{"id":"ad7c64ba-d82a-4fa5-a8de-0c309722e7cb","arxiv_id":"2608.07015","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Language-supervised pretraining on a new multimodal infrared dataset improves infrared small-target detection and cross-domain generalization compared with prior IRST detectors.","lead":"JinSight instruction-tunes a vision-language model on a new 39k-pair infrared dataset, then transfers the learned features to small-target detection, reporting state-of-the-art IoU on WideIRSTD. The work shows a path toward one detector that works across infrared sensor bands and scenes, which matters for surveillance and autonomous systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Capacity confound, not language supervision, likely drives the headline gains: Table 4's same-capacity visual-only backbone (40.72 IoU) already beats all SOTA in Table 2, and cross-dataset tasks have no such control.","rationale":"The reader's weakest assumption points to the missing same-capacity visual-only baseline on cross-dataset tasks; this is exactly the most load-bearing gap. I extend it with an internal observation: Table 4 already provides such a control on WideIRSTD, and it shows the 314M InternVL ViT alone beats all Table 2 baselines. This strengthens the concern that the headline 14% gain is mostly capacity, not paradigm. I do not move the verdict because the reader's CONDITIONAL verdict already requires adding exactly this control. A secondary internal inconsistency (main text uses r=128 while Table 5 reports r=256 as optimal) and the absence of error bars further justify the conditional verdict, but the missing cross-dataset control is decisive. My recommendation is UNCHANGED: keep the paper conditional pending the same-capacity visual-only evaluation and clarified comparison protocol.","tokens_in":17419,"tokens_out":5141,"duration_ms":54052,"concrete_test":"Train the same 314M InternVL2.5-1B ViT + UperNet + LSI on WideIRSTD-Train with r=128 and identical Stage-II settings, but initialized from the raw pretrained ViT (skip Stage I entirely). Evaluate this visual-only control on NUDT-SIRST, NUAA-SIRST, and IRSTD-1K using Table 2's cross-dataset protocol. If this control matches or exceeds JinSight's Pd/IoU on those three shifts, the language-driven domain-invariant generalization claim fails; if it is substantially lower, the paradigm survives. Report results over at least three seeds with mean and std, and also recompute the abstract's '14% IoU improvement' relative to this control rather than to lower-capacity SOTA.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Stage-I language supervision learns domain-invariant semantics that drive the reported improvements is not established, because the comparisons are confounded by backbone scale. JinSight uses the InternVL2.5-1B ViT (314M parameters), whereas the best visual baseline in Table 2 (Text-IRSTD) has 181M and most baselines have under 15M. More tellingly, Table 4's own 'w/o IVIT, w/o LSI' row -- the same 314M ViT plus UperNet, with no language supervision and no LSI -- reaches 40.72 IoU and 70.28 Pd on WideIRSTD, already exceeding every method in Table 2. The 'over 14% IoU improvement' headline is therefore primarily a capacity/backbone effect; the language-plus-LSI increment over that same-capacity visual-only baseline is only 42.94 vs 40.72 (about 5.5% relative). For the three cross-dataset generalization tasks in Table 2, no same-capacity visual-only control is reported at all, so the claim that JinSight 'consistently transfers across heterogeneous domains' cannot be attributed to the understand-before-detect paradigm. Without that control, the central contribution is untested against the most plausible alternative explanation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"JinSight is a two-stage framework for omni-domain infrared small target detection. Stage I instruction-tunes a generative VLM (InternVL2.5-1B) on OmniIRST-VL, a new dataset of 39.7k image-instruction-response pairs derived from WideIRSTD, covering captioning, physical VQA, counting, and grounding. Stage II transfers the tuned ViT to dense prediction via UperNet and a low-rank Latent Semantic Interaction (LSI) module that exchanges global semantics with multi-scale spatial features. The paper claims consistent state-of-the-art detection performance, including over 14% IoU improvement on WideIRSTD, and attributes the gains to language supervision that learns domain-invariant semantics.","tokens_in":17662,"tokens_out":7136,"duration_ms":64318,"significance":"The OmniIRST-VL dataset is a substantial new resource: it spans six instruction tasks, multiple platforms and spectral bands, includes expert verification, and will be released with code and weights. The internal ablation in Table 4 is well designed and shows that instruction tuning and LSI add consistent, if modest, gains over a same-backbone visual baseline. The leave-one-dataset-out evaluation protocol is a good methodological choice. However, the paper's central claim of consistent state-of-the-art performance is not yet established because the main comparisons in Table 2 are confounded by model capacity: the same backbone without any language supervision already beats all published baselines. The cross-dataset generalization claim lacks a same-capacity visual-only control. If the authors can address this capacity confound, the work would be a solid contribution to the IRST community.","major_comments":[{"comment":"The main comparison is confounded by model capacity. JinSight uses a 314M-parameter InternVL ViT backbone, whereas the best visual baseline (Text-IRSTD) has 181M parameters and most baselines have under 15M. Table 4's 'w/o IVIT, w/o LSI' row shows that the same 314M ViT plus UperNet, with no language supervision and no LSI, already reaches 40.72 IoU and 70.28 Pd on WideIRSTD, exceeding every method in Table 2 (best SOTA IoU: Text-IRSTD at 36.62). The full JinSight reaches 42.94 IoU, so the gain attributable to the proposed paradigm is about 2.2 IoU (5.5% relative), not the 'over 14%' claimed in the abstract. For the three cross-dataset generalization tasks in Table 2, no same-capacity visual-only control is reported, so the claim that JinSight 'consistently transfers across heterogeneous domains' is not established against the most plausible alternative explanation. Please add a visual-only baseline using the same InternVL ViT backbone (with and without LSI) on all four evaluation settings.","section":"§5.2 Table 2; §5.3 Table 4"},{"comment":"The implementation details state that the LSI latent dimension is set to r=128, but Table 5 reports r=256 as giving the highest IoU (42.16 vs 41.79 for r=128). If the main results use r=128, the paper should justify that choice explicitly; if r=256 is optimal, the main experimental results in Table 2 should be updated accordingly. This inconsistency matters for reproducibility and for the reported SOTA numbers.","section":"§9.1 vs §5.3 Table 5"},{"comment":"The comparison of JinSight-1B with generic VLMs is not a controlled test of the 'understand before detect' paradigm. JinSight-1B is fine-tuned on OmniIRST-Train, while InternVL2.5-1B, Qwen3-VL-2B, and LLaVA-NeXT-7B are evaluated zero-shot; the improvements could simply reflect in-distribution training. A same-initialization baseline fine-tuned on generic visual instruction data (or on OmniIRST-VL with a subset of instructions) would be needed to attribute the gains to the proposed instruction set.","section":"Table 3"}],"minor_comments":[{"comment":"The phrase 'over 14% IoU improvement' is ambiguous (relative vs absolute) and is misleading in light of the capacity confound; please state the baseline and the relative/absolute interpretation explicitly.","section":"Abstract and Conclusion"},{"comment":"Figure 3 contains typographical errors ('Descirbe' and the malformed closing tag '<\\point>' in the example); please proofread the figure text.","section":"Figure 3"},{"comment":"The memory usage with LSI (e.g., 9.43 GB at r=64) is lower than without LSI (16.92 GB), which is counterintuitive since LSI adds parameters; please explain this result.","section":"Table 5"},{"comment":"The reference for EarthDial (Sagar Soni et al., 2025) has a garbled author list; please correct the citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-written and the dataset is potentially valuable. The central issue is the capacity confound; if the authors can show a same-capacity visual-only baseline performing worse, especially on cross-dataset tasks, the paper would be much stronger. I would encourage the editor to request this revision rather than reject, as the methodological idea is sound and the internal ablation is encouraging."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it ships something genuinely useful: OmniIRST-VL, the first large-scale vision–language dataset for infrared small-target detection, with 39k instruction–response pairs across captioning, physical VQA, counting, and grounding. That alone fills a real gap, and the dataset construction is careful. Second, the central claim of consistent state-of-the-art performance is not supported by the evidence as presented, because the comparisons are confounded by model scale.\n\nThe within-dataset ablations are the honest part. Table 4 shows that a 314M-parameter InternVL ViT plus UperNet, with no language supervision and no LSI, reaches 40.72 IoU on WideIRSTD. That already beats every method in Table 2, including Text-IRSTD's 36.62. Adding instruction tuning and LSI gets you to 42.94 IoU, a real but modest 5.5% relative gain over the same-capacity visual-only baseline. So the paradigm does something, but the paper's \"over 14% IoU improvement\" is almost entirely a backbone-capacity effect, not evidence of domain-invariant semantics. The cross-dataset generalization results in Table 2 have no same-capacity visual-only control at all, so the claim that language supervision drives transfer across heterogeneous domains is untested against the obvious alternative explanation.\n\nThere are also two smaller problems. The hyperparameter reporting is inconsistent: implementation details say LSI uses r=128, but the latent-dimension ablation in Table 5 reports r=256 as optimal. The ablation is run without Stage-I instruction tuning, so the optimal r for the full model is never shown. And no error bars are reported anywhere, so the gap between 40.72 and 42.94 might be within noise, especially on small test sets.\n\nWhat the paper does well: it retrains all baselines on the same split, the ablation of instruction composition in Table 6 is informative, and LSI looks like an efficient and sensible bridge between coarse ViT features and dense prediction. The authors are also upfront that the benchmark is derived from WideIRSTD rather than new imagery, which is honest.\n\nBottom line: this is a solid dataset-plus-framework paper that deserves a serious referee, but the claims need to be re-framed. The authors must add same-capacity visual-only baselines for the cross-dataset tasks, reconcile the r=128/r=256 choice, and stop attributing the headline gains to language supervision. With those changes, the modest but real contribution of the understand-before-detect recipe would be credible. I would bring it to reading group for the capacity-confound discussion alone.","headline":"The OmniIRST-VL dataset and JinSight framework are real contributions, but the headline gains are mostly backbone scale; the language-supervision effect is modest and cross-dataset claims lack a same-capacity control.","tokens_in":18251,"tokens_out":3470,"would_cite":true,"duration_ms":36527,"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":"This paper claims that infrared small-target detection should begin with language-supervised scene understanding, and that its two-stage vision-language detector, JinSight, raises WideIRSTD IoU by over 14% relative to prior methods.","keywords":["infrared small target detection","omni-domain generalization","vision-language learning","visual instruction tuning","dense prediction","low-rank cross-attention","multimodal benchmark"],"falsifier":"Train the same InternVL ViT, initialized from the same pretrained weights, directly into the same dense-prediction decoder on the same training split without Stage I language tuning and with a matching training budget; if its IoU on WideIRSTD approaches or exceeds 42.94, the paradigm's claimed contribution is not supported.","tokens_in":17200,"feed_emoji":"🛰️","tokens_out":12342,"duration_ms":112555,"temperature":0.7,"pith_summary":"The paper argues that infrared small-target (IRST) detection trained only on binary target/background masks loses the scene-level semantics that stay stable across heterogeneous infrared domains, which is why domain-specific detectors degrade under domain shift. It proposes \"understand before detect\": first instruction-tune a generative vision-language model on infrared understanding tasks, then transfer the language-tuned visual backbone to dense detection. Building on that paradigm, JinSight combines an IRST-oriented visual instruction-tuning stage with a low-rank Latent Semantic Interaction module that exchanges global semantics with fine-grained spatial features, and the paper contributes OmniIRST-VL, the first large-scale vision-language dataset for omni-domain IRST detection with 39,701 instruction-response pairs. If the paradigm holds, a single model can replace a collection of domain-specific infrared detectors and generalize to unseen imaging conditions.","feed_headline":"Describe the infrared scene first, then detect: IoU up 14%.","feed_subtitle":"Language-supervised understanding before detection lets one model work across NIR, SWIR, and LWIR sensors.","key_machinery":"The load-bearing object is the two-stage \"understand before detect\" pipeline. In Stage I, the paper takes the InternVL2.5-1B generative vision-language model and fine-tunes its ViT, projector, and language model on the OmniIRST-VL instruction mixture with the autoregressive objective $\\mathcal{L}_{\\text{inst}} = -\\sum_{t=1}^{|y|} \\log p(y_t \\mid I, q, y_{<t})$, then keeps the instruction-tuned ViT as an IRST-aware backbone. In Stage II, the language projector and decoder are dropped, and Latent Semantic Interaction (LSI) modules are interleaved with the ViT blocks: each LSI block runs bidirectional cross-attention in a latent space of dimension $r \\ll D$, transferring local spatial details from multi-scale features into ViT tokens and propagating global semantics back, with a low-rank FFN refining the latent tokens. OmniIRST-VL supplies 39,701 instruction-response pairs across six tasks (detailed captioning, physical VQA, image- and region-level counting, image- and region-level grounding), organized into scene-oriented and target-centric groups. The design's purpose is to make the encoder retain scene context and target-context relations that binary masks throw away, so that downstream dense prediction sees domain-invariant semantics.","core_discovery":"The paper's central claim is that generative language supervision, applied before any dense-prediction training, reshapes an infrared visual encoder so that it carries domain-invariant semantics rather than domain-specific low-level statistics. Under this claim, the autoregressive objective forces the vision transformer to keep spatial and semantic evidence about few-pixel targets, while language ties heterogeneous imaging conditions (NIR/SWIR/LWIR, land/aerial/space platforms) to a shared vocabulary; afterwards, the same backbone can be dropped into a standard dense-prediction head and outperform task-specific detectors. On the WideIRSTD benchmark the paper reports an IoU of 42.94 against 36.62 for the strongest visual baseline, Text-IRSTD, a relative improvement it summarizes as over 14%, and on leave-one-dataset-out tests it reports the highest detection probability and IoU on each of three held-out datasets.","pith_inferences":["Because the encoder is grounded in language, a natural extension is zero-shot target discovery, where a user describes a new target type in text and the model segments it without mask supervision for that type.","The same two-stage recipe could transfer to other sparse-object settings where scene context matters, such as tiny lesions in medical images or small vessels in satellite imagery.","The four-option benchmark format could be complemented by open-ended evaluation of the same grounding and counting questions, since a model may score well on option selection while failing to generate a correct coordinate or count.","The task ordering in OmniIRST-VL suggests a curriculum variant: train on scene-oriented tasks first (captioning, physical VQA), then on target-centric tasks (counting, grounding); the paper trains all six tasks as one shuffled mixture."],"forward_implications":["A single JinSight model can be deployed across land, aerial, and space imaging platforms and across NIR, SWIR, and LWIR bands without per-domain retraining.","Text is used only as training-time supervision: at inference the language branch is removed, and detection runs on infrared images alone.","The low-rank interaction is also an efficiency win: in the paper's latent-dimension ablation, setting $r=128$ roughly halves latency relative to the no-LSI baseline while raising IoU.","Both instruction groups contribute to the gain: removing either scene-oriented or target-centric instructions lowers IoU and detection probability, supporting the claim that holistic understanding, not just localization text, drives the improvement.","OmniIRST-VL provides a shared evaluation ground for large vision-language models on infrared understanding, with a four-option single-choice format for discriminative tracks and open-ended caption generation."],"supporting_citations":[{"why":"Supplies WideIRSTD, the 11,000-image visual source whose images, masks, and metadata OmniIRST-VL is built from.","marker":"Li et al. 2024"},{"why":"Provides InternVL2.5-1B, the pretrained generative VLM whose ViT is instruction-tuned in Stage I and later used for dense prediction.","marker":"Chen et al. 2024b"},{"why":"Contributes the ViT-Adapter design that organizes the LSI output into the multi-scale feature pyramid for the detection head.","marker":"Chen et al. 2022"},{"why":"Supplies UperNet, the default dense-prediction decoder that consumes the LSI feature hierarchy.","marker":"Xiao et al. 2018"},{"why":"Text-IRSTD is the strongest visual-only baseline that JinSight must beat on WideIRSTD.","marker":"Huang et al. 2025"},{"why":"Motivates moving from CLIP-style contrastive language to autoregressive generative supervision by showing CLIP behaves like a bag of words.","marker":"Yuksekgonul et al. 2023"},{"why":"Supplies the zero-shot evaluation protocol used to compare large vision-language models on the OmniIRST-VL benchmark.","marker":"Guo et al. 2024"},{"why":"SCTransNet is a transformer-based IRST baseline used as a comparison point in the cross-dataset generalization tests.","marker":"Yuan et al. 2024"},{"why":"DNA-Net appears both as a baseline detector and as one of the decoders tested in the Stage-II ablation.","marker":"Li et al. 2022"}],"fun_headline_variants":["Language-first IRST: understand scene, then detect, up 14% IoU","One model, all infrared domains: vision-language IR detector","Train on words, detect in any infrared band: IoU +14%","New paradigm: 'Understand before Detect' for cross-domain IR targets","JinSight: language-grounded IR detector beats domain-specific models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that language-guided understanding, not the much larger pretrained backbone, is what lets a single detector generalize across infrared domains; the paper reports no same-capacity visual-only control.","fun_headline_variants_meta":{"raw":{"variants":["Language-first IRST: understand scene, then detect, up 14% IoU","One model, all infrared domains: vision-language IR detector","Train on words, detect in any infrared band: IoU +14%","New paradigm: 'Understand before Detect' for cross-domain IR targets","JinSight: language-grounded IR detector beats domain-specific models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000813,"raw_usage":{"total_tokens":3589,"prompt_tokens":998,"completion_tokens":2591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":2495}},"tokens_in":614,"tokens_out":2591,"duration_ms":19714,"temperature":1.0,"reasoning_tokens":2495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:34:03.796928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same InternVL ViT, initialized from the same pretrained weights, directly into the same dense-prediction decoder on the same training split without Stage I language tuning and with a matching training budget; if its IoU on WideIRSTD approaches or exceeds 42.94, the paradigm's claimed contribution is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies UperNet, the default dense-prediction decoder that consumes the LSI feature hierarchy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Text-IRSTD is the strongest visual-only baseline that JinSight must beat on WideIRSTD."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates moving from CLIP-style contrastive language to autoregressive generative supervision by showing CLIP behaves like a bag of words."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the zero-shot evaluation protocol used to compare large vision-language models on the OmniIRST-VL benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SCTransNet is a transformer-based IRST baseline used as a comparison point in the cross-dataset generalization tests."}],"review_version":1}