{"id":"351a5295-2c0a-4424-ad25-2973a212c190","arxiv_id":"2508.11531","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"MST combines multi-state feature generation, lightweight state-specific enhancement, and cross-state interaction to achieve state-of-the-art accuracy among efficient object trackers on multiple datasets.","lead":"This paper introduces a new object tracking architecture, Multi-State Tracker (MST), which improves tracking accuracy and speed for real-time applications. It reports a 4.5% higher accuracy score than the previous best efficient tracker on one benchmark, with only 0.1 GFLOPs added computation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency-accuracy claim rests on unreported comparison protocol and module-only overhead; reproducibility check needed.","rationale":"The reader's UNVERDICTED verdict is appropriate because the manuscript is available only as an abstract. My stress-test identifies a specific, load-bearing concern: the empirical advantage is not backed by a reproducible protocol in the abstract, and the reported overhead is module-specific rather than end-to-end. This does not change the verdict because the paper is unverified either way; it remains UNVERDICTED. The reader's weakest_assumption emphasizes the HSA-SSD modeling efficacy; my concern is more about the fairness and completeness of the efficiency comparison. Thus, partial agreement. The proposed concrete test would settle whether the 4.5% AO and 0.1 GFLOPs/0.66M claims are real under a controlled comparison.","tokens_in":739,"tokens_out":2525,"duration_ms":26576,"concrete_test":"Download the released code from https://github.com/wsumel/MST. Reproduce the GOT-10K AO for MST and the HCAT baseline under identical conditions: same training data, augmentation, epochs, input resolution, and test script. Use the provided trained models to run the test split. Measure full model FLOPs and parameters using standard tools (e.g., ptflops) with a representative input, and measure real runtime on the same GPU. Compare to the abstract's stated 0.1 GFLOPs/0.66M module overhead. If the AO gap between MST and HCAT shrinks below the reported 4.5% or the end-to-end overhead exceeds the stated numbers by more than 10%, the central efficiency-accuracy claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MST outperforms all previous efficient trackers, specifically a 4.5% AO improvement over HCAT on GOT-10K, while adding only 0.1 GFLOPs and 0.66M parameters in the SSE and CSI modules. The abstract does not state the total model FLOPs, parameter count, or runtime after adding these modules, nor does it specify the comparison protocol (shared backbone, training data, number of epochs, input resolution, inference framework). Since the overhead is reported only for the two introduced modules, the end-to-end efficiency may be higher. Furthermore, the HSA-SSD design is described only by name; without a precise definition or pseudocode, its theoretical cost cannot be independently verified. The 4.5% AO improvement could stem from training details or evaluation settings rather than the proposed architecture. If the comparison is not apples-to-apples, the claimed efficiency-accuracy tradeoff does not hold.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The abstract proposes Multi-State Tracker (MST), an efficient visual object tracker built from three components: multi-state generation (MSG), state-specific enhancement (SSE), and cross-state interaction (CSI). The authors claim that SSE and CSI adopt a highly lightweight hidden state adaptation-based state space duality (HSA-SSD) design, adding only 0.1 GFLOPs and 0.66M parameters. They report that MST outperforms all previous efficient trackers on multiple datasets, with a 4.5% AO improvement over the previous SOTA efficient tracker HCAT on GOT-10K. The manuscript under review contains only the abstract; no full text, equations, figures, tables, or experimental details are available.","tokens_in":973,"tokens_out":2083,"duration_ms":23417,"significance":"If the stated results hold, the proposed lightweight multi-state interaction mechanism would be a practically useful contribution to efficient object tracking, as it claims a favorable accuracy/efficiency trade-off with a very small module-level overhead. The public code link is a positive reproducibility signal. However, because the submission is abstract-only, the technical novelty and the empirical claims cannot be independently verified. The significance is therefore conditional on the full manuscript providing a precise definition of HSA-SSD, complete end-to-end efficiency numbers, and rigorous comparisons.","major_comments":[{"comment":"The central claim, 'MST outperforms all previous efficient trackers across multiple datasets,' is made without any supporting experimental details in the provided manuscript. There are no tables, no error bars, no statistical significance tests, and no description of the evaluation protocol. As presented, the claim is unsupported and cannot be assessed. The full paper must include complete comparison tables, dataset splits, and evaluation metrics for every reported result.","section":"Abstract (overall)"},{"comment":"The reported cost of '0.1 GFLOPs in computation and 0.66 M in parameters' is explicitly attributed to the SSE and CSI modules only. The abstract does not report the total model FLOPs, total parameters, or end-to-end runtime after adding these modules. Since the efficiency-accuracy trade-off is the paper's main selling point, the module-only overhead is insufficient; the reader must know the full model cost and inference speed to judge whether MST is genuinely efficient in practice.","section":"Abstract (efficiency claim)"},{"comment":"The 'hidden state adaptation-based state space duality (HSA-SSD) design' is only named, not defined. No equations, pseudocode, or architectural details are provided in the abstract. Without a precise formulation, the claimed theoretical cost of 0.1 GFLOPs and 0.66M parameters cannot be independently verified, and the novelty of the design cannot be evaluated. The full manuscript must give a formal definition and complexity analysis.","section":"Abstract (HSA-SSD description)"},{"comment":"The 4.5% AO improvement over HCAT on GOT-10K is reported as a headline result, but the abstract gives no information about the comparison protocol: shared backbone, training data, input resolution, number of training epochs, or inference framework. Any of these factors could explain the gain. A controlled, apples-to-apples comparison with the same backbone and training setup, plus ablations isolating the contribution of each module, is required before this improvement can be attributed to the proposed architecture.","section":"Abstract (comparison protocol)"}],"minor_comments":[{"comment":"The phrase 'single-layer features' is vague; it should clarify what is meant by 'state' and how multi-state features differ from standard multi-level or multi-branch features.","section":"Abstract"},{"comment":"The abstract states 'code is available' but gives no license, repository structure, or instructions for reproducing the reported numbers. A reproducibility statement would strengthen the submission.","section":"Abstract"},{"comment":"The name 'Multi-State Tracker' and the acronyms MSG, SSE, CSI, HSA-SSD are introduced without expansion beyond the abstract. The full paper should define all terms early.","section":"General"}],"recommendation":"uncertain","confidential_remarks":"The manuscript provided for review contains only the abstract. There is no full text, and therefore no technical content to referee. The central claims are unverifiable from the submitted material. I recommend obtaining the full manuscript before any further decision. If the full paper is not available, the editor should treat this as an abstract-only submission and not as a complete refereed manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the abstract for MST describes a plausible, well-motivated architecture for efficient object tracking, and the reported 4.5% AO jump over HCAT on GOT-10K is the kind of result that should get a serious look. But right now the evidence is just the abstract, and the efficiency claim is under-specified in a way that could cut either way.\n\nWhat's actually new: the idea of generating multiple target states at different feature-extraction stages, giving each state a lightweight specialized enhancement, and then fusing them with a cross-state interaction module built on a \"hidden state adaptation-based state space duality\" (HSA-SSD). The HSA-SSD design is the interesting piece—if it truly costs only 0.1 GFLOPs and 0.66M parameters for both added modules, that's a meaningful contribution to the efficient-tracking design space. The authors also do the right thing by releasing code and comparing against the previous efficient SOTA.\n\nSoft spots: the numbers reported are module-only. The abstract doesn't state total model FLOPs, parameter count, or end-to-end runtime after integrating SSE and CSI, so we can't yet verify that the efficiency-accuracy tradeoff holds in practice. It also doesn't specify the comparison protocol: shared backbone, training data, number of epochs, input resolution, inference hardware. All of these can shift a 4.5% AO gap. And HSA-SSD is named but not defined, so its theoretical cost and novelty relative to existing state-space trackers can't be assessed. These are gaps in the abstract, not necessarily flaws in the full paper—but they do mean the central claim is unverified.\n\nThe stress-test note about unreported comparison protocol and module-only overhead is a fair concern; I'd want the full paper to explicitly address both. If the full version shows apples-to-apples comparisons and reports end-to-end costs, this could be a solid contribution to efficient tracking. If it doesn't, the headline improvement is just a number.\n\nWho is this for? Researchers working on real-time or resource-constrained tracking, and people interested in applying state-space models to visual feature refinement. It's not a paradigm shift, but it could influence how lightweight fusion modules are designed.\n\nRecommendation: send it to peer review. The significance is high enough and the proposed design is concrete enough to warrant referee time, even though I can't yet vouch for the results. I'd bring it to a reading group once the full text is out, and I wouldn't cite it myself until I've seen the experimental details.\n\nBest,\n\n[Your name]","headline":"MST looks like a real step forward for efficient trackers, but the abstract's headline numbers need the full method and experimental protocol before they convince.","tokens_in":1395,"tokens_out":1500,"would_cite":false,"duration_ms":18778,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that splitting tracking features into multiple specialized states, refining each one, and letting them interact yields state-of-the-art efficient tracking accuracy with only 0.1 GFLOPs added.","keywords":["visual object tracking","efficient tracking","multi-state representation","state-specific enhancement","cross-state interaction","state space model","GOT-10K","lightweight network"],"falsifier":"Run MST and the previous best efficient tracker on GOT-10K using identical backbone, input size, and postprocessing; if the 4.5% AO advantage disappears when SSE and CSI are removed, or if the measured cost of the two modules exceeds 0.1 GFLOPs and 0.66M parameters, the claimed efficiency-accuracy tradeoff is not supported.","tokens_in":695,"feed_emoji":"🎯","tokens_out":5295,"duration_ms":56087,"temperature":0.7,"pith_summary":"The paper introduces Multi-State Tracker (MST), an efficient visual object tracker that aims to close the accuracy gap between efficient and heavyweight trackers without sacrificing speed. Its central claim is that a single-layer feature is not enough to capture a target's state; generating multiple state representations at different feature-extraction stages, enhancing each state separately, and exchanging information between them through a very lightweight interaction mechanism sharply improves robustness. MST reports an average-overlap improvement of 4.5% over the previous best efficient tracker on GOT-10K while adding only 0.1 GFLOPs and 0.66 million parameters via its two new modules. The result, if correct, means efficient tracking can recover much of the representational power that was traded away for speed.","feed_headline":"Tiny 0.1-GFLOP module lifts tracker accuracy by 4.5%","feed_subtitle":"Multi-state feature fusion lets light trackers beat every prior efficient tracker on GOT-10K.","key_machinery":"The central mechanism is HSA-SSD (hidden state adaptation-based state space duality), the shared lightweight design for SSE and CSI. It treats feature refinement and cross-state information exchange as hidden-state updates in a state-space formulation, so that the two modules cost only 0.1 GFLOPs and 0.66M parameters. This is what lets the multi-state representation improve accuracy while keeping the tracker efficient.","core_discovery":"MST is built from three parts working together: multi-state generation (MSG) creates several target-state representations at multiple stages of feature extraction; state-specific enhancement (SSE) refines each representation to bring out target-specific cues; and cross-state interaction (CSI) lets the refined states exchange and integrate complementary information. The SSE and CSI modules share a hidden state adaptation-based state space duality (HSA-SSD) design, which the paper argues keeps the added computation to 0.1 GFLOPs and added parameters to 0.66 million. With this setup, MST is evaluated across multiple tracking benchmarks and is reported to outperform all previous efficient tracke","pith_inferences":["Because SSE and CSI are modular and tiny, a natural next step is to drop them into existing efficient trackers and see whether the gain transfers beyond MST's own backbone.","The state-space formulation invites a direct comparison to other state-space sequence models; testing HSA-SSD with a longer temporal context or in multi-object tracking would show whether the duality is specific to this design or a general property.","The paper's reported overhead counts the two added modules; end-to-end FPS still depends on the feature-extraction backbone, so benchmarking MST with the same backbone as its competitors would isolate where the 4.5% comes from."],"forward_implications":["If MST's results hold, efficient trackers no longer need single-layer features; multi-state specialization can restore much of the lost representation capacity.","The added modules are small enough that the accuracy gain is available on low-compute hardware.","Cross-state interaction lets complementary target cues combine, which the paper says improves robustness in complex environments.","The reported 4.5% AO gain on GOT-10K would establish a new efficiency-accuracy point for lightweight tracking."],"supporting_citations":[],"fun_headline_variants":["Multi-state fusion adds 0.1 GFLOPs, lifts tracker 4.5%","0.1 GFLOPs: Multi-state trick beats every efficient tracker","State-specific interaction gives light trackers a 4.5% edge","Tracker accuracy gains 4.5% with just 0.1 GFLOPs"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The whole efficiency-accuracy tradeoff rests on the assumption that the HSA-SSD state-space update genuinely captures target dynamics well enough to refine and fuse multi-state features for only 0.1 GFLOPs and 0.66M parameters; if that update is either ineffective or more expensive than reported, the central claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Multi-state fusion adds 0.1 GFLOPs, lifts tracker 4.5%","0.1 GFLOPs: Multi-state trick beats every efficient tracker","State-specific interaction gives light trackers a 4.5% edge","Tracker accuracy gains 4.5% with just 0.1 GFLOPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00066,"raw_usage":{"total_tokens":2878,"prompt_tokens":790,"completion_tokens":2088,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":1998}},"tokens_in":534,"tokens_out":2088,"duration_ms":17679,"temperature":1.0,"reasoning_tokens":1998,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:49:41.964458+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MST and the previous best efficient tracker on GOT-10K using identical backbone, input size, and postprocessing; if the 4.5% AO advantage disappears when SSE and CSI are removed, or if the measured cost of the two modules exceeds 0.1 GFLOPs and 0.66M parameters, the claimed efficiency-accuracy tradeoff is not supported.","supporting_citations":[],"review_version":1}