Pith. sign in

REVIEW 2 major objections 4 minor 47 references

Latent reasoning can look equally unfaithful at the finish line while taking different paths there—and the causal role of its hidden steps flips with answer format.

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 →

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

T0 review · grok-4.5

2026-07-11 00:26 UTC pith:D53J2XN3

load-bearing objection Trajectory tracking of latent faithfulness is real progress; the format claim is the softest part of an otherwise careful empirical paper. the 2 major comments →

arxiv 2607.06648 v1 pith:D53J2XN3 submitted 2026-07-07 cs.LG cs.CL

Final Checkpoints Are Not Enough: Analyzing Latent Reasoning Faithfulness Along Training Trajectories

Hengyu Jin , Shu Yang , Di Wang This is my paper
classification cs.LG cs.CL
keywords latent reasoningfaithfulnesstraining trajectoriesactivation patchingcounterfactual editsanswer formatcontinuous thoughtschain-of-thought
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 argues that judging latent reasoning only at a finished checkpoint misses how faithfulness forms. Latent methods keep multi-step work inside continuous hidden states instead of written steps, so the real question is whether those hidden steps still drive the answer. Tracking two training styles across saved checkpoints with verified input flips and noise patches on the latent steps, the authors show both styles can end up similarly unfaithful under counterfactual edits, yet one collapses at a curriculum boundary while the other is retention-like from early training. At the activation level the causal contribution of the latent steps decays on the same examples that stop following those flips. When only the answer format is changed—from binary choice to open-ended decoding—that activation trajectory reverses direction. Anyone building or auditing compact reasoning should care, because stage and format can change whether the latent path is still doing causal work.

Core claim

At the output level, different latent-reasoning training paradigms can look similarly unfaithful at convergence under verified counterfactual edits while following qualitatively divergent trajectories. At the activation level, the causal contribution of latent reasoning steps to the final answer decays across training for both paradigms, and the examples that flip from counterfactual-following to retention are the ones on which that contribution decays. Changing only the answer format reverses the activation-level trajectory: it decays on binary choice and rises on open-ended decoding. Latent reasoning faithfulness is therefore a property of training stage and answer format, not of the archi

What carries the argument

Verified single-edge counterfactual edits on ProsQA graph-reachability questions (so a BFS oracle flips the label) plus all-step content-erasing activation patches (zero, mean, or norm-matched noise) on continuous thought positions, both measured across the full saved-checkpoint trajectory.

Load-bearing premise

The claim hangs on treating all-step noise and mean patches of the continuous thought positions as a fair measure of whether those steps drive the answer, and on patterns from a small GPT-2 backbone and two specific training recipes generalizing beyond that setup.

What would settle it

If later checkpoints of both paradigms still showed high conditional flip rates on the verified ProsQA counterfactual pairs and high output-change rates under all-step norm-noise on the latent positions—or if the activation trajectory failed to reverse when the same GSM problems were switched only between binary choice and open-ended answers—the stage-and-format claim would not hold.

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

If this is right

  • A single final-checkpoint faithfulness score can collapse two different training stories into the same low-faithfulness summary.
  • Curriculum stage boundaries can abruptly destroy counterfactual following even while clean accuracy keeps rising.
  • Binary-choice formats can leave latent steps carrying removable answer bias rather than necessary intermediate computation.
  • Open-ended decoding can increase, rather than decrease, the measured causal role of latent steps along training.
  • Headline accuracy rankings of latent versus explicit reasoning can reverse when the same problems are reformatted as multiple choice versus free answers.

Where Pith is reading between the lines

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

  • If the goal is causal use of latent steps, training recipes may need mid-run interventions or format-aware curricula rather than waiting for the best validation checkpoint.
  • Faithfulness audits for continuous thoughts should report trajectories and answer-format controls, not only endpoint ablation rates.
  • Self-distillation may short-circuit intermediate causal use of latents earlier than staged curricula, which would matter for any method that compresses explicit traces into fixed latent slots.
  • The same checkpoint-trajectory design is a natural next test for other opaque multi-step mechanisms such as pause tokens or recurrent depth.

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

2 major / 4 minor

Summary. The paper argues that latent-reasoning faithfulness cannot be diagnosed from final checkpoints alone. On a shared GPT-2 backbone it tracks COCONUT and CODI (plus CoT/NoCoT baselines) across saved checkpoints, using (i) a BFS-verified single-edge counterfactual edit on ProsQA (405 pairs; Algorithm 1) that yields joint AB/AA label trajectories and CC/CFR metrics, and (ii) all-step content-erasing activation patches (zero / mean / norm-noise) that yield OCR, PWC, and cohort-level IEcontrast. Main findings: the two paradigms reach similarly low final-checkpoint counterfactual faithfulness via qualitatively different output-level trajectories (COCONUT collapses AB o AA at the curriculum boundary; CODI is retention-like early); the causal contribution of latent steps decays along training for both, and the Gdrift cohort that flips on the output side also loses IEcontrast under norm-noise; and the activation-level OCR trajectory reverses when only answer format is changed (decays on ProsQA/GSM-Choice, rises on GSM-open). The authors conclude that faithfulness is a property of training stage and answer format, not of architecture or the final checkpoint alone.

Significance. If the trajectory and format results hold, they change how latent-reasoning faithfulness should be evaluated: final-checkpoint replaceability studies (already common in the literature the paper cites) can collapse qualitatively different training stories into the same low-faithfulness summary, and binary-choice vs open-ended formats can reverse the measured mediation direction. Strengths that support this contribution include a verifiable, oracle-checked counterfactual construction (Algorithm 1), explicit joint-label and cohort analyses that link output flips to activation-level decay (Table 1–2, Table 6), multi-seed OCR/PWC reporting under three ablations, and a content-matched GSM-open vs GSM-Choice design that isolates format. The work is scoped carefully to GPT-2-scale COCONUT/CODI recipes and to settings with deterministic oracles, which is appropriate for a first trajectory study even if it limits immediate generality.

major comments (2)
  1. §4.1 and Appendix A.3: the central activation-level claim (latent steps lose causal contribution; Takeaway 2) rests on all-step content-erasing patches (zero/mean/norm-noise) of continuous-thought positions. Without a control that patches non-latent residual or early-layer positions of matched dimensionality, OCR o0 / PWC o1 is also consistent with the answer path simply bypassing the patched positions (as the shortcut literature the paper cites already suggests). The Gdrift IEcontrast drop in Table 2 is more specific but still uses the same all-step norm-noise and is reported only for COCONUT on ProsQA. A non-latent or single-step control, or an explicit residual-stream baseline, is needed before the mediation interpretation can be treated as load-bearing.
  2. §5 / Figure 5 / Takeaway 3: the claim that the activation-level trajectory reverses with answer format is under-specified relative to the experimental contrast. GSM-open and GSM-Choice differ not only in binary vs open-ended format but in output space (free-form numeric vs two-letter choice) and decoding budget (256 vs 8 tokens). Rising OCR on GSM-open can therefore reflect higher answer fragility under noise rather than rising latent mediation; the paper’s own Table 3 already shows that non-zero OCR on binary choice can be removable bias (w o c > c o w). Either a same-output-space format control or a clearer statement that the result is format-plus-decoding-regime dependence would make the claim proportionate to the evidence.
minor comments (4)
  1. Figure 1 caption and panels (b–d) are illustrative of the RQs but the actual quantitative curves live in later figures; a short note in the caption that (b–d) are schematic would reduce confusion.
  2. §3.2 / Figure 3: the AB o AA collapse at the COCONUT curriculum boundary is striking; pointing readers earlier to the raw counts in Appendix Table 6 would help.
  3. Notation table (Appendix Table 7) is useful; ensuring that IEcontrast / mi are defined once in the main text near Table 2 would improve readability for readers who skip the appendix.
  4. Limitations correctly flags GPT-2 scale and oracle dependence; a one-sentence statement of what would count as a failed generalization (e.g., larger backbone, non-curriculum latent method) would make the scope claim more falsifiable.

Circularity Check

0 steps flagged

No circularity: empirical trajectory measurements against external BFS oracles and interventional effect sizes, not quantities forced by definition or fit.

full rationale

The paper is an empirical analysis of latent-reasoning training dynamics. Its load-bearing claims (divergent output-level trajectories under verified ProsQA counterfactual edits; decaying OCR/PWC under all-step content-erasing patches; format-dependent reversal of the activation-level trajectory) are measurements of model behavior on held-out inputs and under interventions, not derivations that reduce to their own inputs. The ProsQA counterfactual construction (Algorithm 1) is verified by an independent BFS oracle that flips the label by construction of the graph edit, not by fitting to model outputs. OCR, PWC, CC, CFR, and IEcontrast are defined as interventional statistics (Eqs. 1–2 and Appendix A.3) and reported as observed trajectories; they are not parameters fitted to equal a target claim. Training recipes (COCONUT curriculum, CODI self-distillation) and patching techniques are taken from external literature and applied along saved checkpoints; there is no uniqueness theorem, ansatz, or self-citation chain that forces the reported trajectories. Minor self-contained cohort definitions (Table 1) merely partition examples by observed joint labels and do not redefine the metrics. The work is therefore self-contained against external benchmarks and interventions; circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The paper is empirical, not axiomatic theory. Load-bearing premises are standard ML domain assumptions (activation patching as causal probe; BFS oracle as ProsQA ground truth; curriculum/distillation recipes as given) plus analysis constructs (verified single-edge edits, joint labels, cohorts). Free parameters are training and intervention hyperparameters chosen by recipe or hand, not fitted to force the faithfulness conclusion. No new physical entities are postulated.

free parameters (5)
  • COCONUT curriculum schedule (cthought, epochs_per_stage, max_latent_stage)
    Hand-chosen staged latent budget growth that defines the epoch-15–16 boundary where the main AB→AA transition is observed; different schedules could move or remove that transition.
  • num_latent / latent length for CODI and terminal COCONUT stages
    Fixed latent budget (e.g., 6) is a design choice that shapes how much continuous thought capacity exists to ablate.
  • norm-noise construction (isotropic Gaussian rescaled to latent norm; 3 noise seeds)
    Intervention strength and sampling are experimenter choices that define the measured OCR/PWC effect sizes.
  • Activation-patching evaluation grids and 256-question sampling seeds
    Checkpoint subsets and subsample seeds are chosen by the authors; reported means depend on these choices though qualitative directions are stated to be seed-stable.
  • GSM-Choice distractor construction and gold-letter balancing
    Distractor sources and Pr[gold=A]=0.5 are design choices that define the binary-format comparison.
axioms (5)
  • domain assumption Content-erasing activation patches at latent positions measure the causal contribution of latent reasoning steps to the extracted answer.
    Invoked throughout §4–§5 and Appendix A.3 following Vig et al. / Heimersheim & Nanda; central to RQ2/RQ3.
  • domain assumption BFS reachability on the ProsQA graph is the correct oracle for original and edited answers.
    Used in Algorithm 1 and §3.1 to define verified counterfactual flips.
  • ad hoc to paper A model is counterfactually faithful on a pair when it is correct on the original and follows the oracle flip on the edited input (CC/CFR).
    Operational definition of output-level faithfulness in §3.1, adapted from contrast-set / counterfactual literature.
  • domain assumption With architecture, backbone, and dataset fixed, trajectory differences between COCONUT and CODI are attributable to training mechanism.
    Stated in §3.1 experimental setup; underwrites the paradigm-comparison claim.
  • domain assumption Matched GSM-open vs GSM-Choice content with the same COCONUT recipe isolates answer format as the causal factor for trajectory direction.
    §5.1 design claim for RQ3.
invented entities (2)
  • Verified single-edge ProsQA counterfactual edit (Algorithm 1) independent evidence
    purpose: Produce scalable, oracle-checked original/edited pairs that flip the answer with a minimal surface rewrite.
    Construction is new to this paper relative to the latent-faithfulness literature it cites; independent_evidence is true via BFS check on each accepted swap.
  • Transition cohorts Gstable / Gdrift / Gretain no independent evidence
    purpose: Link output-level AB→AA flips to activation-level decay of latent causal contribution.
    Analysis partitions defined in Table 1; useful but paper-internal constructs without external existence claims.

pith-pipeline@v1.1.0-grok45 · 22403 in / 3538 out tokens · 43828 ms · 2026-07-11T00:26:31.067359+00:00 · methodology

0 comments
read the original abstract

Latent reasoning methods perform multi-step inference entirely in the model's continuous hidden states, promising more compact and efficient reasoning. However, these opaque hidden states raise a question of faithfulness: whether these latent reasoning steps causally drive the final answer. Prior work investigates this question at converged checkpoints and reports several unfaithful behaviors, such as latent reasoning steps that can be replaced without changing the answer, but leaves how these behaviors form during training unexamined. We instead track how faithfulness evolves across saved checkpoints for different latent reasoning paradigms, applying a verifiable counterfactual edit on the input and a noise-ablation activation patch on the latent reasoning steps. We find that (i) at the output level, latent reasoning methods can look similarly unfaithful at convergence under counterfactual edits while following qualitatively divergent trajectories; (ii) at the activation level, the causal contribution of latent reasoning steps to the final answer decays across training for both paradigms, with the examples that flip on the output side in (i) also being the examples on which this contribution decays; and (iii) the activation-level trajectory diverges by answer format, decaying on binary choice and rising on open-ended decoding. These findings highlight that latent reasoning faithfulness depends on training stage and answer format.

Figures

Figures reproduced from arXiv: 2607.06648 by Di Wang, Hengyu Jin, Shu Yang.

Figure 1
Figure 1. Figure 1: Analysis along the training trajectory. (a) Counterfactual edit on ProsQA and noise-ablation patch on all latent reasoning steps, each evaluated across saved checkpoints. (b–d) Output-level (RQ1), activation-level (RQ2), and cross-format (RQ3) trajectories under these analyses. checkpoints along training on a shared GPT-2 back￾bone, applying the two interventions shown in Fig￾ure 1a. The first, illustrated… view at source ↗
Figure 2
Figure 2. Figure 2: Clean accuracy, Contrast Consistency (CC), and Conditional Flip Rate (CFR) on the 405 ProsQA oracle pairs at each model’s best checkpoint (CoT ep30, NoCoT ep14, CODI ep37, COCONUT ep48). NoCoT answer-only baseline, even though CO￾CONUT’s clean accuracy is the highest of the four paradigms; both methods’ CFR is roughly an or￾der of magnitude below the explicit-CoT baseline. CC and CFR agree at this snapshot… view at source ↗
Figure 3
Figure 3. Figure 3: Per-epoch joint label distribution on the 405 ProsQA oracle pairs for the four training paradigms. Each column is a stacked area over {AB, AA, BA, BB, AO, BO, O-cells}, where O-cells groups OO, OA, and OB. On the COCONUT panel, AB → AA collapses at the epoch 15–16 stage bound￾ary. rise above it, so the trajectory enters the retention regime within the first few epochs and stays there. The contrast with COC… view at source ↗
Figure 4
Figure 4. Figure 4: Causal analyses along the ProsQA training [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Latent-position dependence along the trajectory under the three all-step content-erasing interventions [PITH_FULL_IMAGE:figures/full_fig_p008_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

47 extracted references · 47 canonical work pages · 11 internal anchors

  1. [1]

    Educational and psychological measurement , volume =

    A coefficient of agreement for nominal scales , author =. Educational and psychological measurement , volume =. 1960 , publisher =

    work page 1960

  2. [2]

    biometrics , pages =

    The measurement of observer agreement for categorical data , author =. biometrics , pages =. 1977 , publisher =

    work page 1977

  3. [3]

    Physical therapy , volume =

    The kappa statistic in reliability studies: use, interpretation, and sample size requirements , author =. Physical therapy , volume =. 2005 , publisher =

    work page 2005

  4. [4]

    OpenAI blog , volume =

    Language models are unsupervised multitask learners , author =. OpenAI blog , volume =

    work page

  5. [5]

    Towards Faithfully Interpretable NLP Systems: How Should We Define and Evaluate Faithfulness?

    Jacovi, Alon and Goldberg, Yoav. Towards Faithfully Interpretable NLP Systems: How Should We Define and Evaluate Faithfulness?. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653/v1/2020.acl-main.386

    work page doi:10.18653/v1/2020.acl-main.386 2020

  6. [6]

    Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    Evaluating models’ local decision boundaries via contrast sets , author=. Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    work page 2020

  7. [7]

    Advances in neural information processing systems , volume =

    Beyond accuracy: quantifying trial-by-trial behaviour of CNNs and humans by measuring error consistency , author =. Advances in neural information processing systems , volume =

    work page

  8. [8]

    Advances in neural information processing systems , volume =

    Investigating gender bias in language models using causal mediation analysis , author =. Advances in neural information processing systems , volume =

    work page

  9. [9]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages =

    ERASER: A benchmark to evaluate rationalized NLP models , author =. Proceedings of the 58th annual meeting of the association for computational linguistics , pages =

    work page

  10. [10]

    International Conference on Learning Representations (ICLR) , year =

    Learning the Difference that Makes a Difference with Counterfactually Augmented Data , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  11. [11]

    Show your work: Scratchpads for intermediate computation with language models , author =

    work page

  12. [12]

    Training Verifiers to Solve Math Word Problems

    Training verifiers to solve math word problems , author =. arXiv preprint arXiv:2110.14168 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets

    Grokking: Generalization beyond overfitting on small algorithmic datasets , author =. arXiv preprint arXiv:2201.02177 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    2022 , journal =

    In-context Learning and Induction Heads , author =. 2022 , journal =

    work page 2022

  15. [15]

    2022 , journal =

    Emergent abilities of large language models , author =. 2022 , journal =

    work page 2022

  16. [16]

    Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

    Interpretability in the wild: a circuit for indirect object identification in gpt-2 small , author =. arXiv preprint arXiv:2211.00593 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Advances in neural information processing systems , volume =

    Locating and editing factual associations in gpt , author =. Advances in neural information processing systems , volume =

    work page

  18. [18]

    Advances in neural information processing systems , volume =

    Chain-of-thought prompting elicits reasoning in large language models , author =. Advances in neural information processing systems , volume =

    work page

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    Memorization without overfitting: Analyzing the training dynamics of large language models , author=. Advances in Neural Information Processing Systems , volume=

    work page

  20. [20]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , month = jul, year =

    Faithfulness Tests for Natural Language Explanations , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , month = jul, year =. doi:10.18653/v1/2023.acl-short.25 , pages =

    work page doi:10.18653/v1/2023.acl-short.25 2023

  21. [21]

    Advances in neural information processing systems , volume =

    Faith and fate: Limits of transformers on compositionality , author =. Advances in neural information processing systems , volume =

    work page

  22. [22]

    Advances in Neural Information Processing Systems , volume =

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author =. Advances in Neural Information Processing Systems , volume =

    work page

  23. [23]

    The Eleventh International Conference on Learning Representations , year =

    Progress measures for grokking via mechanistic interpretability , author =. The Eleventh International Conference on Learning Representations , year =

    work page

  24. [24]

    Measuring Faithfulness in Chain-of-Thought Reasoning

    Measuring faithfulness in chain-of-thought reasoning , author =. arXiv preprint arXiv:2307.13702 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    How to use and interpret activation patching

    How to use and interpret activation patching , author =. arXiv preprint arXiv:2404.15255 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Training Large Language Models to Reason in a Continuous Latent Space

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

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Multiple-Choice Questions are Efficient and Robust LLM Evaluators

    Multiple-choice questions are efficient and robust llm evaluators , author =. arXiv preprint arXiv:2405.11966 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step

    From explicit cot to implicit cot: Learning to internalize cot step by step , author =. arXiv preprint arXiv:2405.14838 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    International Conference on Learning Representations , volume =

    Think before you speak: Training language models with pause tokens , author =. International Conference on Learning Representations , volume =

    work page

  30. [30]

    Advances in Neural Information Processing Systems , volume=

    LLM circuit analyses are consistent across training and scale , author=. Advances in Neural Information Processing Systems , volume=

    work page

  31. [31]

    Towards Faithful Model Explanation in NLP : A Survey

    Lyu, Qing and Apidianaki, Marianna and Callison-Burch, Chris. Towards Faithful Model Explanation in NLP : A Survey. Computational Linguistics. 2024. doi:10.1162/coli_a_00511

    work page doi:10.1162/coli_a_00511 2024

  32. [32]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =

    Codi: Compressing chain-of-thought into continuous space via self-distillation , author =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =

    work page 2025

  33. [33]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Softcot: Soft chain-of-thought for efficient reasoning with llms , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    work page

  34. [34]

    LLM Pretraining with Continuous Concepts

    Llm pretraining with continuous concepts , author =. arXiv preprint arXiv:2502.08524 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    2025 , eprint =

    A Survey on Latent Reasoning , author =. 2025 , eprint =

    work page 2025

  36. [36]

    arXiv preprint arXiv:2505.16782 , year =

    Reasoning beyond language: A comprehensive survey on latent chain-of-thought reasoning , author =. arXiv preprint arXiv:2505.16782 , year =

    work page arXiv

  37. [37]

    arXiv preprint arXiv:2512.21711 , year =

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

    work page arXiv

  38. [38]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    Implicit reasoning in transformers is reasoning through shortcuts , author =. Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    work page 2025

  39. [39]

    arXiv preprint arXiv:2508.03440 , year =

    Llms are single-threaded reasoners: Demystifying the working mechanism of soft thinking , author =. arXiv preprint arXiv:2508.03440 , year =

    work page arXiv

  40. [40]

    Advances in Neural Information Processing Systems , volume =

    Soft thinking: Unlocking the reasoning potential of llms in continuous concept space , author =. Advances in Neural Information Processing Systems , volume =

    work page

  41. [41]

    Advances in Neural Information Processing Systems , volume =

    Scaling up test-time compute with latent reasoning: A recurrent depth approach , author =. Advances in Neural Information Processing Systems , volume =

    work page

  42. [42]

    Advances in Neural Information Processing Systems , volume =

    Hybrid latent reasoning via reinforcement learning , author =. Advances in Neural Information Processing Systems , volume =

    work page

  43. [43]

    Advances in Neural Information Processing Systems , volume =

    Reasoning by superposition: A theoretical perspective on chain of continuous thought , author =. Advances in Neural Information Processing Systems , volume =

    work page

  44. [44]

    arXiv preprint arXiv:2602.00449 , year =

    Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks , author =. arXiv preprint arXiv:2602.00449 , year =

    work page arXiv

  45. [45]

    Dynamics Within Latent Chain-of-Thought: An Empirical Study of Causal Structure

    Dynamics Within Latent Chain-of-Thought: An Empirical Study of Causal Structure , author =. arXiv preprint arXiv:2602.08783 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2602.22441 , year =

    How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision? , author =. arXiv preprint arXiv:2602.22441 , year =

    work page arXiv

  47. [47]

    Are Latent Reasoning Models Easily Interpretable?

    Are Latent Reasoning Models Easily Interpretable? , author =. arXiv preprint arXiv:2604.04902 , year =

    work page internal anchor Pith review Pith/arXiv arXiv