- ing: Alternation between storage and computation in latent reasoning models.NeurIPS Mechanistic Interpretability Workshop, 2025

    Sayam Goyal, Brad Peters, Maria Emilia Granda, Akshath Narmadha, Dharunish Yugeswardeenoo, Cole Blondin, Callum McDougall, Sean O’Brien, Ashwinee Panda, and Kevin Zhu. Scratchpad think- ing: Alternation between storage and computation in latent reasoning models.NeurIPS Mechanistic Interpretability Workshop, 2025

  20. [20]

    Observable pat- terns are not explanations: A causal-geometric analysis of latent reasoning models.arXiv preprint arXiv:2606.12689, 2026

    Darpan Aswal, Thomas Palmeira Ferraz, Yongxin Zhou, and Maxime Peyrard. Observable pat- terns are not explanations: A causal-geometric analysis of latent reasoning models.arXiv preprint arXiv:2606.12689, 2026

  21. [21]

    Demystifying hidden-state recurrence: Switchable latent reasoning with on-policy reinforcement learning.arXiv preprint arXiv:2606.13106, 2026

    Jiayu Yang, Chao Chen, Shengen Wu, Yinhong Liu, Yuxuan Fan, Lujundong Li, Songning Lai, Cheng- wei Qin, and Zhijiang Guo. Demystifying hidden-state recurrence: Switchable latent reasoning with on-policy reinforcement learning.arXiv preprint arXiv:2606.13106, 2026. 32