{"id":"c05d2588-2bdd-4e7b-bcf2-98d49ef3585e","arxiv_id":"2607.21981","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"J-CoT improves LLM reasoning by passing a sparse, vocabulary-indexed coefficient state—not text or full hidden vectors—between recurrent computation cycles.","lead":"The paper introduces J-CoT, a way to let a language model carry its intermediate reasoning state as a compact set of vocabulary-indexed coefficients rather than as words or a dense hidden vector. Across math, science, and coding benchmarks, the authors report that this 'J-thought' state matches or beats previous latent-reasoning methods, and that a lightly trained version scores highest.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unmeasured recovery defect ε_ℓ leaves the J-CoT transport precondition unverified; the claimed interface may be injecting noise rather than carrying coherent coefficients.","rationale":"We read the paper in good faith. The empirical results, particularly for J-CoT-Train, are consistent and the ablations support the design choices. However, the central theoretical justification for the J-space interface is Proposition 1, which is explicitly conditional on small recovery defects. The paper never measures ε_ℓ(a) for the activations it actually transports; the only calibration is on 1,000 generic sequences. This is the load-bearing gap because if recovery is poor, the interface is not doing what is claimed—it is just a noisy recurrent perturbation, and the scientific contribution of 'vocabulary-indexed coefficient transport' is not established. The manuscript's frequent references to an omitted appendix (e.g., §3.2 'Appendices?? and ??', §4.1 'Appendix gives the complete stopping procedure') further prevent independent verification of the estimator, solver, and proof. These are missing-support flags that weigh against full acceptance. The proposed test—measuring ε on actual reasoning states—directly settles the concern. Because the reader's CONDITIONAL verdict already captures this uncertainty and the empirical trend is promising, we do not recommend changing the verdict.","tokens_in":12460,"tokens_out":12485,"duration_ms":119155,"concrete_test":"For each benchmark input and recurrent cycle, compute a = Φ_ℓw(Z_t^w), reconstruct at the read layer as Ψ_ℓr(a), re-extract ã = Φ_ℓr(Ψ_ℓr(a)), and report the distribution of ‖ã−a‖₂/‖a‖₂ and of ‖h−D_ℓ a‖₂/‖h‖₂ (reconstruction error). If the median coefficient change exceeds ~10%, the transport guarantee of Proposition 1 is violated for actual reasoning states, confirming the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's recurrent-interface mechanism depends on a conditional transport guarantee: Proposition 1 (Eq. 3) states that multi-hop transport of a coefficient vector a is close to the direct reconstruction only if the recovery defect ε_ℓ(a) in Eq. 2 is small. The abstract and §4.2 claim J-CoT-Zero matches or exceeds latent baselines, but the paper never reports ε_ℓ for any activation. The dictionaries D_ℓ are estimated from an averaged downstream Jacobian on 1,000 calibration sequences (§3.2, Eq. 5); the resulting columns may not sparsely span the carrier activations that arise during actual reasoning. If ε_ℓ is large, the written coefficients do not survive the read-layer reconstruction: the 'same coefficient identity' is lost, and the read–write step degenerates into adding dictionary noise to the residual stream. Since J-CoT-Zero has no learned interface, its empirical gains would then be attributable to extra computation depth or stochastic perturbation, not to the J-space geometry the paper claims. The ablations (Table 2) compare to learned latent states but do not measure ε or include a random-dictionary control. The missing appendices—referenced repeatedly in §3.2 and §4.1—prevent auditing the Jacobian estimator, the elastic-net solver, and the proof of Proposition 1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":12773,"tokens_out":7649,"duration_ms":72434,"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":[{"comment":"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.","section":"§3.2, Eq. (2)–(3)"},{"comment":"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'.","section":"§4.2, Table 1"},{"comment":"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.","section":"§3.2, §4.1, §4.3 (missing appendix)"},{"comment":"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.","section":"§4.3, Fig. 2 and §4.4, Fig. 3"}],"minor_comments":[{"comment":"'therotical' should be 'theoretical'.","section":"§3.1"},{"comment":"Resolve the 'Appendices??and??' placeholders with actual cross-references.","section":"§3.2"},{"comment":"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.","section":"Table 1"},{"comment":"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.","section":"§4.1 vs §4.2"},{"comment":"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.","section":"§4.3"},{"comment":"Gurnee et al. (2026) is cited as a web publication; include a stable identifier beyond the Transformer Circuits URL.","section":"References"},{"comment":"The claim of 'matched sequential-compute budget' is not substantiated in Table 1; list the actual FLOPs or cycle budgets for each method.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript shows promise for J-CoT-Train, but the missing appendix and the unverified transport precondition are blocking issues. The editor may wish to ask the authors for the appendix and a random-dictionary control before further consideration. Also, the J-CoT-Zero claim should be toned down."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: J-CoT is a genuinely new recurrent interface — sparse, nonnegative coefficients over vocabulary-indexed J-space directions carried between cycles — and J-CoT-Train's 2–3 point gains over the strongest latent baseline SIM-Coconut are the most defensible results in the paper. Across eight benchmarks the gains are consistent (2.1–3.6 points) and exceed the reported standard deviations. That is a real empirical contribution. The zero-shot variant's claim to 'match or exceed every benchmark' is true only numerically: margins of 0.3–0.6 points are inside one standard deviation, so it should be reported as a tie. The abstract overstates it.\n\nWhat's actually new: the interface itself, with carrier positions, elastic-net read/write maps, and a learned read gate. None of the cited latent-reasoning papers use a vocabulary-indexed coefficient state, and the J-space paper didn't use it recurrently. The experimental setup is careful about matched backbone, shared reasoning-adapted checkpoint, and matched compute for linguistic baselines. J-CoT-Train's consistent gains across math, science, and code merit attention.\n\nWhere it's soft: Proposition 1 is conditional on the recovery defect εℓ being small, and εℓ is never measured on real activations. Without that measurement you can't tell whether the read–write step carries coherent coefficients or just adds dictionary noise. A random-dictionary control would have been the cheapest way to address this; the paper doesn't include one. Second, the scaling study only varies J-CoT's own budget and backbone size — there's no matched-compute baseline like Coconut or CoT at the same FLOPs, so it's a scaling curve, not a comparison. Third, the appendices describing the Jacobian estimator, the elastic-net solver, and the proof are missing, so the method can't be audited from this version.\n\nThe circularity concern is muted: dictionaries are fixed from 1,000 unlabeled sequences and the read/write maps are deterministic, so the main empirical claim is not a fitted artifact. These are fixable issues: measure εℓ, run a random-dictionary control, add significance tests, and release the appendix.\n\nThis paper is for people working on latent reasoning and recurrent interfaces. It deserves a serious referee — the idea is plausible, the trained results are consistent, and the weaknesses are addressable. I'd expect major revision, and I wouldn't rely on the zero-shot claim until the control and measurements are in.","headline":"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.","tokens_in":13297,"tokens_out":3699,"would_cite":true,"duration_ms":38855,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Carrying intermediate reasoning states as vocabulary-indexed coefficients outperforms both text chains and dense hidden vectors on the tasks tested.","keywords":["chain-of-thought","latent reasoning","J-space","recurrent reasoning","intermediate states","vocabulary-indexed representations","language model inference","read-write interface"],"falsifier":"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.","tokens_in":12323,"feed_emoji":"🧠","tokens_out":6699,"duration_ms":71998,"temperature":0.7,"pith_summary":"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.","feed_headline":"A word-tagged 'J-thought' out-reasons text chains and hidden vectors","feed_subtitle":"Training-free, it matches the best latent baseline everywhere; light tuning makes it the top scorer on math, science, and code.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Sparse word-tagged thoughts beat text chains and dense vectors in reasoning","J-CoT: reasoning state as vocabulary coefficients, not sentences or hidden vectors","The interface is the bottleneck: J-thoughts outperform full text and latent baselines","Training-free J-CoT matches latent baselines; with tuning, tops math, science, code","Carry reasoning as sparse word coefficients—no fluent rationale needed"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Sparse word-tagged thoughts beat text chains and dense vectors in reasoning","J-CoT: reasoning state as vocabulary coefficients, not sentences or hidden vectors","The interface is the bottleneck: J-thoughts outperform full text and latent baselines","Training-free J-CoT matches latent baselines; with tuning, tops math, science, code","Carry reasoning as sparse word coefficients—no fluent rationale needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1190,"prompt_tokens":774,"completion_tokens":416,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":327}},"tokens_in":518,"tokens_out":416,"duration_ms":4814,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:08:03.007362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}