Pith. sign in

REVIEW 4 major objections 7 minor 13 references

Carrying intermediate reasoning states as vocabulary-indexed coefficients outperforms both text chains and dense hidden vectors on the tasks tested.

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 ↗

J-CoT improves LLM reasoning by passing a sparse, vocabulary-indexed coefficient state—not text or full hidden vectors—between recurrent computation cycles.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection J-CoT is a real new interface with a solid trained variant; the zero-shot claim is noise-level and the transport precondition is unverified. the 4 major comments →

arxiv 2607.21981 v1 pith:UQWJGVY2 submitted 2026-07-24 cs.CL cs.AI

J-CoT: Chain-of-Thought in J-Space

Junde Wu , Jiayuan Zhu , Fengling Liu , Minhao Hu , Jiazhen Pan This is my paper
classification cs.CL cs.AI
keywords chain-of-thoughtlatent reasoningJ-spacerecurrent reasoningintermediate statesvocabulary-indexed representationslanguage model inferenceread-write interface
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.

The reading

Chain-of-thought drives a language model by passing decoded sentences between reasoning steps; J-CoT argues that this over-commits the model to language, while dense latent recurrence over-commits to an unstructured vector. The paper proposes a third interface: at each cycle boundary, the model writes only a sparse set of vocabulary-indexed coefficients — a J-thought — extracted from a few non-decoded carrier positions, and reads it back at an earlier layer to seed the next cycle. The authors claim that under matched backbone and inference settings, the training-free J-CoT-Zero matches or exceeds the strongest evaluated latent-reasoning baseline on every benchmark, and that the lightly trained J-CoT-Train gets the highest scores across math, science, coding, and structured path-reasoning tasks. A reader should care because this is evidence that a model's own hidden geometry can serve as a recurrent workspace without either verbalizing intermediate states or forcing the whole hidden vector across the boundary.

Core claim

The claim: the interface, not the model, is the bottleneck in stepwise reasoning. J-CoT carries reasoning state between cycles as a J-thought — sparse vocabulary-indexed coefficients extracted from non-decoded carrier positions and read back at an earlier layer — instead of decoded sentences or dense hidden vectors. The authors report that with matched backbone and inference settings, the training-free variant matches or exceeds the strongest latent-reasoning baseline on every benchmark, and the trained variant tops all eight benchmarks, lifting the average from 47.5 to 50.2. A conditional stability bound (Proposition 1) guarantees multi-hop transport when each layer's recovery defect is sma

What carries the argument

J-space: a vocabulary-indexed coordinate system in the model's hidden representations, where each token names a residual direction per layer. Dictionaries D_ℓ are built by pulling the unembedding directions through an averaged downstream Jacobian, so the same coefficient vector refers to the same vocabulary-indexed identity in different layers. A read map reconstructs the J-thought through the read-layer dictionary; non-decoded carrier positions carry it through ordinary Transformer blocks; a write map re-extracts coefficients with a nonnegative sparse regression at a later layer. Proposition 1 bounds the transport error by the per-layer recovery defect ε_ℓ(a), making small reconstruction er

Load-bearing premise

Everything rests on the assumption that a coefficient vector extracted from a carrier activation at the write layer reconstructs well enough at the read layer — a recovery defect the paper bounds conditionally but never measures on the actual states produced during reasoning.

What would settle it

Measure the recovery defect ε_ℓ(a) from Eq. 2 on the carrier activations J-CoT itself produces across the benchmark tasks, and compare against the same quantity on random or calibration states. If the defect is large for the reasoning states (or no better than for random coefficient vectors), the 'same coefficient identity' fails to survive transport and the recurrent read-write reduces to injecting noise; equivalently, an ablation that substitutes random vocabulary-indexed coefficients of matched sparsity would expose whether the specific J-space geometry is doing the work.

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

If this is right

  • Because the J-space dictionaries are fixed and estimated once per backbone, J-CoT-Zero adds a recurrent interface to an already fine-tuned model without any interface-specific training.
  • The trained variant shows that optimizing only the carrier read-in (embeddings and a gate), while freezing the Transformer and dictionaries, buys a further 2.3-point average gain, so the interface itself is a recoverable source of performance.
  • Measured inference FLOPs grow with recurrent depth and backbone scale, and gains from deeper reasoning widen from 7B to 405B, implying the method does not saturate at current scales.
  • The interface-spectrum experiment, where the recurrent state is continuously blended between dense, J-thought, and linguistic, finds the best accuracy at the J-thought operating point, suggesting an optimal middle ground between full verbalization and full density.
  • The method's adaptive stopping rule means recurrent depth is per-example rather than fixed, which can trade compute against accuracy at inference time.

Where Pith is reading between the lines

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

  • Editorial: the paper's transport guarantee is conditional on small recovery defect but never measures ε_ℓ on actual reasoning states; a direct measurement on collected carrier activations, or a control using random coefficient vectors of the same sparsity, would show whether the specific J-space geometry is what drives the gains.
  • Editorial: because the dictionaries are built once on an unlabeled calibration corpus and reused across tasks, the same interface should transfer to other Transformer backbones and to other recurrent loops (e.g., multi-pass tool use or retrieval), a direction the paper leaves unstated.
  • Editorial: the interface-spectrum result suggests task-dependent optima; one testable extension is to learn or search the interpolation point per benchmark rather than fixing full J-thought state.
  • Editorial: the method's nonnegative coefficients give a natural degree of interpretability — active columns can be labeled by vocabulary strings — which could be used for post-hoc rationalization of intermediate states.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 7 minor

Summary. The paper introduces J-CoT, a recurrent reasoning interface that stores information between computation cycles as vocabulary-indexed coefficient vectors ('J-thoughts') in 'J-space'. The dictionaries are derived from the averaged downstream Jacobian on unlabeled calibration data, and read/write operations use nonnegative elastic-net decoding/encoding. Two variants are evaluated: J-CoT-Zero, which uses fixed interface components, and J-CoT-Train, which optimizes carrier embeddings and a read gate. On a shared Qwen3-8B-Base backbone, J-CoT-Train outperforms all baselines on eight benchmarks, while J-CoT-Zero is reported to match or exceed the SIM-Coconut baseline on all benchmarks. Results are also presented for scaling with model capacity and for an interpolation between dense, J-space, and linguistic recurrent states.

Significance. If the transport mechanism is validated, J-CoT proposes a genuinely intermediate interface between explicit CoT and dense latent recurrence, with a training-free variant that does not require fitted parameters. The paper ships a formal conditional transport bound (Prop. 1) and a reproducible evaluation protocol (shared checkpoints, fixed templates, matched compute). The reported J-CoT-Train gains (2.1–3.6 points) are consistent and exceed standard deviations on most tasks. However, the contribution as presented is undercut by the unmeasured recovery defect, the statistically unsupported zero-shot claim, and the missing appendix; the significance of the paper depends on resolving these.

major comments (4)
  1. [§3.2, Eq. (2)–(3)] The recovery defect ε_ℓ(a) is the load-bearing precondition of Proposition 1, but it is never measured on any activation, calibration or task. Since J-CoT-Zero has no learned interface, the observed gains could in principle be caused by the extra recurrent compute or by the stochastic/dictionary noise, rather than by J-space geometry. Please (i) report ε_ℓ for the actual carrier activations in at least one benchmark, and (ii) add a control that replaces D_ℓ with random directions (or a permuted vocabulary index) to show that the transport property is specific to the J-space dictionaries.
  2. [§4.2, Table 1] J-CoT-Zero's improvements over SIM-Coconut are 0.3–0.6 points on seven benchmarks and a tie on AIME 2024; with the reported standard deviations (0.5–3.5), all these differences are within one standard deviation. The abstract's statement that J-CoT-Zero 'matches or exceeds the strongest evaluated latent-reasoning baseline on every benchmark' is therefore an overstatement. Please provide paired significance tests (e.g., bootstrap over problems) or rephrase the claim as 'is competitive with' / 'differences are not statistically significant'.
  3. [§3.2, §4.1, §4.3 (missing appendix)] The manuscript defers to an Appendix that is not included in the submission. The proof of Proposition 1, the randomized Jacobian estimator, the elastic-net solver, the stopping procedure, and the complete evaluation rules are all said to be there; explicit placeholders 'Appendices??and??' appear in §3.2. As submitted, these details cannot be audited. The appendix must be part of the revision.
  4. [§4.3, Fig. 2 and §4.4, Fig. 3] The scaling and interface-spectrum experiments report point estimates without standard deviations or significance tests. The non-monotonic peak at λ=0.5 and the Heavy-over-Instant gains could be within noise, especially given the small AIME-style sample sizes elsewhere in the paper. Report variances, sample sizes, and paired comparisons for these figures.
minor comments (7)
  1. [§3.1] 'therotical' should be 'theoretical'.
  2. [§3.2] Resolve the 'Appendices??and??' placeholders with actual cross-references.
  3. [Table 1] AIME 2024 standard deviations (3.0–3.8) are large enough that row-level comparisons are uninformative; consider reporting the number of problems and exact-match counts.
  4. [§4.1 vs §4.2] The paper says hyperparameters are 'selected on development data' and elsewhere 'selected once using the aggregate development set'; clarify whether these are the same split and whether any benchmark-level adaptation occurred.
  5. [§4.3] Specify whether the scaling study uses J-CoT-Zero or J-CoT-Train and how the dictionaries are estimated for Qwen2.5/Llama-3.1 backbones.
  6. [References] Gurnee et al. (2026) is cited as a web publication; include a stable identifier beyond the Transformer Circuits URL.
  7. [§4.2] The claim of 'matched sequential-compute budget' is not substantiated in Table 1; list the actual FLOPs or cycle budgets for each method.

Circularity Check

0 steps flagged

No circularity: J-CoT's central claim is empirical, and its dictionaries and read–write maps are fixed from unlabeled calibration data rather than fitted to the reported benchmarks.

full rationale

I walked the derivation chain and found no step in which a 'prediction' is equivalent to a fitted input by construction. The J-space dictionaries D_l are defined in Eq. 5 from the model's own averaged downstream Jacobian and unembedding, estimated once on an unlabeled calibration corpus ('The dictionaries are estimated once for each backbone on an unlabeled calibration corpus and remain fixed during J-CoT training and evaluation', §3.2), and J-CoT-Zero adds no interface-specific optimization (§4.2). The reported benchmark numbers are obtained by actually running this fixed interface on held-out tasks; they are not solutions of the defining equations. Proposition 1 (Eq. 3) is explicitly conditional on the recovery defect epsilon_l(a) (Eq. 2); it is a triangle-inequality bound stating that small epsilon gives small transport error, not a claim that epsilon is small. The paper never measures epsilon for actual reasoning states, and the proof is deferred to an absent appendix ('The proof is given in Appendix', §3.1; 'Appendices??and??', §3.2) — these are verification gaps, not circularity. The only self-referential aspect is that the coordinate system is built from the model's own unembedding, but this is not used to fit the evaluation outcomes and does not make the empirical claim reduce to its inputs. No load-bearing self-citation appears: Gurnee et al. (2026) is an external citation, and the paper constructs its own dictionaries rather than importing a uniqueness theorem. Thus the central claims are self-contained relative to the circularity patterns considered; score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 2 invented entities

The central claim depends on a small set of hand-chosen hyperparameters (elastic-net penalties, read/write layers, carrier count, stopping rule) and on three domain assumptions: the Jacobian linearization is faithful, the recovery defect is small, and the prior J-space construction is valid. The transport guarantee is conditional on an unmeasured quantity, and the missing appendix prevents auditing the dictionary estimation. No code or data are released.

free parameters (5)
  • Elastic-net penalties λ1, λ2 = λ1=0.05, λ2=1e-3
    Control sparsity and stability of J-thought extraction; chosen by hand in §4.1 and affect which coefficients survive the recurrent boundary.
  • Read/write layers ℓ_r, ℓ_w = ℓ_r=12, ℓ_w=28 for Qwen3-8B
    Define the recurrent computation interval; selected on development data (§4.1). Transport properties depend on this choice.
  • Carrier count M = 8
    Number of parallel coefficient channels; a design choice evaluated only partially in ablations (§4.5).
  • Stopping threshold and max cycles = r_t < 0.02 for two consecutive cycles, max 8
    Adaptive recurrence hyperparameters that trade compute for accuracy (§4.1).
  • J-lens calibration corpus = 1,000 sequences of 128 tokens
    Dictionary estimation set; details of the randomized Jacobian estimator are in a missing appendix, so this choice cannot be audited (§4.1).
axioms (3)
  • domain assumption The averaged downstream Jacobian J_ℓ is an adequate linear surrogate for how layer-ℓ residual changes affect final-layer states.
    Equation (4) treats the map as linear; if the linearization is inaccurate for carrier activations, the dictionaries D_ℓ (Eq. 5) will not align across layers.
  • domain assumption The nonnegative elastic-net extraction has a solution, and the recovery defect ε_ℓ is small for task carrier activations.
    Proposition 1 only guarantees transport under small ε_ℓ; the paper never measures or reports ε_ℓ for actual reasoning states.
  • domain assumption The prior J-space result (Gurnee et al., 2026) is valid: vocabulary-indexed residual directions are shared across layers.
    J-CoT treats J-space as a usable coordinate system and cites Gurnee et al. without independent verification in this paper.
invented entities (2)
  • J-thought no independent evidence
    purpose: A matrix of nonnegative vocabulary-indexed coefficients serving as the recurrent boundary state between reasoning cycles.
    No external falsifiable handle; its existence is evidenced only by the paper's own experiments on the same model family.
  • Carrier positions no independent evidence
    purpose: Non-linguistic appended positions that expose residual activations for reading and writing J-thoughts.
    Similar to pause tokens, but without an external prediction or measurement that would validate them independently of J-CoT.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of J-CoT: Chain-of-Thought in J-Space." pith.science (2026). https://pith.science/paper/UQWJGVY2

@misc{pith2026260721981,
  author       = {Pith},
  title        = {Pith review of: J-CoT: Chain-of-Thought in J-Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQWJGVY2}},
  note         = {Machine review of arXiv:2607.21981}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Chain-of-thought prompting improves language-model reasoning by carrying intermediate states across successive computation steps. However, relying on natural language as the only recurrent interface is overly restrictive, since many transient computations do not need to be fully verbalized. Existing latent-reasoning methods remove this constraint by recurrently propagating continuous hidden states. However, these methods pass a dense hidden vector as a whole, without an explicit mechanism for selecting and organizing the information needed by the next reasoning step. This motivates an intermediate interface that remains linguistically grounded without requiring a decoded sentence. We introduce \textbf{J-CoT}, a recurrent reasoning framework built on \emph{J-space}, a vocabulary-indexed coordinate system within the model's hidden representations. Within each cycle, the model computes in its full hidden space. At the cycle boundary, J-CoT expresses the intermediate state as vocabulary-indexed coefficients, carries these coefficients forward as a \emph{J-thought}, and maps them back into the model's hidden representation for the next cycle. J-CoT therefore requires neither a fluent intermediate rationale nor recurrence over the complete hidden state. Under matched backbone and inference settings, J-CoT-Zero matches or exceeds the strongest evaluated latent-reasoning baseline on every benchmark, while J-CoT-Train obtains the highest score across the evaluated mathematical, scientific, coding, and structured path-reasoning tasks.

Figures

Figures reproduced from arXiv: 2607.21981 by Fengling Liu, Jiayuan Zhu, Jiazhen Pan, Junde Wu, Minhao Hu.

Figure 1
Figure 1. Figure 1: Overview of recurrent reasoning interfaces. Explicit chain-of-thought passes decoded [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: J-CoT scaling with model capacity and recurrent depth. B, L, XL, XXL, and H denote [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Reasoning behavior across the recurrent-interface spectrum. We continuously interpolate [PITH_FULL_IMAGE:figures/full_fig_p011_3.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

13 extracted references · 7 linked inside Pith

  1. [5]

    URLhttps://arxiv.org/abs/2310. 02226. Alex Gu, Baptiste Roziere, Hugh Leather, Armando Solar-Lezama, Gabriel Synnaeve, and Sida I. Wang. CRUXEval: A benchmark for code reasoning, understanding and execution.arXiv preprint arXiv:2401.03065,

    Pith/arXiv arXiv

  2. [7]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt

    URLhttps://arxiv.org/abs/2412.06769. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. In Advances in Neural Information Processing Systems: Datasets and Benchmarks Track,

    Pith/arXiv arXiv

  3. [8]

    LiveCodeBench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974,

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. LiveCodeBench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974,

    Pith/arXiv arXiv

  4. [12]

    org/abs/2404.15758

    URLhttps://arxiv. org/abs/2404.15758. Qwen Team. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    Pith/arXiv arXiv

  5. [13]

    CODI: Compressing chain-of-thought into continuous space via self-distillation

    Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. CODI: Compressing chain-of-thought into continuous space via self-distillation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 677–693,

    2025

  6. [15]

    An Yang et al

    URLhttps://arxiv.org/abs/2509.20317. An Yang et al. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115,

    arXiv

  7. [17]

    Hui Zou and Trevor Hastie

    URLhttps://arxiv.org/abs/2505.12514. Hui Zou and Trevor Hastie. Regularization and variable selection via the elastic net.Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2):301–320,

    arXiv

  8. [2005]

    doi: 10.1111/ j.1467-9868.2005.00503.x. 14

    arXiv 2005

  9. [2019]

    2024 American Invitational Mathematics Examination,

    Mathematical Association of America. 2024 American Invitational Mathematics Examination,

    2024

  10. [2022]

    Let’s verify step by step.arXiv preprint arXiv:2305.20050,

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step.arXiv preprint arXiv:2305.20050,

    Pith/arXiv arXiv

  11. [2023]

    URLhttps://arxiv.org/abs/2303.08112. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Pith/arXiv arXiv

  12. [2024]

    org/abs/2405.14838

    URLhttps://arxiv. org/abs/2405.14838. Abhimanyu Dubey et al. The Llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Pith/arXiv arXiv

  13. [2025]

    URLhttps://aclanthology.org/2025

    doi: 10.18653/v1/2025.emnlp-main.36. URLhttps://aclanthology.org/2025. emnlp-main.36/. Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. InProceedings of the 61st Annual Meeting of the Association for Computational Lin...

    doi:10.18653/v1/2025.emnlp-main.36 2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.