{"id":"9e8f9135-d785-43d6-9444-31b53c1b9fca","arxiv_id":"2607.27902","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Single anchor visual tokens chosen by reinforcement learning improve MLLM scene text spotting over multi-patch and coordinate-text grounding on Total-Text, CTW1500, and ICDAR 2015.","lead":"SPaTS makes an MLLM route each text instance through a single image-patch token selected by reinforcement learning, then decodes geometry from the full image. It reports large F1 gains over frontier MLLMs on three text-spotting benchmarks, but the paper's own appendix shows it still trails classical expert spotters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multi-Patch baseline is implausibly weak vs PaDT, undermining the controlled evidence for single-patch superiority","rationale":"The reader's weakest_assumption focuses on the patch-dominance reward's use of a fixed overlap threshold to define correct patches, arguing this may reinforce suboptimal anchors. This is a valid internal concern about whether the RL signal truly identifies the most informative patch. However, it does not directly threaten the paper's central comparative claim; even a suboptimal reward could yield a model that outperforms baselines. The more serious weakness is the Multi-Patch baseline in Table 2, which is the controlled evidence for 'One Patch Is Enough.' The paper describes Multi-Patch as following PaDT's pipeline, yet it scores far below PaDT's own reported numbers. Such a large deficit cannot be explained by architecture alone and suggests an implementation gap. If that baseline is not a faithful multi-patch method, the central claim loses its empirical foundation. The concrete test of re-running the baseline with PaDT's official implementation would settle this. The reader's rationale already calls for explicit comparison to PaDT, so our verdict remains CONDITIONAL (unchanged), but the reasoning shifts to the baseline validity issue rather than the reward proxy.","tokens_in":22531,"tokens_out":7560,"duration_ms":97826,"concrete_test":"Reproduce the Multi-Patch baseline using PaDT's official code/checkpoint fine-tuned on the same SFT data as SPaTS, with identical hyperparameters. Verify whether Total-Text F-measure reaches ≈30 (PaDT-3B level). If it does, Table 2's claimed single-patch advantage disappears or shrinks, invalidating the central claim. Alternatively, re-run Table 2 using PaDT-7B's reported numbers as the Multi-Patch baseline; if single-patch still wins, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2's Multi-Patch baseline is implausibly weak. The paper states (Appendix A.5) that Multi-Patch uses the same SPaTS architecture with five routed patches per instance following PaDT, trained under the same SFT setting. After SFT it scores F=21.6 on Total-Text, while Table 11 reports PaDT-3B and PaDT-7B at F=30.7 and 32.2 on the same benchmark. Since Multi-Patch is a strict superset of PaDT's routing (same backbone, more patches), it should achieve at least PaDT's numbers. A ~9–11 point gap suggests the Multi-Patch implementation was not held to the same standard (e.g., different patch sampling, token sequence, or temperature). If so, the headline conclusion 'Single-Patch selection consistently outperforms Multi-Patch formulations' rests on a strawman baseline, and the paper's central contribution — that one patch is enough — lacks controlled support. This is more load-bearing than the reward-proxy concern because it attacks the comparative result that motivates the paradigm, not just the optimization signal.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SPaTS, an MLLM-based scene text spotter that routes each text instance through a single anchor visual token and recovers geometry by full-image cross-attention refinement. To learn the anchor without oracle labels, the authors introduce SPaSO, a GRPO-based reinforcement-learning stage with two patch-level rewards (generation quality and top-k patch dominance), together with DEA, which decouples feature magnitude from direction, and PED, which fuses the routed patch with language hidden states. The method is evaluated on Total-Text, CTW1500, and ICDAR 2015 against closed-source MLLMs, open-source MLLMs, OCR-specialized MLLMs, and expert spotters, with ablations of the grounding formulation, reward components, and architectural modules. The claimed central finding is that a single patch, selected by RL and refined by full-image decoding, is a better grounding interface than coordinate text or multi-patch routing.","tokens_in":22826,"tokens_out":6272,"duration_ms":86058,"significance":"If the central comparison were fully supported, the single-anchor design would be a useful and parameter-efficient contribution to visual-token grounding for scene text spotting. The paper provides a concrete RL formulation for discrete visual-token selection, a reasonable architectural rationale for DEA and PED, extensive ablations, efficiency measurements, and a public code release. However, the advertised 'state-of-the-art' claim is contradicted by the paper's own tables, and the controlled evidence for 'one patch is enough' rests on a Multi-Patch baseline whose behavior is not yet explained. The conceptual idea is promising, but the evidence as presented is not yet at the level of the claims.","major_comments":[{"comment":"The central claim that SPaTS 'consistently and significantly outperforms both frontier closed-source MLLMs and OCR MLLMs' is contradicted by the paper's own tables. In Table 1, on CTW1500 'None', HunyuanOCR reaches F=63.8, while SPaTS-2B scores 47.0 and SPaTS-4B 51.2; on 'Full', HunyuanOCR (77.7) beats SPaTS-2B (75.1). Table 10 shows OmniParser ahead of both SPaTS variants on all reported metrics (e.g., Total-Text None 84.0 vs. 64.9; ICDAR 2015 S 89.6 vs. 73.4). The abstract and §4.2 should be rewritten to state the actual comparison scope, and the acknowledgment in Appendix B.1 that SPaTS trails the strongest specialist systems should be reflected in the main text. This is not merely wording: the advertised result is not the result the tables establish.","section":"Abstract; §4.2, Table 1; Appendix B.1, Table 10"},{"comment":"The controlled comparison for the central 'one patch is enough' claim depends on the Multi-Patch baseline. After SFT, Multi-Patch gives F=21.6 on Total-Text (Table 2). Table 11 reports PaDT-3B and PaDT-7B at F=30.7 and 32.2 on the same benchmark, and Appendix A.5 states that Multi-Patch uses the same SPaTS architecture with five routed patches per instance following PaDT under the same SFT setting. A correct implementation should not fall roughly 9–11 F points below PaDT. Please provide the PaDT-equivalent configuration, or an explanation of why DEA/PED, designed for single-patch routing, cause such a large drop when applied to multi-patch routing. Without this, Table 2's conclusion that Single-Patch consistently outperforms Multi-Patch is not a controlled comparison.","section":"Appendix A.5; Table 2; Table 11"},{"comment":"The Patch Dominance Reward uses a candidate set G_i defined as every grid cell whose overlap with the ground-truth polygon mask exceeds a fixed threshold. The threshold value is never given. Since Figure 6 demonstrates that peripheral or mismatched patches can cause immediate grounding degradation, the reward may reinforce boundary-overlapping but visually uninformative patches whenever the overlap proxy is not faithful to patch informativeness. Please specify the threshold, report sensitivity to it, and ideally validate the overlap proxy against an alternative (e.g., crop-based recognition accuracy). This is load-bearing for the claim that SPaSO learns to select the most informative patch without oracle labels.","section":"§3.2, Eq. (9); Appendix A.6"}],"minor_comments":[{"comment":"Table 4 reports CTW1500 Patch F=81.2, whereas Table 1 reports SPaTS-4B CTW1500 Full F=79.6. Appendix A.7 says different matching protocols are used for format-specific ablations, but the main text does not say which stage or protocol Table 4 uses. Please clarify.","section":"Table 4 vs. Table 1"},{"comment":"The table mixes two factor groups (L2 Norm/Scale/DEA and h_t/p_att/PED) in a way that makes it unclear whether the lower-block rows also include DEA. Please separate the two ablations explicitly.","section":"Table 5"},{"comment":"The Total-Text F=63.7 in Table 2 for Single-Patch is identical to the SPaTS-4B row in Table 11. Readers cannot tell which model variant Table 2 uses. Please state the model size and training stage in the caption.","section":"Tables 2 and 11"},{"comment":"The edit-distance threshold in r_gen is reported as 0.2, but the normalization (character-level? word-level? normalized by length?) is not defined. Please specify the exact metric so the reward is reproducible.","section":"Table 8"},{"comment":"The qualitative claim that patch selection drives grounding quality would be stronger with a quantitative plot (e.g., selection distance to GT center vs. final F/IoU) over the validation set.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper contains strong ideas and a fair amount of useful ablation, but the main text's sweeping claims overstate what the tables show. The Multi-Patch baseline issue is the most important technical point: it directly affects the central 'single patch is enough' claim, and the authors need to either re-run the comparison in a PaDT-equivalent setting or provide a convincing analysis of why their multi-patch variant is so much weaker than PaDT. I would not reject the paper, but the claims must be corrected and the controlled comparison strengthened before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the core idea is real. Routing each text instance through one anchor patch and recovering geometry with full-image cross-attention is a clean departure from the multi-patch routing in PaDT and ClawMachine, and SPaSO's patch-level RL with the generation and dominance rewards is a sensible way to avoid oracle patch labels. DEA and PED are not deep, but they are well-motivated and the ablations are mostly internally consistent. Table 11 shows the single-patch formulation at least merits a serious look.\n\nThe soft spots are in the framing and in one baseline. The abstract says SPaTS \"consistently and significantly outperforms\" OCR MLLMs, but Table 1 shows HunyuanOCR beating SPaTS on CTW1500 by wide margins (63.8 vs 47.0 on None, 77.7 vs 75.1 on Full). The contribution bullet claims state-of-the-art, yet Table 10 shows OmniParser beats SPaTS by 20+ F points on Total-Text and CTW1500. The appendix concedes this, but the headline and the contribution list do not.\n\nThe stress-test note about Multi-Patch is on target. Table 2's Multi-Patch gets F=21.6 on Total-Text after SFT; Table 11 gives PaDT-3B 30.7 and PaDT-7B 32.2 on the same protocol. The paper says Multi-Patch uses the same SPaTS architecture with five routed patches per instance and the same SFT setting. A strict superset should not be 9–11 points worse than the method it claims to follow. That gap almost certainly reflects an implementation mismatch (patch sampling, token sequence, or prompt), not the routing count. If so, the central empirical claim — that single-patch beats multi-patch — does not have controlled support. The missing overlap threshold for the candidate set G_i in Appendix A.6 is a lesser but real issue: the RL reward treats any GT-overlapping patch as correct, while Figure 6 shows that peripheral patches degrade grounding. The threshold is never specified, so the reward coverage is under-validated.\n\nOverall: a paper with a good idea and a clean method, published with comparisons that do not currently back the headline claims. It needs a redo of the baselines with a uniform prompt protocol, a faithful Multi-Patch implementation (or direct use of PaDT's official code), error bars, and removal of the unqualified SOTA language. I would send it to peer review — the single-patch idea deserves referee time — but I would not trust the reported numbers until the baselines are fixed.","headline":"Genuine single-patch grounding idea with a clean RL formulation, but the headline claims overreach and the Multi-Patch baseline is likely undercooked.","tokens_in":23364,"tokens_out":3084,"would_cite":false,"duration_ms":42114,"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":"A single, reinforcement-selected visual patch per text instance, refined against the full image, is the most accurate grounding interface for MLLM-based scene text spotting.","keywords":["Scene Text Spotting","Multimodal Large Language Models","Visual Token Grounding","Single-Patch Selection","Reinforcement Learning","Visual Patch Decoding","Directional Embedding Alignment"],"falsifier":"Train the Multi-Patch formulation with the same SPaSO rewards and the same full-image decoder used in Single-Patch, then compare end-to-end F-measure on Total-Text, CTW1500, and ICDAR 2015; if the multi-patch policy matches or beats the single-patch policy, the central claim collapses.","tokens_in":22405,"feed_emoji":"📝","tokens_out":6370,"duration_ms":85675,"temperature":0.7,"pith_summary":"This paper claims that the noise and ambiguity in MLLM-based scene text spotting come largely from grounding on multiple image patches at once, and that routing each text instance through a single anchor visual token—then recovering geometry by attending to the whole image—is both more accurate and more reliable. To find that anchor without extra labels, the authors train the model with reinforcement learning rewards that measure final spotting quality, and they stabilize the selected patch features by separating feature direction from magnitude. If correct, this means a lightweight 2B or 4B model can outperform far larger general-purpose MLLMs and specialized OCR systems on text spotting benchmarks, shifting the bottleneck from model scale to the grounding format.","feed_headline":"One patch is enough: single-anchor grounding wins text spotting","feed_subtitle":"A 2B/4B model with one routed visual token and full-image refinement beats far larger MLLMs on curved and dense text.","key_machinery":"The central machinery is the learned single-patch selector under the SPaSO objective: the model autoregressively emits an index into the image's patch-token vocabulary for each text instance, and the policy is optimized with two patch-level rewards — a generation reward that measures end-to-end spotting quality and a dominance reward that keeps ground-truth patches in the top-k. DEA ensures that the visual-token logits reflect directional alignment rather than raw magnitude, and PED turns the selected token into an object query that reads from the full image to decode boxes, Bézier curves, points, and scores. The single anchor is therefore the point where recognition and localization are cou","core_discovery":"SPaTS is built on the claim that the correct granularity for visual grounding in scene text spotting is one anchor image patch per text instance, not a set of patches or a set of coordinates. The paper's central discovery is that this single-patch interface is learnable without oracle labels: SPaSO formulates patch selection as a discrete reinforcement-learning problem, rewarding both the final spotting outcome (a harmonic mean of F1 and average IoU) and the presence of a ground-truth-overlapping patch among the top-k predictions. Two architectural pieces make the single patch usable: DEA parameterizes each visual prototype as a normalized direction times a learned scale, so token selection","pith_inferences":["If single-anchor grounding is as general as the paper suggests, the same routing-plus-full-image-refinement recipe should transfer to other dense multimodal tasks, such as document key-information extraction or referring-expression grounding, where multi-patch aggregation is currently the default.","A label-free extension would be to learn the patch-selection reward from self-consistency instead of ground-truth polygon overlap: sample several candidate patches, decode geometry from each, and reward patches whose transcriptions and boxes agree; this would remove the need for a fixed overlap threshold.","A natural stress test is to evaluate SPaTS on images with severe occlusion or specular highlights, where the most informative patch may not overlap the ground-truth polygon; if the magnitude branch is frozen after warmup, feature norms under extreme illumination could become a failure mode that current benchmarks do not expose."],"forward_implications":["Multi-patch grounding, a common approach in patch-based MLLM spotting, is actively harmful for scene text: a single routed patch yields higher precision, recall, and F-measure on the tested benchmarks.","Discrete patch selection can be trained end-to-end with task-level rewards, so no oracle best-patch annotation is needed; the model discovers which patch is informative from spotting quality alone.","Decoupling feature magnitude from direction in the unified head is a transferable mechanism: it can prevent high-norm patches from dominating any visual-vocabulary selection task, not just text spotting.","The reported results imply that a 2B or 4B MLLM with a well-designed grounding interface can beat much larger general MLLMs and specialized OCR models on end-to-end spotting, making the grounding format a primary design lever."],"fun_headline_variants":["One patch per text instance beats multi-patch MLLMs","SPaTS: reinforcement-optimized token grounding for text spotting","Single patch, full refinement: scene text spotting's new edge","Patch-level RL hones MLLM text spotting to one token","SPaTS condenses text spotting to a single anchor patch"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that overlap with the ground-truth polygon mask is a faithful proxy for how informative an image patch is for recognition and localization; if the reward reinforces patches that only skim the polygon but are visually uninformative, the selection policy can settle on suboptimal anchors.","fun_headline_variants_meta":{"raw":{"variants":["One patch per text instance beats multi-patch MLLMs","SPaTS: reinforcement-optimized token grounding for text spotting","Single patch, full refinement: scene text spotting's new edge","Patch-level RL hones MLLM text spotting to one token","SPaTS condenses text spotting to a single anchor patch"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000328,"raw_usage":{"total_tokens":1681,"prompt_tokens":766,"completion_tokens":915,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":837}},"tokens_in":510,"tokens_out":915,"duration_ms":12071,"temperature":1.0,"reasoning_tokens":837,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T03:15:58.901799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the Multi-Patch formulation with the same SPaSO rewards and the same full-image decoder used in Single-Patch, then compare end-to-end F-measure on Total-Text, CTW1500, and ICDAR 2015; if the multi-patch policy matches or beats the single-patch policy, the central claim collapses.","supporting_citations":[],"review_version":2}