{"id":"472c9625-404b-4e5b-a9a7-54a4ebcc8197","arxiv_id":"2507.06203","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey that organizes latent reasoning methods into vertical recurrence, horizontal recurrence, and infinite-depth diffusion, arguing that silent reasoning can beat explicit chain-of-thought.","lead":"This survey maps the field of AI systems that reason silently inside their own mathematical spaces instead of writing out thoughts in words. It groups that research into three families and argues that silent reasoning can be faster and more expressive than text-based reasoning.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed equivalence between temporal and depth recurrence in §5.2 is the load-bearing bridge of the taxonomy, but it is asserted rather than derived; the Eq. 13 update changes its objective with every token, so longer sequences are not shown to be more iterations of the same layer.","rationale":"The reader's weakest assumption correctly identifies the depth-time commensurability claim in Section 5.2 as the main soft spot, and I agree that it is the load-bearing bridge of the taxonomy. My stress-test sharpens this into a concrete technical objection: the gradient-state update in Eq. 13 is an online optimization over a changing data stream, not repeated application of the same layer to the same input, so the claimed mathematical equivalence is not established and, in the stated form, is unlikely to hold for general inputs. The survey is otherwise candid: it explicitly flags the absence of empirical comparison and documents failed reproductions (e.g., Infini-attention), which is good scholarly practice. The issue is therefore not fatal to the survey's descriptive value, but it does mean the unifying claim in Section 5.2.2 should be presented as an open hypothesis rather than a mathematical consequence. Since the reader's CONDITIONAL verdict already asks for exactly this reframing, I do not recommend changing the verdict; the concern reinforces it but does not move it.","tokens_in":29943,"tokens_out":3889,"duration_ms":47741,"concrete_test":"Re-derive Section 5.2's equivalence on a minimal case: let the horizontal recurrence be S_t = S_{t-1} - eta grad_S l(S_{t-1}; k, v) with a fixed (k, v), and let the vertical recurrence be x_{t+1} = f(x_t) for a single shared block f. Solve both recurrences in closed form for T = 2 and T = 3; if S_T cannot be expressed as f^T(x_0) for any choice of f, then the equivalence in Section 5.2 is not a mathematical identity and must be reframed as an open hypothesis. If the closed-form comparison is inconclusive, run the same check numerically with random linear maps and a fixed target (k, v).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The survey's central organizing move is to treat vertical (activation-based) and horizontal (hidden-state-based) recurrence as two implementations of the same latent reasoning resource. Section 5.2 states that 'processing a longer sequence is mathematically equivalent to running the same layer for more optimization iterations' (around Eq. 13). This is the load-bearing bridge of the taxonomy, and it is asserted without derivation or empirical support. In the actual update S_t = alpha_t S_{t-1} - eta_t grad_S l(S_{t-1}; k_t, v_t) (Eq. 13, Section 3.2.2), each new token supplies a new (k_t, v_t), so the objective being optimized changes from step to step. Vertical recurrence (Eq. 2) instead applies the same transformation f repeatedly to the same input representation. These are different dynamics: one is online learning with a moving target, the other is fixed-point iteration on a single objective. Nothing in the paper shows that the number of tokens is a valid proxy for the number of layer executions, or that one extra token equals one extra 'optimization iteration' of the same implicit layer. The survey itself concedes (Section 3.2) that hidden-state models 'have not yet produced evidence demonstrating enhanced reasoning capabilities,' which makes the claimed equivalence especially load-bearing: without it, the horizontal branch is not a demonstrated form of latent CoT, and the taxonomy is a classification rather than a unification. The bandwidth argument in Figure 1 is a related but separate weakness: raw FP16 bits are not a measure of usable reasoning capacity, and the caption's 'approx.' does not repair the inference.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This survey organizes the emerging literature on latent reasoning in large language models into a taxonomy with two principal branches: vertical recurrence (activation-based methods that loop over layers) and horizontal recurrence (hidden-state-based methods that propagate compressed states over time), extended by a third branch on diffusion-based infinite-depth reasoning. It proposes a unified mathematical framework in Section 2, an optimization-based equivalence between temporal and depth recurrence in Section 5.2, and a review of mechanistic interpretability evidence in Section 4. The survey is candid about the absence of direct empirical comparisons and includes critical discussion of Infini-attention's reported practical failures.","tokens_in":30253,"tokens_out":6299,"duration_ms":70169,"significance":"If taken as a map of an emerging field, the survey is timely and useful. It covers a wide range of recent work, provides a structured categorization, and includes welcome critical coverage (e.g., the Infini-attention reproduction issues in Section 5.2.1) and explicit limitation statements (Section 6). Its main contribution is the conceptual organization, not new empirical results. The central unification claim—that temporal and depth recurrence are equivalent—is, however, asserted rather than demonstrated, and the horizontal branch is admitted to lack evidence of enhanced reasoning. With that claim appropriately tempered, the survey would be a valuable resource; as written, the significance depends on an unsupported equivalence.","major_comments":[{"comment":"The claim that 'processing a longer sequence is mathematically equivalent to running the same layer for more optimization iterations' is not supported by the update rule in Eq. (13), where each new token supplies a fresh (k_t, v_t) and therefore a different objective; vertical recurrence in Eq. (2), by contrast, iterates the same transformation f on a refined representation. The paper provides no derivation or empirical test of this equivalence, and Section 3.2 itself concedes that hidden-state models 'have not yet produced evidence demonstrating enhanced reasoning capabilities.' This makes the unifying claim load-bearing and currently unsupported; I recommend either proving the equivalence under stated conditions, citing evidence for it, or recasting the taxonomy as a classification with open questions about whether horizontal recurrence constitutes latent reasoning.","section":"Section 5.2, Eq. (13)"},{"comment":"The optimization-based unification is overgeneralized. Only DeltaNet and Gated DeltaNet (and the gradient-state models in Table 2) have updates that are exactly gradient steps on a regression objective; the linear-state updates for Linear Attention, RetNet, GLA, Mamba-2, and HGRN-2 are associative decay/add rules that do not, in general, minimize a shared objective. The text claims a conceptual unification of temporal and depth recurrence based on DeltaNet, but the table itself shows that the linear-state family is not an optimizer. Please restrict the unification claim to gradient-state and delta-rule models, or show the objective for each update rule.","section":"Section 3.2.1, Table 2"},{"comment":"The survey labels hidden-state sequence models as a principal branch of latent reasoning despite admitting that these methods have not yet demonstrated enhanced reasoning capabilities (Section 3.2). This conflates memory/linear recurrence with reasoning. Similarly, Section 4.3's title promises Turing completeness of layer-based latent CoT, but the cited results concern Turing completeness of Transformers, RNNs, and prompting, not of latent CoT specifically. Please align the claims with the evidence, for example by explicitly distinguishing demonstrated latent reasoning from architectures that merely have potential for it.","section":"Section 3.2 and Section 4.3"}],"minor_comments":[{"comment":"The bandwidth comparison compares bits per token for explicit CoT with bits per hidden state for latent reasoning, but the two quantities are not matched in time granularity or information content; please state the assumptions and compare over a full reasoning step.","section":"Figure 1"},{"comment":"The indexing mixes time steps—x^{l+1}_{t+1} is computed from x^l_{t+1} and S^l_t; please define the time convention consistently so the spatial and temporal updates are unambiguous.","section":"Section 2.1, Eq. (1)"},{"comment":"The text says the same transformation f is applied to 'the same input,' but each call passes a different hidden state S^{l+i-1}_t; please rephrase as shared parameters with an evolving state.","section":"Section 3.1.1, Eq. (2)"},{"comment":"The critical assessment of Infini-attention relies on a blog post whose first author is listed as 'neuralink'; please verify the author attribution and, if possible, cite a more archival source or clearly mark the blog as non-peer-reviewed.","section":"Section 5.2.1, Ref. [73]"},{"comment":"DoT-SEDD and DoT-Plaid are both cited to Ref. [124] even though the text treats them as distinct frameworks; please clarify the relationship.","section":"Section 5.1.1 and 5.1.2, Ref. [124]"},{"comment":"The subsection title promises Turing completeness of layer-based latent CoT, while the cited results concern Transformers, RNNs, and prompting; consider retitling to avoid overstatement.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The survey's own-group masked diffusion citations (refs 74–76, 139) are prominent in Table 4 and Section 5.1; this is not fatal, but the editor may want to check citation balance. The paper also uses a Hugging Face blog post as a key critical source; that is appropriate as a reproduction report, but its non-archival status should be visible in the reference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this survey is worth reading and worth citing. The vertical/horizontal recurrence split and the Pre/Loop/Coda breakdown give the field a usable vocabulary, and Table 2's unification of linear-state and gradient-state memory updates is the most useful single page. The Infini-attention section is honest about a failed reproduction, and Section 6 admits there is no empirical comparison. Those are real merits.\n\nThe soft spots are real but addressable. The big one is the claim in Section 5.2 that processing a longer sequence is mathematically equivalent to running the same layer for more optimization iterations. In the actual update (Eq. 13), each token supplies a new (k_t, v_t), so the objective moves every step; that is online learning with a changing target, not fixed-point iteration on a single objective. Nothing in the paper shows that token count is a valid proxy for layer executions. The taxonomy doesn't fall apart without that bridge — it works fine as a classification — but the paper currently presents the equivalence as established when it is an open hypothesis. That should be flagged explicitly. The Figure 1 bandwidth ratio is also an illustration, not a measure of usable reasoning capacity; the 'approx.' doesn't fix the inference. And the survey never states its inclusion criteria, which matters because several diffusion papers anchoring the taxonomy come from the authors' own group.\n\nNone of this is fatal. The survey is candid, broad, and organized around distinctions that will actually help people navigate the literature. For a grad student entering latent reasoning, or a researcher who wants to place a new method in context, this is a good map. It deserves a serious referee. My recommendation: send it to review, and ask for three changes — weaken or prove the time-depth equivalence, qualify the bandwidth claim, and add an inclusion-criteria paragraph. With those, I'd be glad to see it out.","headline":"A genuinely useful taxonomy of latent reasoning, with a central unification claim that is currently asserted rather than shown.","tokens_in":30920,"tokens_out":2024,"would_cite":true,"duration_ms":24915,"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":"This survey claims that latent chain-of-thought — reasoning in continuous hidden states rather than words — can overcome the expressive limits of explicit CoT, and that all current methods share a depth-from-optimization principle.","keywords":["latent chain-of-thought","continuous hidden state","activation-based recurrence","hidden-state recurrence","masked diffusion models","test-time training","loop transformers","mechanistic interpretability"],"falsifier":"Run a matched comparison on a fixed reasoning task, such as multi-hop question answering or GSM8K-style arithmetic, with two models of identical parameter count: one that loops the same layer stack for T iterations and one that processes T hidden-state updates along the sequence. If doubling T improves only the looped model, or if the two models diverge in accuracy at equal T, the depth-time equivalence at the core of the survey's unification is false for real transformers.","tokens_in":29735,"feed_emoji":"🧠","tokens_out":14532,"duration_ms":127267,"temperature":0.7,"pith_summary":"Explicit chain-of-thought makes an LLM spell out each reasoning step in words, and that constraint costs expressive power: a token carries roughly 15 bits, while one 2560-dimensional hidden state carries about 40,000 bits. The survey argues that latent chain-of-thought, which keeps the reasoning process in the continuous hidden state via $z_{t+1} = \\mathrm{Transform}(z_t, S_t)$, can sidestep that bottleneck and may exceed what word-based reasoning can express. It organizes the field into vertical recurrence (looping the same layers to deepen computation), horizontal recurrence (evolving a compressed hidden state across a sequence), and infinite-depth reasoning via diffusion and optimizer-style updates. The organizing claim is that all three families are one phenomenon: depth emerges from optimization over time, so longer sequences and more loop iterations are the same resource. If this framing holds, non-linguistic reasoning paths and test-time scaling that does not generate more words become available to any model, not just new architectures.","feed_headline":"Thinking in hidden states can beat text chain-of-thought","feed_subtitle":"Looped, recurrent, and diffusion models share one recipe: depth from optimization over time.","key_machinery":"The central object is the hidden state $S^{l}_{t}$ together with the latent update rule $z_{t+1} = \\mathrm{Transform}(z_t, S_t)$ that replaces the explicit-CoT rule $y_{t+1} = \\mathrm{Decode}(\\mathrm{Transform}(x_t, S_t))$. The load-bearing identity of the survey is the gradient-state recurrence $S_t = S_{t-1} - \\eta_t \\nabla_S \\ell(S_{t-1}; k_t, v_t)$, which makes each incoming token one step of an online optimizer refining a fast-weight layer; DeltaNet's closed-form update is shown to be exactly one gradient step on a least-squares objective. Around this core, the survey places three mechanism families: vertical recurrence (the same layer stack run for multiple iterations), horizontal recurrence (matrix or vector state updated per token), and diffusion-based spatial refinement (masked or embedding-level denoising over the whole sequence). These mechanisms are connected by the claim that depth can be traded for time, making the hidden-state update the common substrate of latent reasoning.","core_discovery":"The paper's central claim is that reasoning need not be verbalized to be performed: a transformer can carry out multi-step inference entirely in continuous representations by eliminating the decoding step that forces each thought into a token. In the survey's unified formalism, every method is a spatial transformation across layers plus a temporal transformation of a hidden state $S^{l}_{t}$ that can be a key-value (KV) cache, a fixed-size linear-attention matrix, or a recurrent vector. Activation-based methods (Universal Transformer, looped and recursive transformers, Coconut) deepen the computational graph by reusing layers; hidden-state methods (linear attention, RetNet, Mamba-2, TTT, Titans, DeltaNet) evolve a compressed memory; and text diffusion models refine the whole sequence bidirectionally, with confidence-thresholded KV caches that make the process reversible and self-correcting. The survey further claims that these branches are unified by a gradient-based view in which the hidden state is a fast-weight layer updated by an optimizer, so processing a longer sequence is mathematically equivalent to running the same layer for more optimization iterations. The conclusion is that latent reasoning is a coherent design space with a common mathematical core, and that the field's main open problems are training stability, memory efficiency, and the absence of standardized benchmarks.","pith_inferences":["If the depth-time equivalence is real, then current linear-attention and state-space models should show measurable latent-reasoning gains on multi-step benchmarks, not just language modeling and recall; the survey notes that such evidence does not yet exist, so this is a testable prediction rather than a reported result.","The bandwidth argument implies a sharp interpretability trade-off: latent thoughts are unreadable, so safety and verification may require auxiliary probes or reconstructive decoders rather than inspection of the reasoning trace itself.","The same equivalence suggests that deployed reasoning models could replace explicit thinking tokens with silent hidden-state iterations, changing the cost and latency profile of test-time compute scaling.","One concrete extension: a controlled comparison of looped transformers and linear-recurrent models with matched parameters and step counts would settle whether vertical and horizontal recurrence are truly the same resource; the survey's taxonomy makes this experiment natural."],"forward_implications":["If latent chain-of-thought works as claimed, models can reason about problems that have no clean linguistic expression, because thoughts no longer have to pass through a finite vocabulary.","Long-context models become deep reasoners: feeding a model more tokens can serve as extra optimization steps, so sequence length and network depth become interchangeable reasoning resources.","Diffusion-based reasoning can revise earlier decisions, unlike autoregressive generation which commits to tokens and cannot revisit them, enabling globally consistent and self-correcting chains.","Existing pretrained transformers can acquire latent reasoning through training alone, for example with Coconut's continuous thought vectors, CODI's self-distillation, or pause tokens, without architectural changes, so the capability is accessible to current LLMs.","The unification predicts that performance on reasoning tasks scales with the number of hidden-state refinement steps, independent of whether those steps come from loops, recurrent states, or denoising iterations."],"supporting_citations":[{"why":"Universal Transformer, the foundational loop-based latent reasoning architecture with adaptive computation time.","marker":"[22]"},{"why":"Coconut, the first training-induced latent CoT method that loops the last hidden state back into the model.","marker":"[44]"},{"why":"DeltaNet's state update is shown to be exactly one gradient step, grounding the linear/gradient-state duality.","marker":"[123]"},{"why":"Introduces gradient-state recurrence by running SGD on the hidden state during inference.","marker":"[100]"},{"why":"Establishes that a K-layer transformer doing m-step CoT can be simulated by an (L+O(1))-layer transformer, supporting the view of layer depth as latent CoT.","marker":"[86]"},{"why":"Shows that log-depth transformers gain expressive power, making depth a primary bottleneck for latent reasoning.","marker":"[70]"},{"why":"Generalizes chain-of-thought to masked diffusion models, the survey's infinite-depth reasoning paradigm.","marker":"[124]"},{"why":"Shows implicit fixed-point RNNs recover general RNN expressivity and give another route to unbounded depth.","marker":"[87]"}],"fun_headline_variants":["Latent reasoning: no tokens, just hidden-state depth","All latent reasoning is optimization over time","Loops, memory, diffusion: one latent reasoning recipe","Why explicit CoT can be a bottleneck","From verbalized steps to hidden-state inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that depth and time are interchangeable: Section 5.2 asserts that processing a longer sequence is mathematically equivalent to running the same layer for more optimization iterations, but the survey does not prove this equivalence for real transformers, and the unification of vertical and horizontal recurrence rests on it.","fun_headline_variants_meta":{"raw":{"variants":["Latent reasoning: no tokens, just hidden-state depth","All latent reasoning is optimization over time","Loops, memory, diffusion: one latent reasoning recipe","Why explicit CoT can be a bottleneck","From verbalized steps to hidden-state inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00032,"raw_usage":{"total_tokens":1835,"prompt_tokens":1011,"completion_tokens":824,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":753}},"tokens_in":627,"tokens_out":824,"duration_ms":7725,"temperature":1.0,"reasoning_tokens":753,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:08:23.237145+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a matched comparison on a fixed reasoning task, such as multi-hop question answering or GSM8K-style arithmetic, with two models of identical parameter count: one that loops the same layer stack for T iterations and one that processes T hidden-state updates along the sequence. If doubling T improves only the looped model, or if the two models diverge in accuracy at equal T, the depth-time equivalence at the core of the survey's unification is false for real transformers.","supporting_citations":[],"review_version":1}