{"id":"b4536668-31d5-4c79-a495-d19676ca2eb6","arxiv_id":"2607.18856","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A compact, window-level thought state can be expanded into dense queries that refine super-resolution outputs over four shared ticks, but the V1 report's numbers are inconsistent and its implementation is undisclosed.","lead":"ThinkSR stretches the Continuous Thought Machine idea to image super-resolution: shared 'thought ticks' refine local image windows and emit progressively sharper estimates. The paper is an honest V1 feasibility report, but its headline numbers are internally inconsistent and no code or implementation details are provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unspecified DQ-CTM update in §2.3 leaves the central feasibility claim unfalsifiable; an ablation that severs the thought state from the dense update would settle whether the progressive trajectory is due to CTM dynamics or to generic shared residual refinement.","rationale":"The reader's weakest assumption — that the unspecified DQ-CTM update could be equivalent to a standard residual MLP/attention — matches the most load-bearing concern. The paper's own boundary statements (§3.5, §4.3) concede that matched recurrent baselines and evaluation details are still pending, so the central empirical claim is not independently supported. The numeric inconsistency between the abstract and Table 2 is real but secondary: if the implementation evidence were supplied, re-evaluation would resolve it; without that evidence, the architecture-level concern alone justifies REJECT. My recommendation is UNCHANGED because the reader's verdict already captures the evidentiary gap, and the proposed ablation would transform the unfalsifiable claim into a testable one.","tokens_in":6529,"tokens_out":5087,"duration_ms":45443,"concrete_test":"Obtain the exact Uθ implementation and trained checkpoint, then run a single control: freeze s(t) and h(t) at their T=0 values for all ticks while keeping Uθ, encoder, decoder and the shared-parameter unrolling unchanged. If the T=4 PSNR-Y stays within 0.1 dB of the reported 30.47 dB and monotone improvement from T=1 to T=4 persists, the progressive trajectory does not depend on the temporal thought state, and the feasibility claim reduces to generic shared residual refinement. If the PSNR-Y drops materially or monotonicity breaks, the thought state is load-bearing and the concern is rebutted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a compact DQ-CTM thought state, not merely a shared residual refinement loop, produces the reported progressive SR reconstruction. The manuscript intentionally withholds the actual update: §2.3 calls Eqs. (2)–(3) \"the intended interface rather than every operation\" and says the factorization, intermediate dimensions, coefficient construction, and update order are \"intentionally omitted.\" This makes the central claim unfalsifiable: any token-count-preserving shared update (two-layer MLP, per-token cross-attention, SwinIR stage) satisfies the stated equations and could yield a similar monotone trajectory if the encoder/decoder and residual structure do most of the work. The authors themselves acknowledge in §3.5 that \"parameter-matched recurrent baselines\" and \"verified evaluation details\" remain ongoing, and §4.3 that \"recurrent comparisons must also be completed\" before strong empirical conclusions are appropriate. Without a matched recurrent baseline or a demonstration that the synchronization state s(t) actually modulates X(t), the feasibility of sparse latent thought for dense vision is not established. The reader's REJECT is warranted; the defect is missing implementation evidence, not an internal contradiction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a continuous-thought extension to dense prediction, called Dense-Query Continuous Thought Machine (DQ-CTM) and its super-resolution instantiation ThinkSR. The idea is to maintain a persistent dense feature field while a compact, shared-parameter thought state evolves over discrete ticks and modulates the field via a low-rank, position-aligned query mechanism. The authors report a progressive reconstruction trajectory on 100 DIV2K validation images (T=0..4), with PSNR/SSIM improving monotonically, and position the work as a feasibility study for sparse latent thought in dense vision. The paper is explicit that it is a preliminary V1 report and does not claim adaptive stopping or state-of-the-art performance.","tokens_in":6949,"tokens_out":6097,"duration_ms":58002,"significance":"If the central claim were substantiated, the paper would provide a novel interface for continuous-thought models in dense prediction, with potential implications for adaptive or test-time refinement in image restoration. The formulation of the sparse-thought–dense-output mismatch is clear, and the decision to supervise only the final tick is a methodological strength: the reported progression is not directly enforced by a per-tick loss. The authors also deserve credit for openly stating limitations and for not overclaiming beyond the trained horizon. However, in its current form the paper is closer to a research proposal than a complete empirical study. The defining update in §2.3 is intentionally unspecified, the quantitative results are internally inconsistent, and no matched recurrent baseline is provided. These issues block an assessment of the central feasibility claim.","major_comments":[{"comment":"The central claim—that a compact CTM state can govern dense reconstruction—is unfalsifiable as presented. Equations (2)-(3) define only an interface: Tθ is a compact transition and Uθ is any token-count-preserving update. The text states that the factorization, intermediate dimensions, coefficient construction, and update order are 'intentionally omitted,' and §3.5 defers 'implementation-level interface verification.' Under this specification, any shared-parameter residual block (a two-layer MLP, per-token cross-attention, or a SwinIR stage) satisfies Eq. (3), and the progressive trajectory could reflect generic iterative refinement rather than CTM thought dynamics. To support the abstract's feasibility conclusion, the paper must either specify Uθ and the construction of s(t), or provide a control that severs the thought state from the dense update (e.g., replacing s(t) with a fixed or r","section":"§2.3, Eqs. (2)-(3)"},{"comment":"The quantitative record is internally inconsistent, which undermines the headline empirical claim. The abstract reports PSNR-Y 28.1045→30.2817 dB, PSNR-RGB 26.6271→28.7781 dB, and mean ℓ1 0.034602→0.023545; Table 2 reports T=0: PSNR-Y 28.91, PSNR-RGB 25.45, ℓ1 0.0386; T=4: PSNR-Y 30.47, PSNR-RGB 28.98, ℓ1 0.0229. Section 3.3 further states that 'PSNR-Y increases by 1.4611 dB' from T=1 to T=4 and that the tick-wise increment 'decreases from 0.6928 dB between T=1 and T=2 to 0.2025 dB between T=3 and T=4.' Table 2's increments are 0.30, 0.20, and 0.09 dB over T=1→2, T=2→3, T=3→4, with a total T=1→4 increase of 0.59 dB. These are not minor rounding differences; they are different numbers. The monotone-improvement claim and the 'all 100 images improve' statement cannot be evaluated until the numbers are reconciled and the evaluation protocol is identical.","section":"Abstract vs. Table 2 vs. §3.3"},{"comment":"No matched recurrent or iterative-refinement baseline is provided. The paper acknowledges this: §3.5 lists 'Parameter-matched recurrent baselines' as ongoing, and §4.3 states that 'recurrent comparisons must also be completed before strong empirical conclusions are appropriate.' This is load-bearing because the progressive trajectory may be a generic property of shared-weight unrolling, not of the thought state. A matched control—e.g., a DRCN-style recursive network or a shared-weight residual network with the same parameter count and no synchronization state—is necessary to attribute the improvements to CTM dynamics. Without such a baseline, the feasibility claim for 'sparse latent thought' is not established.","section":"§3.5, §4.3"},{"comment":"The experimental evidence is a single run with no error bars, no multiple seeds, and no released code or artifacts. Section 3.1 reports one trained model ('1.13M parameters, Adam, Ttrain=4') and a single evaluation sweep. Section 3.5 lists 'verified evaluation details' as ongoing. For a paper whose central claim is empirical feasibility, this is insufficient to support statements such as 'All 100 evaluated images improve.' The authors should provide standard deviations across at least three seeds, or release code and the exact evaluation script. Otherwise the reported numbers cannot be independently checked and the monotone improvement could be within run-to-run variation.","section":"§3.1, §3.5"}],"minor_comments":[{"comment":"The caption says outputs from T=0 to T=4 are shown, but the displayed panels appear to start at T=1. Please align the caption with the actual figure.","section":"§3.3 / Fig. 4 caption"},{"comment":"The loss notation is malformed; use a proper ℓ1 norm, e.g., L = || D(W^{-1}(X(4))) - I_HR ||_1.","section":"Eq. (5)"},{"comment":"The 'SwinIR-style encoder' is not specified architecturally (number of blocks, heads, window size, etc.), which makes the 1.13M parameter count non-reproducible. Please provide the full configuration or point to the exact code version.","section":"§2.2 / Table 1"},{"comment":"The note says 'evaluated using four thought steps' while the text uses T=0..4 and Ttrain=4. Clarify whether 'four thought steps' means T=4 or four shared updates after T=0, and be consistent with the tick indexing.","section":"Table 1 note"},{"comment":"The 'relative attention' visualization is not defined. If it is derived from the model's attention maps, specify how and acknowledge that it is not causal; the current caption is too vague.","section":"Supplementary Fig. S1"},{"comment":"The repeated use of 'V1 report' and 'preliminary' is appropriate for an arXiv preprint, but for a journal submission the methodology must be presented as complete. Consider removing the 'V1' framing and describing the status of the work once the missing details are added.","section":"§5 / throughout"},{"comment":"The claim 'parameter-efficient' appears several times, but no FLOPs or parameter breakdown for the DQ-CTM module is provided. Please report the module-level parameter count and computational cost to support that term.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: This is an unusual submission: a deliberately incomplete 'V1' report whose central mechanism is left unspecified. The idea is timely and the problem formulation is interesting, but the current manuscript does not yet support its feasibility claim. I recommend major revision rather than outright rejection because the missing implementation details, consistent numbers, and matched baselines could in principle be supplied in a revised version. However, if the author intends the omission of the DQ-CTM update to be permanent (e.g., a position paper), then the central claim remains unfalsifiable and the paper would not be suitable for publication in a serious journal. I would also check the internal numerical inconsistencies carefully in any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know about this paper: it's a short V1 that proposes a reasonable way to adapt Continuous Thought Machines to dense prediction, but the central mechanism is intentionally left unspecified and the numbers inside the paper disagree with each other. I'd treat it as a position note, not an empirical result.\n\nThe paper is at its strongest when it frames the problem. The 'sparse-thought–dense-output mismatch' is a real challenge, and the idea of using a compact window summary to generate position-specific queries without collapsing the spatial carrier is a sensible route. The authors are appropriately cautious: they call it V1, don't claim SOTA, and explicitly say no extrapolation beyond the trained horizon. That level of self-awareness is rare and worth crediting.\n\nThe problems start where the evidence lives. Section 2.3 says the compact-to-dense update is an 'intended interface' and that factorization, intermediate dimensions, coefficient construction, and update order are 'intentionally omitted.' That makes the core claim unfalsifiable: any shared residual or cross-attention step would satisfy the stated equations. Without a matched recurrent baseline, you can't attribute the monotone trajectory to CTM thought dynamics rather than generic iterative refinement. The authors acknowledge in §3.5 that 'parameter-matched recurrent baselines' are ongoing and in §4.3 that 'recurrent comparisons must also be completed' — so this gap is on record, but it isn't filled.\n\nThe numbers are a separate concern. The abstract reports T=0 PSNR-Y 28.1045 and L1 0.034602; Table 2 has 28.91 and 0.0386. The abstract T=4 PSNR-Y is 30.2817; Table 2 has 30.47. And the text says the T=1→2 PSNR-Y gain is 0.6928 dB, while Table 2 gives 0.30 dB. Even allowing for different metric computations, these are not consistent with each other in a single manuscript. With no code/checkpoints, the reported trajectory isn't checkable.\n\nWho is this for? Someone working on CTM, iterative refinement, or adaptive computation in low-level vision might want to see the idea, but they should not rely on the numbers. As a reviewer, I'd send it back asking for the missing implementation details, a matched recurrent baseline, and corrected tables. In its current form, I wouldn't send it to peer review as a complete paper — the evidence base isn't there. If the author provides code and fixes the reporting, the framing is worth revisiting.","headline":"A promising framing with an omitted core: the DQ-CTM update is unfalsifiable and the headline numbers don't match — as a V1 it's honest but not yet a verifiable result.","tokens_in":7309,"tokens_out":4319,"would_cite":false,"duration_ms":38750,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A compact thought state converts into per-position queries and progressively refines super-resolved images over four shared ticks.","keywords":["Continuous Thought Machine","dense prediction","image super-resolution","iterative refinement","Dense-Query CTM","sparse-thought–dense-output mismatch","ThinkSR","test-time computation"],"falsifier":"Run the reported validation sweep and check two things: (a) whether PSNR-Y at T=0 is 28.10 dB (abstract) or 28.91 dB (Table 2), and (b) whether replacing the compact thought state with a fixed vector at every tick preserves the trajectory; either check can falsify the feasibility claim.","tokens_in":6446,"feed_emoji":"🖼️","tokens_out":3884,"duration_ms":39498,"temperature":0.7,"pith_summary":"This paper asks whether a compact, temporal 'thought' process originally designed for low-dimensional decisions can coordinate a dense, location-indexed prediction such as image super-resolution. It proposes DQ-CTM, a low-rank interface that turns each window's compact thought summary into distinct queries for every spatial position, while a persistent dense feature carrier keeps local evidence intact. In the ThinkSR instantiation, shared thought parameters are unrolled for four ticks, and reconstruction quality improves monotonically on all 100 DIV2K validation images within this trained horizon. The paper's claim is feasibility: sparse latent thought can govern dense spatial reconstruction without collapsing spatial evidence, and the evolving estimate is measurable at every tick.","feed_headline":"Sparse thought state sharpens images over four ticks","feed_subtitle":"A compact window summary becomes per-pixel queries, and all 100 test images improve at every shared refinement step.","key_machinery":"The Dense-Query Continuous Thought Machine (DQ-CTM): a structured, low-rank, parameter-efficient compact-to-dense query mechanism that maps each window-level CTM summary into distinct queries aligned with every position in the window. It works alongside a persistent dense visual carrier — a token-count-preserving feature field — and shared thought dynamics, so that the compact state reads from and modulates dense tokens without collapsing spatial identities. Equations (2)–(3) define the intended interface: a compact thought transition and a dense, token-count-preserving update with shared parameters across ticks.","core_discovery":"The central claim is that the sparse-thought–dense-output mismatch is surmountable: a compact continuous-thought representation, derived from neuron-level histories and synchronization, can be expanded by a structured low-rank projection into window-aligned dense queries, each position receiving its own query while the thought parameters are shared across ticks. In ThinkSR, this drives a progressive reconstruction trajectory — PSNR-Y rises from about 28.91 dB at T=0 to 30.47 dB at T=4 on DIV2K validation (the abstract quotes 28.1045 to 30.2817), with all 100 images improving from T=1 to T=4. The author frames the result as establishing feasibility within the learned four-tick horizon, not as","pith_inferences":["If the dense update's functional role is implemented exactly as a low-rank expansion of the thought state rather than an off-the-shelf residual block, the progressive gains suggest a new axis of scaling: extending the thought horizon at test time could refine estimates further, though the paper explicitly disclaims this.","A parameter-matched recurrent baseline that shares weights but omits the compact thought state would isolate whether the gains come from CTM dynamics or generic iterative refinement; this is the paper's own stated next step.","The monotone per-image improvement hints that early ticks may act as a coarse-to-fine prior; if true, this could transfer to other ill-posed inverse problems where a stable coarse estimate aids convergence.","The unstated implementation details are the main risk: if the low-rank query projection is equivalent to a standard per-token linear layer, the 'thought' framing may reduce to iterative refinement with a different name."],"forward_implications":["If the interface is real, continuous-thought mechanisms can be applied to dense prediction tasks without first pooling visual evidence into a single global descriptor.","Shared thought parameters across ticks mean the model exposes a measurable internal reconstruction trajectory, not just a final output, enabling inspection of how estimates change.","The progressive refinement within the four-tick horizon suggests an implicit curriculum where early ticks produce coarse estimates and later ticks refine detail.","The DQ-CTM interface is architecturally agnostic to the output modality; the same persistent-carrier plus compact-thought separation could in principle carry segmentation logits, depth, or flow.","Because all 100 validation images improve from T=1 to T=4, the trajectory is not an artifact of a few favorable samples."],"fun_headline_variants":["Sparse thoughts sharpen images in four ticks","Continuous thought ticks improve every test image","Think sparse, predict dense: super-resolution via thought","Progressive thought refinement boosts image quality"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the concrete DQ-CTM update really implements the described compact-to-dense functional interface; the paper intentionally omits implementation details, so if the dense update is just a standard residual or attention block, the reported trajectory reflects generic iterative refinement rather than continuous-thought dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Sparse thoughts sharpen images in four ticks","Continuous thought ticks improve every test image","Think sparse, predict dense: super-resolution via thought","Progressive thought refinement boosts image quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1840,"prompt_tokens":832,"completion_tokens":1008,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":953}},"tokens_in":576,"tokens_out":1008,"duration_ms":8124,"temperature":1.0,"reasoning_tokens":953,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:06:35.311388+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the reported validation sweep and check two things: (a) whether PSNR-Y at T=0 is 28.10 dB (abstract) or 28.91 dB (Table 2), and (b) whether replacing the compact thought state with a fixed vector at every tick preserves the trajectory; either check can falsify the feasibility claim.","supporting_citations":[],"review_version":1}