{"id":"3a9323a8-18f2-4919-bd3a-4765f2f8a91a","arxiv_id":"2507.00454","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A visual-language tracker with fine-grained attribute decomposition and a cross-modal token achieves top results on TNL2K, but its ablation data is internally inconsistent.","lead":"ATSTrack is a visual-language tracker that splits language descriptions into categories, appearance, action, and location, and uses a memory token to guide visual extraction. It reports state-of-the-art tracking on TNL2K and competitive results on LaSOT, but the paper contains several internal inconsistencies and missing implementation details.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core mechanism depends on an unspecified LLM attribute decomposition; without evidence that the segmentation is accurate, the reported gains from fine-grained alignment cannot be attributed to the proposed temporal/spatial alignment.","rationale":"The reader's weakest_assumption identifies the same load-bearing point: the framework depends on decomposing every language description into four fixed attributes using an unspecified LLM, and if that decomposition is noisy or inconsistent, fine-grained alignment collapses to standard cross-attention. I agree with this assessment. I considered alternative concerns, such as the contradiction in Table 2c (text says α=25 is best while the table shows α=50) and the OTBlang overclaim (the table lists ATSTrack as second, not first, among VL trackers), but those are claim-wording and reproducibility issues that can be repaired editorially. The LLM decomposition is the one assumption that, if false, would invalidate the mechanism itself rather than just the wording of a contribution sentence. A concrete check is available: compare the paper's LLM segmentation against human annotation and re-run the TNL2K evaluation under a fixed parser. Because the reader already made the paper's acceptance conditional on addressing precisely this missing component, my stress-test does not move the verdict; it sharpens the condition by requiring a segmentation-quality check in addition to releasing the LLM name and prompt.","tokens_in":11635,"tokens_out":10222,"duration_ms":124252,"concrete_test":"Run a controlled TNL2K experiment with three segmentation conditions: (i) the exact LLM and prompt used in the paper, (ii) a fixed rule-based parser over the raw sentence, and (iii) human-annotated attribute spans for the same 200 randomly sampled descriptions, with two annotators and reported inter-annotator agreement. Compare the resulting TNL2K AUCs. If conditions (i) and (ii) are within noise while (iii) differs, the LLM is not the source of the fine-grained benefit. If (i) is required to reproduce the reported 66.2 AUC, the paper must specify the LLM and release the segmentations; otherwise the reported results are unreproducible and the fine-grained alignment claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is that each language description is split into category, appearance, action, and location phrases, and each attribute is routed to a different visual input in the FGM module (§3.2–§3.3). This routing is what supposedly realizes 'aligning temporal and spatial scales.' But the LLM that performs the split is never specified: §4.1 states only 'we use LLM to segment the language descriptions into different attributes,' with no model name, prompt, temperature, or post-processing, and no evaluation of segmentation quality. This omission is load-bearing rather than cosmetic. If the LLM misassigns phrases across the four slots, then the VFM module will suppress the wrong visual regions, the LFA module will gate on the wrong language tokens, and the action feature will attend to the wrong template sequence. The architecture then degenerates to an ordinary coarse-grained cross-attention tracker, and the Table 2a ablation gap (w/ FGM 72.0 vs w/o FGM 71.1 on LaSOT) cannot be credited to fine-grained alignment. The paper provides no evidence that the decomposition is reliable: Table 3 removes entire attributes, but removing an attribute does not test whether the LLM put the right words into that attribute. Consequently, the central claim that ATSTrack outperforms SOTA VL trackers on three datasets depends on an undocumented component whose failure would remove the proposed mechanism's advantage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ATSTrack, a vision-language tracker that addresses temporal and spatial scale misalignment between visual and language inputs. Language descriptions are decomposed by an unspecified LLM into four attributes (category, appearance, action, location), each routed to different visual inputs in a Fine-Grained Modification module: category suppresses background in the template, action attends over a template sequence, location is gated against the search region, and appearance is refined by the modified template. A Visual-Language token, formed by concatenating the average of modified language features with the visual CLS token, is fed into the visual backbone of the next frame. Experiments on TNL2K, LaSOT, and OTBlang report an AUC of 66.2 on TNL2K (3.1 points above prior language trackers), 72.6 on LaSOT, and 71.0 on OTBlang (second best). Ablations on LaSOT are presented for FGM components, gating weight α, VL-token components, and attribute removal. The stated contributions claim state-of-the-art performance on three tracking datasets.","tokens_in":11908,"tokens_out":2124,"duration_ms":20203,"significance":"If the results hold, ATSTrack would be a strong empirical contribution to vision-language tracking: the TNL2K improvement (66.2 AUC vs. 63.1 for UVLTrack-B) is substantial, and the LaSOT number is competitive with strong visual-only trackers. The conceptual framing of attribute-level temporal/spatial correspondence is appealing and leads to a modular architecture (FGM, VL token) that is easy to understand. The paper also includes several useful sanity ablations (attribute removal, token modality removal) and states that code will be released. The main significance is contingent on reproducibility of the LLM-based attribute decomposition and on the soundness of the ablation/hyperparameter choices, as detailed below.","major_comments":[{"comment":"The central mechanism depends on decomposing each language description into four attributes, but the LLM used for this decomposition is never specified. §4.1 states only that 'we use LLM to segment the language descriptions into different attributes,' with no model name, prompt, temperature, or post-processing, and no evaluation of segmentation quality. This omission is load-bearing because the FGM routing, the LFA gating, and the VL token all operate on this decomposition. If the LLM misassigns phrases, the reported gains from fine-grained alignment cannot be attributed to the proposed temporal/spatial alignment mechanism. The authors should specify the LLM and prompt, provide segmentation examples with failure cases, and ideally report a human-evaluated accuracy on a sample of the TNL2K and LaSOT training descriptions.","section":"§4.1 and §3.2–3.3"},{"comment":"The gating weight α is selected on the LaSOT benchmark (α=50 is reported in §3.3, while Table 2c's caption states the best value is α=25, and the text says 'LFA achieves the best performance with α = 25'), and the same LaSOT benchmark is then used for the final state-of-the-art comparison in Table 1. This is test-set fitting in a mild form: the final LaSOT AUC (72.6) is reported after choosing a hyperparameter on the same benchmark. The authors should either (a) report a validation split or cross-validation procedure for choosing α, (b) use a different benchmark or validation set for hyperparameter selection, or (c) at minimum show that the final LaSOT result is stable across a range of α values and report the sensitivity honestly.","section":"§4.3, Table 2c"},{"comment":"There is a direct internal contradiction: Table 2c reports the best AUC at α=50 (72.0), but the text states that 'LFA achieves the best performance with α = 25.' Moreover, §3.3 sets α=50 in the method description, which is consistent with Table 2c but inconsistent with the text. This contradiction must be resolved: the authors must correct the text or the table and state unambiguously which α is used in the final model.","section":"§4.3, Table 2c and text"},{"comment":"The ablation narrative for the Visual-Language token contradicts its own table. The text says 'Without using the Visual-language token (w/o token), the model decreases in the AUC score by 0.6%,' which matches the 72.0→72.6 improvement. However, it then says 'Using the Language token independently (w/o V token) leads to a decrease in the AUC score by 0.3%,' while Table 2b lists w/o V token at 71.7, which is a 0.9-point drop from the full model (72.6), not 0.3. Table 2b also shows w/o L token at 72.0 (0.6 below full), while the text says 'Using the visual token independently (w/o L token) does not lead to notable changes.' The text and table should be made consistent, and the conclusions about the contribution of each token should be re-derived from the correct numbers.","section":"§4.3, Table 2b and text"}],"minor_comments":[{"comment":"The gating formula uses a scaling factor α with α=50 in the main text, but the ablation section reports α=25 and α=50 inconsistencies as noted in the major comments; the notation should be consistent and the chosen value should be stated once in the method.","section":"§3.3, Eq. for LFA"},{"comment":"The caption says 'LaSOText' but the table columns list LaSOT and OTBlang; the dataset names should be corrected for consistency.","section":"Table 1 caption"},{"comment":"The overview text refers to 'TSATrack' in §3.1 ('Fig. 2 shows the general framework of the TSATrack'), which should be 'ATSTrack'.","section":"§3.1, Figure 2"},{"comment":"The LaSOT paragraph claims ATSTrack 'outperforms the second best visual-language tracker by 1.8% in term of AUC,' but the table shows the second best VL tracker on LaSOT AUC is DecoupleTNL at 71.2, giving a gap of 1.4 points, not 1.8; this should be corrected or clarified which metric is meant.","section":"§4.2"},{"comment":"The 'Effect of Each attribute' section describes results, but the table rows are not labeled with the corresponding AUC in the text in a way that allows a reader to trace each claimed drop; I recommend adding the numeric deltas directly in the text or in the table.","section":"§4.3, Table 3"},{"comment":"There are several reference formatting errors (e.g., [9] author list 'Vitaly Feng Qi and, Ablavsky, Qinxun Bai', [15] and [21] missing venue page details); the authors should proofread the bibliography.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising empirical result on TNL2K and a clear conceptual framing, but the reproducibility of the LLM attribute decomposition is a genuine load-bearing gap, and the ablation text/table contradictions suggest the experimental narrative needs careful verification. The α selection on LaSOT is also a correctness-risk issue that should be addressed head-on. I would be happy to re-review a revised version that specifies the LLM, fixes the ablation inconsistencies, and reports a proper hyperparameter selection procedure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a reasonable VLT paper with a strong TNL2K result and a sensible extension of attribute-decomposition ideas, but not ready as written. The contributions overclaim on OTBlang, the ablation text contradicts Table 2c, and the LLM-based attribute decomposition the whole method relies on is undocumented. These are fixable, so I would send it to review rather than desk-reject.\n\nWhat is actually new: the per-attribute routing—category/appearance to the latest template, action to the template sequence, location to the search region—and the cross-modal token fed back into the visual backbone. CiteTracker already had the four-attribute decomposition and QueryNLT had prompt-based modulation, but the specific combination of fine-grained matching to different visual scales is a legitimate extension. The TNL2K AUC of 66.2 is a real jump over the previous best language tracker (63.1), and the LaSOT results are competitive with strong visual-only trackers. The ablations mostly support the modules' value, and the paper cites relevant prior work fairly.\n\nSoft spots, in increasing severity. First, the contributions bullet says 'outperforms SOTA on three datasets,' but the body correctly notes ATSTrack is second on OTBlang (71.0 vs 73.8 for DecoupleTNL). Easy fix. Second, the text says the LFA gating weight α=25 is best, but Table 2c shows α=50 gives AUC 72.0 vs 71.3 for α=25. That is a factual contradiction. Third, α is tuned on LaSOT ablations and then the same benchmark is used for final numbers; that is mild test-set fitting and should be acknowledged or split.\n\nThe big one: the LLM attribute decomposition. Section 4.1 only says 'we use LLM to segment the language descriptions'—no model, prompt, or post-processing, and no evaluation of segmentation quality. Every module in the FGM depends on phrases being assigned to the right attribute; a noisy decomposition would collapse the method to a standard coarse-grained cross-attention tracker. Table 3 removes whole attributes, which does not test whether the LLM put the right words in each slot. The stress-test note is right: without evidence of segmentation reliability, the reported gains cannot be attributed to temporal/spatial alignment.\n\nWho this is for: anyone in vision-language tracking or multi-modal feature modulation. Not a new direction, but a solid incremental step if the details hold. I would accept it for peer review and request a major revision: specify the LLM and evaluate decomposition accuracy, correct the α contradiction, release code, and tighten the claims.","headline":"Strong TNL2K result and a reasonable architectural extension, but the undocumented LLM decomposition and internal contradictions make this a conditional accept rather than a clean one.","tokens_in":12448,"tokens_out":4843,"would_cite":true,"duration_ms":49143,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mismatched temporal and spatial scales, not missing cross-modal interaction, is what limits visual-language tracking.","keywords":["visual-language tracking","single object tracking","temporal and spatial scale alignment","fine-grained feature modulation","visual-language token","multimodal transformer","natural language specification"],"falsifier":"A direct test would be to take the trained ATSTrack and replace the LLM attribute parsing with a fixed randomized assignment of phrases to the four attributes, or with one unified attribute, while keeping everything else identical, then measure AUC on TNL2K; if the score does not drop below the reported 66.2 by a margin comparable to the FGM ablation gain of about 1.4 AUC on LaSOT, then the fine-grained alignment itself is not doing the work.","tokens_in":1740,"feed_emoji":"🎯","tokens_out":3834,"duration_ms":89037,"temperature":0.7,"pith_summary":"The paper tries to establish that the remaining bottleneck in visual-language tracking is not a lack of cross-modal interaction but a mismatch in temporal and spatial scales between language phrases and visual evidence. It claims that decomposing each language description into four attributes, category, appearance, action, and location, and modifying each attribute's features with the visual input that shares its scale improves tracking. It also claims that feeding a visual-language token from the previous frame into the visual backbone guides feature extraction toward language-relevant content. If right, this would make natural-language descriptions a reliably stronger tracking cue, especially in long-term and adversarially described sequences, and it would explain the reported 66.2 AUC on TNL2K, 3.1 points above the prior best language tracker.","feed_headline":"Language tracker aligns text scales to hit 66.2 AUC on TNL2K","feed_subtitle":"Splitting text into four attributes, each matched to the right visual evidence, lifts accuracy 3.1 points.","key_machinery":"The central mechanism is the four-way attribute decomposition together with the Fine-Grained Modification (FGM) module and the Visual-Language (VL) token. The four attributes, category, appearance, action, and location, are defined by which visual input matches their temporal and spatial scale; FGM then applies a custom operation per attribute, with the category feature used to suppress background in the latest template, action features cross-attending over the whole template sequence, and location features gated against the search frame. The VL token, formed by concatenating the average of the modified language features with the visual class token, is fed into the visual backbone of the next frame to bias feature extraction toward language-relevant content.","core_discovery":"On its own terms, the central discovery is that explicit alignment of the temporal and spatial granularity of language and visual inputs is what makes feature modification work in visual-language tracking. Rather than treating all language features and all visual features as two wholes, ATSTrack decomposes each description into category, appearance, action, and location phrases, then routes each phrase to a matching visual source: category and appearance to the latest template, action to the whole template sequence, and location to the search frame. The Fine-Grained Modification module applies a different operation per attribute, including a visual feature modification that suppresses background in the template using category features and a language feature ablation that gates out location tokens unrelated to the search image. A visual-language token assembled from the previous frame's modified language features and the visual class token is injected into the next frame's backbone, so visual features are extracted under linguistic guidance. The reported consequence is state-of-the-art or competitive results on TNL2K, LaSOT, and OTBlang.","pith_inferences":["The fixed four-attribute decomposition is a heuristic tied to how existing datasets phrase descriptions; datasets with richer relational or temporal language would likely need additional attribute types or a learned decomposer for the benefit to transfer.","If the LLM segmentation is noisy, part of the FGM gain could come from an implicit regularization effect rather than from true scale alignment; replacing LLM parsing with human-annotated attribute labels could separate these two explanations.","The VL-token mechanism is a lightweight way to propagate language through a visual backbone, and a similar token could be applied to other vision-language tasks such as referring expression segmentation or video grounding, where spatial-scale mismatch also occurs.","The finding that ungated location descriptions can interfere suggests future datasets should annotate location relative to changing scene context, and the LFA module may partly compensate for annotation bias."],"forward_implications":["On TNL2K, ATSTrack reaches 66.2 AUC, 3.1 points above the previous best language tracker and 3.5 points above the best visual tracker, suggesting scale alignment specifically helps adversarial and camera-switch samples.","On LaSOT, the 72.6 AUC is competitive with top visual-only trackers and 1.8 points above the second-best language tracker, indicating a benefit for long-term tracking.","Ablations on LaSOT attribute roughly 1.4 AUC to the fine-grained modification module and 0.6 AUC to the visual-language token, so both components contribute to the overall gain.","Attribute-level ablations show appearance descriptions matter most, while action and location descriptions contribute less in the current datasets.","Concatenating visual and language information into a single VL token outperforms cross-attention between the two tokens in the reported experiments."],"supporting_citations":[{"why":"Provides the single-stream visual tracker OSTrack whose backbone and prediction head ATSTrack adopts, and serves as the visual-only baseline.","marker":"[39]"},{"why":"QueryNLT is the prior prompt-modulation visual-language tracker whose whole-feature modification ATSTrack argues is limited and whose results it surpasses.","marker":"[29]"},{"why":"CiteTracker is the source of the four-attribute decomposition idea and of the CLIP-based attribute and color prediction used for TrackingNet.","marker":"[19]"},{"why":"DecoupleTNL is a long- and short-term context-decoupling visual-language tracker that ATSTrack compares against and extends with explicit temporal alignment.","marker":"[25]"},{"why":"UVLTrack is the previous best vision-language tracker on TNL2K that ATSTrack's 66.2 AUC surpasses by 3.1 points.","marker":"[26]"},{"why":"TNL2K is the language-tracking benchmark with manual descriptions used for training and for the headline comparison.","marker":"[31]"},{"why":"LaSOT is the long-term tracking benchmark where the ablations and the 72.6 AUC comparison are carried out.","marker":"[8]"},{"why":"CLIP is used as the language backbone and as the predictor of target color for the category-only TrackingNet annotations.","marker":"[28]"},{"why":"SNLT introduced the OTBlang language-annotated OTB benchmark and is one of the earliest Siamese natural-language trackers used for comparison.","marker":"[9]"}],"fun_headline_variants":["Text scale alignment lifts visual-language tracking accuracy","Splitting text into attributes matches visual evidence in tracking","ATSTrack aligns temporal and spatial scales for VLT","Visual-language token from prior frame guides next-frame features","Fine-grained text-visual alignment raises TNL2K AUC to 66.2"],"cache_read_input_tokens":14592,"weakest_assumption_plain":"The whole method depends on a language model cleanly splitting each sentence into exactly four parts, category, appearance, action, and location, and if that split is noisy, the fine-grained alignment gains vanish and the model degrades to ordinary cross-attention.","fun_headline_variants_meta":{"raw":{"variants":["Text scale alignment lifts visual-language tracking accuracy","Splitting text into attributes matches visual evidence in tracking","ATSTrack aligns temporal and spatial scales for VLT","Visual-language token from prior frame guides next-frame features","Fine-grained text-visual alignment raises TNL2K AUC to 66.2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000286,"raw_usage":{"total_tokens":1686,"prompt_tokens":955,"completion_tokens":731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":648}},"tokens_in":571,"tokens_out":731,"duration_ms":8788,"temperature":1.0,"reasoning_tokens":648,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:14:21.528413+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to take the trained ATSTrack and replace the LLM attribute parsing with a fixed randomized assignment of phrases to the four attributes, or with one unified attribute, while keeping everything else identical, then measure AUC on TNL2K; if the score does not drop below the reported 66.2 by a margin comparable to the FGM ablation gain of about 1.4 AUC on LaSOT, then the fine-grained alignment itself is not doing the work.","supporting_citations":[{"cited_title":"Joint feature learning and relation modeling for tracking: A one-stream framework","cited_arxiv_id":null,"evidence_quote":"Provides the single-stream visual tracker OSTrack whose backbone and prediction head ATSTrack adopts, and serves as the visual-only baseline."},{"cited_title":"Context-aware integration of lan- guage and visual references for natural language tracking","cited_arxiv_id":null,"evidence_quote":"QueryNLT is the prior prompt-modulation visual-language tracker whose whole-feature modification ATSTrack argues is limited and whose results it surpasses."},{"cited_title":"Citetracker: Correlating image and text for visual tracking","cited_arxiv_id":null,"evidence_quote":"CiteTracker is the source of the four-attribute decomposition idea and of the CLIP-based attribute and color prediction used for TrackingNet."},{"cited_title":"Tracking by natural language specification with long short-term context decoupling","cited_arxiv_id":null,"evidence_quote":"DecoupleTNL is a long- and short-term context-decoupling visual-language tracker that ATSTrack compares against and extends with explicit temporal alignment."},{"cited_title":"Unifying visual and vision-language tracking via contrastive learning","cited_arxiv_id":null,"evidence_quote":"UVLTrack is the previous best vision-language tracker on TNL2K that ATSTrack's 66.2 AUC surpasses by 3.1 points."},{"cited_title":"Towards more flexible and accurate object tracking with natural language: Algo- rithms and benchmark","cited_arxiv_id":null,"evidence_quote":"TNL2K is the language-tracking benchmark with manual descriptions used for training and for the headline comparison."},{"cited_title":"Lasot: A high-quality benchmark for large-scale single ob- ject tracking","cited_arxiv_id":null,"evidence_quote":"LaSOT is the long-term tracking benchmark where the ablations and the 72.6 AUC comparison are carried out."},{"cited_title":"Siamese natural language tracker: Tracking by nat- ural language descriptions with siamese trackers","cited_arxiv_id":null,"evidence_quote":"SNLT introduced the OTBlang language-annotated OTB benchmark and is one of the earliest Siamese natural-language trackers used for comparison."}],"review_version":1}