{"id":"5d331582-234d-423f-bcfb-5ba9b64a9726","arxiv_id":"2607.06552","paper_version":1,"verdict":"CONDITIONAL","confidence":"UNKNOWN","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"MonoIR-RS synthesizes 600K infrared remote-sensing images from visible sources, rewrites captions to be IR-aware, and shows that IR-aware fine-tuning improves CLIP retrieval by up to 12.8 points and drives VLM infrared-cue coverage to 100% with near-zero RGB-color leakage.","lead":"The paper creates a large dataset of synthetic infrared remote-sensing images paired with infrared-aware text captions, then fine-tunes CLIP and VLM models on it. It provides a controlled testbed for aligning infrared imagery with language, useful for nighttime and low-visibility Earth observation.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The VLM captioning claim (100% IR-cue rate) is near-fully circular: the training prompt and evaluation keyword sets overlap almost exactly, and the only semantic metric (class-hit) is near-zero across all tasks.","rationale":"The reader identified the correct load-bearing concern: the VLM lexical evaluation is circular with the training caption style. I sharpen this by noting that for the captioning task specifically, the circularity is near-complete because the training prompt vocabulary (Fig. 4) and evaluation keyword set K_IR (Sec. 5.3) overlap almost exactly. The 100% IR-cue rate is almost guaranteed by construction.\n\nI also note an additional problem the reader did not emphasize: the class-hit rate — the only metric that could indicate semantic understanding — is near-zero across all tasks and backbones (0.2–1.2%, Table 16). The paper acknowledges this is 'conservative by design,' but the practical consequence is that the VLM results contain no positive evidence of correct scene or object recognition. The VLM claims are therefore unsupported beyond style adoption.\n\nOn the CLIP side, the retrieval evaluation is more genuine: bidirectional recall over 9,720 candidates with zero-shot calibration provides real evidence of adaptation. The concern here is narrower — the headline +12.8 gain comes from a single run of ViT-L/14 with an extended stage-2 schedule not used by other backbones (Table 18), making the top-line number not directly comparable to the other four backbones' gains. The seed stability check covers only ViT-B/32.\n\nThe paper is notably honest about these limitations (Sec. 5.3, Sec. 6, Appendix D), which prevents a REJECT. The CLIP results provide sufficient evidence for a conditional acceptance of the dataset and adaptation claims. The VLM claims need human-verified evaluation to be fully supported. CONDITIONAL is the appropriate verdict.","tokens_in":26881,"tokens_out":5900,"duration_ms":226667,"concrete_test":"Sample 200 held-out IR images with VLM-generated captions. Have 3 human annotators judge for each caption: (1) whether each IR-cue term used (e.g. 'high contrast,' 'low-texture region') is actually visible and correctly applied in the image, and (2) whether the scene/object label is correct. If IR-cue term accuracy is below 50% or scene-label accuracy is near chance, the 100% IR-cue rate reflects style reproduction rather than infrared grounding, and the VLM contribution weakens from 'evidence of IR understanding' to 'evidence of style transfer.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies the VLM evaluation circularity, but I think it is sharper than 'partial.' The IR-aware caption generation prompt (Fig. 4) explicitly instructs: 'Explicitly describe infrared-style visual cues, such as grayscale intensity, bright/dark intensity responses, low-texture regions, structural outlines, and contrast.' The evaluation (Sec. 5.3, Eqs. 33–35) checks for keywords K_IR = {infrared, thermal, grayscale, intensity, contrast, texture}. These are essentially the same vocabulary lists. The 100% IR-cue rate for captioning (Table 9, Table 11) is thus almost guaranteed by construction: the model is trained to produce specific terms and then evaluated on whether it produces those terms. This is not a partial circularity; for the captioning task it is nearly complete.\n\nMore importantly, the only metric that could indicate genuine semantic understanding — class-hit rate — is near-zero across all tasks and backbones: 0.4–1.2% for captioning, 0.2–0.6% for scene questions, 0.4–1.0% for object presence (Table 16). The paper dismisses this as 'conservative by design' (Sec. 5.3), but the consequence is that the VLM results provide no evidence of correct scene or object recognition in IR images. The VLM claims thus rest entirely on lexical style adoption, not infrared grounding.\n\nThe CLIP retrieval results are more substantive: the model must distinguish among 9,720 candidate captions, and consistent gains over zero-shot baselines across five backbones reflect real adaptation. However, the headline +12.8 gain (ViT-L/14) is a single run with an extended stage-2 schedule (6 extra epochs at lower LR, Table 18) not applied to other backbones, making the top-line number partly attributable to training budget rather than the IR-aware recipe alone. The seed check (±0.02) covers only ViT-B/32, not the headline backbone.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper introduces MonoIR-RS, a large-scale infrared remote-sensing vision-language dataset and benchmark. The dataset comprises 600,000 synthetic infrared images generated from visible-band remote-sensing sources via DiffV2IR, paired with 59,032 IR-aware caption records rewritten by Qwen2.5-VL-72B to emphasize grayscale structure and thermal contrast rather than RGB appearance. The authors fine-tune five CLIP backbones for contrastive retrieval and six VLM backbones for instruction-following, evaluating against zero-shot baselines. The CLIP retrieval results show consistent gains (+3.5 to +12.8 mR) across all backbones, with a seed-stability check on one backbone. The VLM results are evaluated via lexical proxy diagnostics (IR-cue rate, RGB-color rate, overclaim rate, class-hit rate). The paper also validates synthetic IR realism against the AVIID benchmark (FID 85.2 vs 126.3 for grayscale baseline) and includes an auxiliary paired-retrieval transfer check on AVIID.","tokens_in":27187,"tokens_out":1599,"duration_ms":134217,"significance":"The paper addresses a genuine gap: infrared remote-sensing vision-language resources are scarce, and existing RGB-centric captions are inappropriate for the infrared modality. The CLIP retrieval evaluation is the strongest contribution: it uses a genuine retrieval task with zero-shot baselines providing a non-circular comparison point, consistent gains across five heterogeneous backbones, and a seed-stability check (11.88±0.02 on ViT-B/32). The synthetic-IR realism check against AVIID (Table 2) is reasonable and the auxiliary transfer sanity check (Table 14, +6.5 mR on real-IR queries) provides supporting evidence. The dataset construction pipeline—synthetic IR generation, IR-aware caption rewriting, split hygiene, and RGB-leakage auditing—is reproducible and documented with algorithmic pseudocode and hyperparameter tables. The release package including images, captions, split files, and evaluation scripts is a positive signal for reproducibility.","major_comments":[{"comment":"§5.3, Eqs. (33)–35, Fig. 4, Table 9: The VLM captioning evaluation is substantially circular. The IR-aware caption generation prompt (Fig. 4) explicitly instructs the model to use terms like 'grayscale intensity,' 'bright/dark intensity responses,' 'low-texture regions,' 'structural outlines,' and 'contrast.' The evaluation keyword set K_IR (Appendix A.7) checks for {infrared, thermal, grayscale, intensity, contrast, texture}. These vocabulary lists overlap almost entirely. The 100% IR-cue rate for captioning (Table 9) is thus nearly guaranteed by construction: the model is trained to produce specific terms and then evaluated on whether it produces those same terms. The paper acknowledges this ('a high IR-cue rate indicates infrared-style phrasing rather than verified grounding,' §5.3), but the abstract and conclusion still prominently feature the 100% IR-cue claim as a central result. I","section":null},{"comment":"§5.3, Table 9, Table 16: The only metric that could indicate genuine semantic understanding—class-hit rate—is near-zero across all tasks and backbones (0.4–1.2% for captioning, 0.2–0.6% for scene questions, 0.4–1.0% for object presence, Table 16). The paper dismisses this as 'conservative by design' (§5.3), but the consequence is that the VLM results provide no positive evidence of correct scene or object recognition in IR images. The VLM claims thus rest entirely on lexical style adoption, not infrared grounding. The paper should either (a) add a semantic evaluation with human-verified labels on a subset, or (b) substantially reframe the VLM contribution as a style-transfer diagnostic rather than an infrared understanding result. As it stands, the VLM half of the paper's central claim is not independently supported.","section":null},{"comment":"§4.2, Table 5: The CLIP fine-tuning uses different batch sizes across backbones (768 for OpenAI B/32, 128 for OpenAI L/14, 1024 for OpenCLIP/RemoteCLIP/GeoRSCLIP) and OpenAI ViT-L/14 receives an additional stage-2 resume (6 epochs, lr 5e-7). This makes cross-backbone comparisons in Table 6 not fully controlled: the strongest model (ViT-L/14, 19.2% mR) also has the most favorable training schedule. The paper should acknowledge this confound when interpreting the relative ranking, or ideally re-run with matched effective batch sizes.","section":null}],"minor_comments":[{"comment":"Abstract: 'drives VLM captioning IR-cue coverage to 100%' is technically accurate but misleading given the circularity discussed above. Consider qualifying this claim in the abstract.","section":null},{"comment":"Table 1: The 'Scale' column mixes units (15K pairs, 1.2K images, 5M pairs). Standardizing to a consistent format would improve readability.","section":null},{"comment":"§3.1: The claim that DiffV2IR 'attains strong fidelity (low FID/PSNR/SSIM against real thermal benchmarks)' cites [25] but the paper itself only reports FID and histogram metrics (Table 2). PSNR/SSIM values are not reported in this manuscript.","section":null},{"comment":"Table 8: The supervision ablation uses a 'smaller development retrieval split' that is 'not directly comparable to the formal 9,720-sample test split.' The absolute numbers (35.1, 30.2, 43.0) are therefore not interpretable alongside Table 6. Consider adding the development-split zero-shot baseline for context.","section":null},{"comment":"§5.2: 'This gap shows that the reported scores reflect genuine infrared adaptation rather than residual visible-domain priors' is slightly overstated. The zero-shot baseline uses original pretrained weights on synthetic IR images; the gain could also reflect adaptation to the synthetic IR domain rather than to infrared evidence per se. The AVIID transfer check partially addresses this but is a small-scale paired retrieval.","section":null},{"comment":"Table 7 caption: 'Small pools (RSICD, RSITMD) inflate recall and are not directly comparable across sources' — this is a good caveat but could be stated more prominently, perhaps in the main text rather than only the caption.","section":null},{"comment":"Appendix C.5: The DiffV2IR generation uses seed 1234 for all images. It would be useful to note whether any qualitative or quantitative sensitivity to seed was checked.","section":null},{"comment":"Fig. 1: The workflow diagram is dense and some text is difficult to read. Consider simplifying or enlarging.","section":null},{"comment":"References: Several 2025–2026 references are cited; ensure these are final published versions rather than arXiv preprints where published versions exist.","section":null}],"recommendation":"major_revision","confidential_remarks":"The CLIP retrieval results are solid and would likely meet the bar for publication on their own. The VLM evaluation is the main weakness: the circularity is real and the paper's own acknowledgment of it does not fully resolve the tension between the acknowledged limitation and the prominence of the 100% IR-cue claim in the abstract and conclusion. If the authors reframe the VLM contribution as a diagnostic/style-transfer result rather than an infrared understanding result, and add even a small human-verified semantic evaluation, the paper would be substantially strengthened. The alternative is to de-emphasize the VLM results and center the paper on the dataset and CLIP retrieval, which are the stronger contributions."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. The referee identifies three substantive issues: (1) circularity in the VLM captioning evaluation due to vocabulary overlap between the generation prompt and the IR-cue keyword set, (2) near-zero class-hit rates undermining the VLM semantic-understanding claim, and (3) uncontrolled batch sizes and an extra training stage for ViT-L/14 confounding cross-backbone CLIP comparisons. We agree with all three points in substance. For (1), we will reframe the 100% IR-cue claim as a style-adoption result rather than evidence of infrared grounding, and adjust the abstract and conclusion accordingly. For (2), we will add a human-verified semantic evaluation on a subset and reframe the VLM contribution as a style-transfer diagnostic with a limited semantic check. For (3), we will add an explicit confound acknowledgment and, if compute permits, re-run with matched effective batch sizes. No standing objections remain.","responses":[{"response":"The referee is correct that the vocabulary overlap between the caption generation prompt and the K_IR keyword set makes the 100% IR-cue rate largely guaranteed by construction for the captioning task. We acknowledge that the abstract and conclusion currently present this number as a central result in a way that overstates its evidential value. We will revise the manuscript as follows: (1) The abstract will reframe the 100% IR-cue claim as evidence of style adoption—i.e., that IR-aware fine-tuning reliably shifts VLM output language from RGB-centric to infrared-style phrasing—rather than as evidence of infrared understanding. (2) The conclusion will be revised to match. (3) Section 5.3 will add an explicit statement that the captioning IR-cue rate is near-tautological given the prompt–keyword overlap, and that the meaningful diagnostic signal lies in the zero-shot comparison (Table 11), where IR-cue rates rise from as low as 4.0% to 100% and RGB-color leakage drops to zero across all six backbones. The zero-shot comparison demonstrates that the fine-tuning, not the base model, is responsible for the style shift, which is a non-trivial behavioral result even if it does not establish semantic grounding. We agree this should not be presented as a central understanding result.","revision_made":"yes","referee_comment":"§5.3, Eqs. (33)–35, Fig. 4, Table 9: The VLM captioning evaluation is substantially circular. The IR-aware caption generation prompt (Fig. 4) explicitly instructs the model to use terms like 'grayscale intensity,' 'bright/dark intensity responses,' 'low-texture regions,' 'structural outlines,' and 'contrast.' The evaluation keyword set K_IR (Appendix A.7) checks for {infrared, thermal, grayscale, intensity, contrast, texture}. These vocabulary lists overlap almost entirely. The 100% IR-cue rate for captioning (Table 9) is thus nearly guaranteed by construction: the model is trained to produce specific terms and then evaluated on whether it produces those same terms. The paper acknowledges this ('a high IR-cue rate indicates infrared-style phrasing rather than verified grounding,' §5.3), but the abstract and conclusion still prominently feature the 100% IR-cue claim as a central result."},{"response":"We agree that the near-zero class-hit rates mean the VLM results currently provide no positive evidence of semantic scene or object recognition in IR images. The paper's existing framing of class-hit as 'conservative by design' is accurate but insufficient—it does not address the referee's core point that the VLM contribution lacks an independent semantic evaluation. We will take both actions the referee suggests: (a) We will add a human-verified semantic evaluation on a subset of 200–300 held-out infrared images, where annotators judge whether VLM-generated captions and scene classifications are semantically correct (not just lexically IR-styled). This will be reported as a new table in the revised manuscript. (b) Pending the outcome of that evaluation, we will substantially reframe the VLM contribution as a style-transfer diagnostic with a limited semantic check, rather than claiming infrared understanding. The abstract and conclusion will be revised to match: the VLM results will be described as demonstrating that IR-aware fine-tuning shifts model language toward infrared-style evidence and eliminates RGB-color leakage, with semantic grounding assessed only on the annotated subset. We agree that, as currently written, the VLM half of the paper's central claim is not independently supported by the lexical diagnostics alone.","revision_made":"partial","referee_comment":"§5.3, Table 9, Table 16: The only metric that could indicate genuine semantic understanding—class-hit rate—is near-zero across all tasks and backbones (0.4–1.2% for captioning, 0.2–0.6% for scene questions, 0.4–1.0% for object presence, Table 16). The paper dismisses this as 'conservative by design' (§5.3), but the consequence is that the VLM results provide no positive evidence of correct scene or object recognition in IR images. The VLM claims thus rest entirely on lexical style adoption, not infrared grounding. The paper should either (a) add a semantic evaluation with human-verified labels on a subset, or (b) substantially reframe the VLM contribution as a style-transfer diagnostic rather than an infrared understanding result. As it stands, the VLM half of the paper's central claim is not independently supported."},{"response":"The referee is correct that the differing batch sizes and the additional stage-2 resume for ViT-L/14 introduce a confound in the cross-backbone comparison in Table 6. The strongest model (ViT-L/14, 19.2% mR) does have the most favorable training schedule, and the current manuscript does not acknowledge this confound when interpreting the relative ranking. We will add an explicit acknowledgment in Section 5.2 that the ViT-L/14 result benefits from both a larger model capacity and an extended training schedule, and that the cross-backbone ranking should therefore not be interpreted as a controlled comparison of backbone quality under matched budgets. Regarding the suggestion to re-run with matched effective batch sizes: we will attempt this for at least the two strongest backbones (ViT-L/14 and GeoRSCLIP ViT-B/32) under a common batch size, and report the results as a supplementary table. However, we note that memory constraints on a single GPU (as specified in our setup) may limit the feasible batch size for ViT-L/14, so a fully matched comparison may require gradient accumulation tuning that we cannot guarantee will be ready for the next revision. At minimum, the confound will be explicitly stated in the main text.","revision_made":"partial","referee_comment":"§4.2, Table 5: The CLIP fine-tuning uses different batch sizes across backbones (768 for OpenAI B/32, 128 for OpenAI L/14, 1024 for OpenCLIP/RemoteCLIP/GeoRSCLIP) and OpenAI ViT-L/14 receives an additional stage-2 resume (6 epochs, lr 5e-7). This makes cross-backbone comparisons in Table 6 not fully controlled: the strongest model (ViT-L/14, 19.2% mR) also has the most favorable training schedule. The paper should acknowledge this confound when interpreting the relative ranking, or ideally re-run with matched effective batch sizes."}],"tokens_in":27116,"tokens_out":1603,"duration_ms":140996,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things matter here. First, the CLIP retrieval results are genuine: consistent gains across five backbones with appropriate zero-shot baselines, and the synthetic-IR-vs-grayscale comparison on AVIID is a reasonable modality check. Second, the VLM evaluation is more circular than the paper admits — the 100% IR-cue rate is almost guaranteed by construction, and the only semantic metric (class-hit) is near-zero everywhere. The stress-test note is right to sharpen this from 'partial' to 'nearly complete' circularity for captioning specifically. The caption generation prompt (Fig. 4) explicitly instructs the model to use terms like 'grayscale intensity,' 'contrast,' 'texture,' and the evaluation keyword set (Eqs. 33–35) checks for essentially the same vocabulary. So 100% IR-cue rate tells you the model learned to reproduce the training style, not that it grounds infrared evidence. The paper does acknowledge this in Sec. 5.3, but the headline still foregrounds the 100% number. What is genuinely new and useful: the integrated pipeline of synthetic IR generation, IR-aware caption rewriting, and controlled split hygiene is a real contribution to a gap in remote-sensing VL resources. The supervision ablation (Table 8) showing IR-aware captions improve development retrieval from 30.2% to 43.0% over original captions is clean evidence that the text side matters. The AVIID transfer sanity check (+6.5 mR for real-IR queries) is small but non-trivial. Split hygiene — zero train/test overlap, 280 RGB-named paths flagged and removed — is good practice. Soft spots in proportion: The headline +12.8 gain (ViT-L/14) comes from a run with an extended stage-2 schedule (6 extra epochs at lower LR, Table 18) not applied to other backbones. The seed check (±0.02) covers only ViT-B/32, not the headline backbone. So the top-line number partly reflects training budget, not the IR-aware recipe alone. This is a real confound but not fatal — the other four backbones still show consistent gains without the extended schedule. Most experiments are single-run, which the paper admits. The VLM claims need human-verified labels or a semantic accuracy metric to be substantive. As it stands, the VLM section demonstrates style adoption, not infrared understanding. This paper is for researchers working on remote-sensing vision-language learning who need an IR-focused dataset and benchmark. The CLIP results and dataset construction deserve scrutiny; the VLM evaluation needs a major revision before those claims hold. It deserves a serious referee because the dataset and CLIP pipeline are real contributions, but the VLM evaluation framework needs to be rebuilt with non-circular metrics before publication.","headline":"CLIP retrieval gains are real; VLM evaluation is near-fully circular and the headline +12.8 is confounded by extra training budget","tokens_in":28041,"tokens_out":647,"would_cite":false,"duration_ms":100403,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Rewriting captions for infrared evidence lifts CLIP retrieval by up to 12.8 points","keywords":[],"falsifier":"If IR-aware captions produced no retrieval gain over original captions in the ablation, or if the lexical diagnostics showed no difference between fine-tuned and zero-shot VLMs, the central claim that supervision rewriting is the decisive lever would be unsupported.","tokens_in":26960,"feed_emoji":"","tokens_out":1383,"duration_ms":88559,"temperature":0.7,"pith_summary":"This paper argues that the core bottleneck in infrared remote-sensing vision-language learning is not just a visual domain shift but a supervision shift: captions written for visible-band images describe color and surface appearance that infrared imagery cannot support. The authors construct MonoIR-RS, a dataset of 600,000 synthetic infrared images paired with 59,032 IR-aware captions that rewrite supervision around grayscale structure, intensity contrast, and object-background separation rather than RGB appearance. They then fine-tune five CLIP backbones and six VLM backbones under a train-only protocol, keeping the infrared image as the sole model-facing modality. The central mechanism is the IR-aware caption: by replacing color-centric language with descriptions of what infrared evidence actually shows, the text side of the image-text alignment problem becomes consistent with the visual input. The authors show that this rewrite is load-bearing — in a development ablation, switching from original captions to IR-aware captions raises CLIP retrieval mean recall from 30.2% to 43.0% averaged across five backbones. On the formal filtered test split, IR-aware adaptation yields +3.5 to +12.8 mean-recall gains over zero-shot baselines, with the best model reaching 19.2%. For VLMs, fine-tuning drives the rate of infrared-vocabulary usage in captioning to 100% across all six backbones while reducing visible-color word leakage to zero. The paper also validates that the synthetic infrared images produced by DiffV2IR are closer to real thermal imagery than simple grayscale conversion under FID and histogram-distance metrics on the AVIID benchmark, and reports a small auxiliary transfer check showing the fine-tuned CLIP improves paired retrieval with real infrared queries.","feed_headline":"Rewriting captions for infrared evidence lifts CLIP retrieval by up to 12.8 points","feed_subtitle":"MonoIR-RS isolates the infrared modality and rewrites RGB-centric captions to match what thermal imagery actually shows, proving supervision","key_machinery":"The IR-aware caption — a text description rewritten from visible-centric source captions to describe grayscale intensity patterns, thermal contrast, object-background separation, and scene layout instead of color — is the central object carrying the argument. The DiffV2IR diffusion model generates the synthetic infrared images, and the Qwen2.5-VL-72B model performs the caption rewriting. The evaluation protocol separates CLIP retrieval (bidirectional image-text matching on a filtered 9,720-image test split) from VLM lexical diagnostics (keyword-match rates for infrared terms, color terms, overclaim terms, class tokens, and response length).","core_discovery":"The paper's central finding is that infrared vision-language alignment fails when captions describe RGB evidence the model cannot see, and succeeds when captions are rewritten to match the infrared modality's actual visual evidence. This supervision shift — not just the pixel-level domain shift — is what the authors identify as the decisive factor. The ablation showing IR-aware captions nearly halve the gap to a joint RGB-IR diagnostic while outperforming original captions by 12.8 mean-recall points on development retrieval is the cleanest evidence for this claim. The paper also establishes that synthetic infrared imagery from a diffusion model carries enough modality-specific structure to支持","pith_inferences":[],"forward_implications":["If supervision-text rewriting is the decisive lever, then any modality with a systematic mismatch between visual evidence and caption vocabulary — synthetic aperture radar, hyperspectral, depth maps — could benefit from the same IR-aware-caption approach, not just infrared.","The finding that RGB-pretrained remote-sensing models (RemoteCLIP) transfer poorly to infrared even after fine-tuning suggests that domain-specific RGB pretraining can encode color shortcuts that actively interfere with non-visible-band adaptation.","The modest absolute retrieval scores (best 19.2% mean recall) indicate that current CLIP-scale vision-language models still lack the visual capacity to discriminate many remote-sensing scenes from infrared evidence alone, leaving substantial headroom for architectures designed around intensity-structure rather than color.","The partially circular VLM evaluation — training on captions written to contain infrared vocabulary, then measuring whether outputs contain that vocabulary — means the 100% IR-cue rate may overstate genuine infrared grounding, and future work will need semantic correctness benchmarks to separate vocabulary adoption from visual evidence understanding."],"fun_headline_variants":["IR-aware captions cut RGB-color leakage to near zero in vision-language models","Rewriting captions around infrared evidence lifts CLIP recall by 12.8 points","Infrared vision-language alignment fails when captions describe unseen RGB cues","MonoIR-RS shows supervision shift matters more than pixel shift for infrared CLIP","IR-aware supervision drives VLM captioning coverage to 100% with near-zero RGB leakage"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The VLM evaluation assumes that lexical keyword matching — counting whether model outputs contain infrared vocabulary terms like 'grayscale,' 'intensity,' 'contrast,' 'thermal' — is a meaningful proxy for infrared grounding quality. Because the training captions are specifically written to contain these same vocabulary terms, the 100% IR-cue rate may largely reflect that models learned to reproduce the vocabulary style of their training data rather than correctly grounding红外","fun_headline_variants_meta":{"raw":{"variants":["IR-aware captions cut RGB-color leakage to near zero in vision-language models","Rewriting captions around infrared evidence lifts CLIP recall by 12.8 points","Infrared vision-language alignment fails when captions describe unseen RGB cues","MonoIR-RS shows supervision shift matters more than pixel shift for infrared CLIP","IR-aware supervision drives VLM captioning coverage to 100% with near-zero RGB leakage"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":732,"prompt_tokens":631,"completion_tokens":101,"prompt_tokens_details":null},"tokens_in":631,"tokens_out":101,"duration_ms":41357,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T01:51:42.973010+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If IR-aware captions produced no retrieval gain over original captions in the ablation, or if the lexical diagnostics showed no difference between fine-tuned and zero-shot VLMs, the central claim that supervision rewriting is the decisive lever would be unsupported.","supporting_citations":[],"review_version":1}