{"id":"1c48aecd-ec7e-43b8-be90-88dcbcd1ce80","arxiv_id":"2607.03143","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Aligning images to multi-view caption cores while suppressing orthogonal residual text and disagreement-aware temperature improves robust zero-shot recognition and LVLM transfer.","lead":"TPC trains vision–language models by treating captions as incomplete: align images to a multi-view semantic core, suppress leftover “unsaid” residual directions, and soften confidence when captions disagree. That framing yields stronger clean and adversarial zero-shot accuracy and better LVLM hallucination metrics under a fixed LLaVA stack.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Synthetic multi-view construction may mis-define the consensus core and residual, so headline gains may not isolate the partial-constraint principle.","rationale":"The reader correctly flags the faithfulness of the multi-view consensus and synthetic views as the weakest assumption. That is load-bearing: without a faithful core/residual split, the method reduces to multi-text contrastive training plus a linear filter, and the partial-constraint story no longer explains the margins over Adv-W2S and other robust baselines. Human multi-caption diagnostics (Fig. 2, Fig. 7–8) and theory are supportive when views are genuine, but the paper’s own training recipe relies on synthetic views for single-caption corpora, and no experiment disentangles that recipe from the claimed mechanism. A human-only vs. synthetic-view retrain is a concrete, decisive check; until then CONDITIONAL remains appropriate rather than ACCEPT or REJECT. I agree with the reader’s diagnosis and do not raise a stronger independent objection.","tokens_in":28191,"tokens_out":657,"duration_ms":6348,"concrete_test":"Retrain TPC and a matched multi-caption CLIP baseline under two regimes with identical steps/batch: (A) human multi-captions only (COCO/Flickr, no synthesis); (B) web-scale with B.2 synthesis. Report ImageNet clean/robust, Avg-14, residual leakage, and controlled-deletion Recall@1/ECE (Fig. 11). If the TPC–baseline gap shrinks by >50% under (A) or if synthesis-only TPC loses attribute/relation sensitivity more than human-only TPC, the load-bearing assumption fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that multi-view consensus ci (Eq. 2) is a faithful latent-core target and that residuals ri,k (Eq. 5) are truly “unsaid” rather than task-relevant detail. On human multi-caption sets this is plausible. On single-caption web-style data—which drives large-scale dual-encoder training—Appendix B.2 synthesizes views via core-only deletion, paraphrase, and back-translation, filtered only by frozen-encoder cosine ≥0.25 and no new named entities. That filter is weak: it can keep paraphrases that drop attributes/relations/counts the model later needs, or keep surface variants that still share residual directions. Then Lcore (Eq. 7) and Lnc (Eq. 6) can reward ignoring useful detail while Lagree (Eq. 8) teaches ψθ to reproduce an artifact mean. Theory (Thm. 5.1–5.2) assumes views are noisy observations of a fixed μi; synthetic deletion systematically removes structure rather than adding zero-mean noise, so the DRO/denoising justification does not transfer. Ablations in Table 3 and residual-leakage Fig. 5 do not isolate human-only vs. synthetic-view training, so the reported 81.42/64.05 and LVLM gains may partly reflect multi-text augmentation rather than core–residual alignment.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Text as Partial Constraint (TPC), a core–residual vision–language alignment framework that treats multi-view captions as incomplete supervision. It defines a multi-view consensus core ci (Eq. 2) as the alignment target, trains a linear single-view core filter ψθ (Eq. 4) for one-text inference, penalizes image–residual correlation via Lnc (Eqs. 5–6), and softens InfoNCE with disagreement-based temperature τi=τ0(1+γui) (Eqs. 3, 7). Theory motivates residual non-commitment as a DRO-style worst-case bound (Thm. 5.1) and multi-view consensus as denoising (Thm. 5.2). Empirically, TPC reports 81.42/64.05 ImageNet clean/robust Top-1, 76.19/52.03 on Avg-14, and improved LLaVA-1.5-7B transfer (85.16 POPE F1, 59.57 OKVQA), with ablations, residual-leakage, risk–coverage, caption-sensitivity, controlled-deletion, and backbone analyses.","tokens_in":28640,"tokens_out":1374,"duration_ms":25749,"significance":"If the gains isolate the partial-constraint principle rather than multi-text augmentation, the work is a clear and practical contribution to robust open-vocabulary alignment and LVLM grounding under underspecified language. Strengths include a diagnostic preliminary study (Sec. 3), single-text deployability via ψθ, multi-front empirical stress tests (residual leakage Fig. 5, risk–coverage Fig. 6, controlled deletion Fig. 11, caption-choice sensitivity Fig. 7), seed variability (Table 5), and compute disclosure. The theory is motivational rather than machine-checked, but Thm. 5.1 cleanly links core alignment and residual sensitivity to worst-case similarity. The main significance risk is whether headline numbers credit core–residual alignment or simply richer multi-view text supervision.","major_comments":[{"comment":"The central claim that gains come from modeling text as a partial constraint is not isolated from multi-text augmentation. Table 3 removes TPC components but does not include a matched multi-view baseline that uses the same K captions (human or synthetic) with standard multi-positive contrastive loss and no core/residual machinery. Without that control, the ImageNet/Avg-14/LVLM improvements (Tables 1–2) may partly reflect extra text views rather than consensus-core alignment and residual non-commitment (Eqs. 6–9).","section":"§6.2–6.3, Tables 1–3"},{"comment":"Appendix B.2 synthesizes views for single-caption web data via core-only deletion, paraphrase, and back-translation with a weak frozen-encoder filter (cosine ≥0.25). Thm. 5.2 models views as zero-mean noise around a latent core μi, but systematic deletion removes structure rather than adding noise, so the denoising justification does not transfer to the large-scale regime that drives dual-encoder training. The paper should either (i) report human-multi-caption-only vs. synthetic-view training splits for residual leakage and main metrics, or (ii) substantially strengthen faithfulness checks and explicitly limit the theory’s scope.","section":"Appendix B.2; §5.2, Thm. 5.2; Eqs. 2–6"},{"comment":"Training data composition, step budget, and compute matching for the headline ImageNet/Avg-14 numbers versus robust baselines (TeCoA, FARE, Adv-W2S) remain underspecified beyond “match when feasible” (§6.1, B.4–B.5). Because adversarial methods incur different costs and TPC’s multi-view text path changes supervision density, a compute- and data-matched block (or clear external-checkpoint labeling) is needed before attributing +5–7 point robust gains primarily to core–residual alignment.","section":"§6.1, Table 1; Appendix B.4–B.5"}],"minor_comments":[{"comment":"Adversarial radius is inconsistent: Table 1 caption states AutoAttack ϵ=2/255, while Appendix B.3 defaults to ϵ=4/255. Please unify the protocol and restate which radius was used for all methods.","section":"Table 1; Appendix B.3"},{"comment":"Title/method casing is broken in several places (“TEXT ASPARTIAL CONSTRAINT”, “TEXT ASPARTIALCONSTRAIN”). Normalize to “Text as Partial Constraint (TPC)” throughout.","section":"Abstract; §1; §4; §6"},{"comment":"Figure 4 axis labels show encoding artifacts (“/glyph1197umber of views”, “/glyph1197on-com m itm ent”). Clean the PDF text layer.","section":"Fig. 4"},{"comment":"Related Work cites three concurrent ICLR 2026 papers (Xiao et al., Zhang et al., Zhou et al.) as “adjacent”; a one-sentence differentiation of partial-constraint alignment vs. continual/compositional adaptation would help readers.","section":"§2"},{"comment":"In §4.6, state explicitly that τ(y) calibration preserves ranking while only changing softmax sharpness; this is said once but should be repeated near the LVLM transfer protocol so readers do not confuse confidence calibration with re-ranking.","section":"§4.6; §6.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is above the usual conference bar in diagnostic thoroughness, but the missing multi-view non-TPC baseline is the single issue most likely to change how much credit the community assigns the method. If the authors can add that control (even on COCO/Flickr + a smaller web subset) and clarify data/compute matching, this could become a strong accept on revision. Citation cluster of same-group 2026 ICLR papers is noticeable but not disqualifying if differentiation is clear."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: treat multi-caption variation as incomplete supervision, align images to a consensus core, learn a single-view core filter for one-text inference, and actively suppress residual commitment while softening contrastive sharpness when views disagree. That package is coherent and new enough as a training principle, not just another robust-CLIP recipe.\n\nWhat they do well is the multi-axis evidence. Clean/robust ImageNet and Avg-14 numbers beat strong adversarial baselines by clear margins; LLaVA drop-in transfer improves POPE and OKVQA; ablations hit every term; residual-leakage and risk–coverage diagnostics actually speak to the claimed mechanism; controlled deletion and caption-sensitivity tests are the right stress. The DRO-style residual bound and multi-view denoising PAC argument are motivational rather than load-bearing, but the proofs are clean and match the design. Citations cover CLIP-scale training, robust VL, and LVLM hallucination without obvious gaps.\n\nThe soft spot is real but not fatal. On human multi-caption data the consensus/residual story is plausible. On single-caption web data they synthesize views by core-only deletion, paraphrase, and back-translation with a weak frozen-encoder filter (≥0.25). That can erase task-relevant attributes and turn “unsaid residual” into useful detail the model is told to ignore. Theory assumes zero-mean noise around a latent core; systematic deletion is not that. They never isolate human-only vs synthetic-view training, so some of the headline lift may be multi-text augmentation rather than pure core–residual alignment. Hyperparameters are free but sensitivity plots look flat enough. No public code/data yet, so the large margins stay unverified.\n\nThis is for people who train dual encoders or swap vision towers into LVLMs and care about paraphrase brittleness and overconfident grounding. It is not a full safety solution and may trade some fine-grained attribute sensitivity. I would still send it to referees: the problem is real, the method is clean, and the empirics are broad enough to deserve scrutiny rather than a desk reject. Engage if you work on robust VL or LVLM grounding; otherwise skim the method figure and Table 1–3.","headline":"Solid dual-encoder recipe that treats captions as partial constraints; gains look real, but synthetic multi-view construction is the soft underbelly.","tokens_in":29201,"tokens_out":535,"would_cite":true,"duration_ms":6500,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Natural captions are incomplete constraints: align vision to the consensus core, suppress residual over-commitment, and confidence follows how strongly text pins the image.","keywords":["vision-language alignment","partial language constraints","core-residual decomposition","multi-view consensus","uncertainty-aware contrastive learning","zero-shot robustness","LVLM hallucination"],"falsifier":"Train and evaluate with only genuine multi-caption human data versus the paper’s synthesized-view recipe on single-caption corpora; if the clean/robust ImageNet and Avg-14 margins and residual-leakage reduction disappear when synthetic views are removed or replaced by unfiltered generation, the partial-constraint claim does not hold under realistic supervision.","tokens_in":29086,"feed_emoji":"🔗","tokens_out":707,"duration_ms":6537,"temperature":0.7,"pith_summary":"Vision–language models trained on natural captions treat each description as a complete target, yet real captions omit details, paraphrase freely, and leave much unsaid. That mismatch makes similarity brittle under rewording and overconfident when language is weak. This paper argues that multi-view captions of the same image expose a shared semantic core and an orthogonal “unsaid” residual, and that robust alignment should match images only to the core while explicitly discouraging residual commitment. TPC distills a multi-view consensus core as the training target, learns a lightweight single-view core filter so standard one-query inference still works, penalizes image–residual correlation, and softens contrastive sharpness when caption views disagree. The resulting representations improve clean and adversarial zero-shot accuracy and transfer as a drop-in vision encoder that reduces hallucination in large vision–language models. A sympathetic reader cares because underspecified language is the default in open-vocabulary recognition, retrieval, and interactive multimodal systems; calibrating alignment to partial constraints is a practical route to more reliable matching.","feed_headline":"Captions are partial: align only the consensus core","feed_subtitle":"Suppressing the unsaid residual lifts clean and robust accuracy and cuts LVLM hallucination","key_machinery":"Core–residual alignment (TPC): consensus core ci as the normalized mean of multi-view text embeddings; single-view core filter ψθ; unsaid residual ri,k orthogonal to the predicted core; non-commitment loss that squares image–residual correlation; and uncertainty-aware InfoNCE with temperature τi = τ0(1 + γui).","core_discovery":"Treating multi-view natural captions as incomplete supervision—aligning images to a consensus semantic core, predicting that core from a single caption at test time, suppressing correlation with the orthogonal unsaid residual, and scaling contrastive temperature by caption-view disagreement—yields vision–language representations whose matching is more stable under paraphrase and whose confidence better tracks how strongly text constrains the image, with measured gains on zero-shot recognition, adversarial robustness, and LVLM transfer.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Align images to consensus core, suppress unsaid residual","Text as partial constraint: distill multi-view semantic core","Core-residual VL alignment for stable caption matching","Treat captions incomplete: predict core, scale by view disagreement","Suppress residual dependence for robust open-vocab VL reps"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The normalized average of several captions of the same image is a faithful target for the shared meaning, and the extra captions built for single-caption web data (shortened cores, paraphrases, back-translations) truly expose underspecification rather than train on artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Align images to consensus core, suppress unsaid residual","Text as partial constraint: distill multi-view semantic core","Core-residual VL alignment for stable caption matching","Treat captions incomplete: predict core, scale by view disagreement","Suppress residual dependence for robust open-vocab VL reps"]},"model":"grok-4.5","effort":"low","cost_usd":0.004416,"raw_usage":{"total_tokens":1338,"prompt_tokens":807,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":44160000,"prompt_tokens_details":{"text_tokens":807,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":467,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":807,"tokens_out":64,"duration_ms":5947,"temperature":1.0,"reasoning_tokens":467,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T04:34:43.417194+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate with only genuine multi-caption human data versus the paper’s synthesized-view recipe on single-caption corpora; if the clean/robust ImageNet and Avg-14 margins and residual-leakage reduction disappear when synthetic views are removed or replaced by unfiltered generation, the partial-constraint claim does not hold under realistic supervision.","supporting_citations":[],"review_version":1}