{"id":"d6373b54-a31e-4913-8d18-32df9dfbcee9","arxiv_id":"2607.09443","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Continuous metadata conditioning of CLIP prompts during training improves longitudinal animal ReID under identity and temporal shifts without needing metadata at test time.","lead":"This paper adapts CLIP for long-term animal re-identification by injecting continuous numerical metadata (like capture year or body length) into training prompts only. The result is better robustness to growth and seasonal change while keeping inference purely visual and parameter-efficient.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper’s strongest claim is empirical and systems-level: continuous numerical metadata inside the prompt during training improves longitudinal ReID while leaving a purely visual inference pipeline. The reader correctly flags the non-obvious transfer from text-side conditioning to visual geometry after text is removed, yet the multi-protocol ablations, continuous-vs-discrete comparison, cross-dataset results, and public code already address that transfer. No hidden assumption (e.g., identity-token leakage, inference-time metadata, or discrete binning) is required for the reported numbers. Sparse TACS/TAOS statistics and attribute-selection freedom are real but do not overturn the claim; they are the expected cost of ecological longitudinal data. Consequently the ACCEPT verdict with high confidence remains appropriate; no adjustment is warranted.","tokens_in":32394,"tokens_out":512,"duration_ms":7716,"concrete_test":"Re-train the exact Ours + Meta (Year+FiLM) and DaySeq+Sin configurations of Tables 4–5 five times with the released code and splits, then recompute mAP on TACS/TAOS head and body; if the continuous-vs-base gap falls below ~1 mAP or loses statistical separation from the discrete baseline of Tables 6–7, the geometric-modulation claim would need qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption is real but not load-bearing against the central claim. Continuous metadata is injected only into the frozen CLIP text prompt (sinusoidal PE or FiLM on attribute basis tokens, Sec. 3.4); the visual encoder is adapted via LoRA + dual-space triplet + auxiliary CE + symmetric CCE, then all text/metadata are discarded at inference. Tables 4–7 show consistent gains of continuous over discrete and over the no-meta baseline under TACS/TAOS on Melops and SeaTurtleID2022; ablations (Tables 8–11) isolate LoRA, prompt ensembling, identity tokens, and dual-space supervision. Temporal-decay analysis (Fig. 3) further supports that appearance drift is continuous, matching the design choice. Sparse time-aware query counts (Table 14) and free choices of which attribute/encoding wins are acknowledged limitations of ecological data and of the recipe, not internal contradictions. The claim that continuous training-time conditioning yields a stronger purely visual embedding therefore stands on the reported evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a parameter-efficient CLIP adaptation framework for longitudinal animal re-identification. It jointly trains LoRA modules in a frozen ViT vision encoder with learnable prompt context tokens and symmetric cross-modal alignment, while introducing continuous metadata conditioning (sinusoidal encoding or FiLM modulation of numerical attributes such as capture year, day-sequence, and body length) that is injected into the prompt representation only during training. All text and metadata components are discarded at inference, yielding a purely visual nearest-neighbor pipeline. The primary claimed contribution is that preserving the continuous structure of numerical metadata, rather than discretizing it into textual bins, induces smooth geometric modulation of the embedding space that improves robustness under identity and temporal distribution shifts. Evaluation uses a seven-year Melops corkwing-wrasse dataset under closed-set, open-set, time-aware closed-set, and time-aware open-set protocols, plus transfer to multiple wildlife benchmarks including SeaTurtleID2022.","tokens_in":32708,"tokens_out":1044,"duration_ms":28925,"significance":"If the empirical claims hold, the work supplies a practical recipe for ecological ReID: metadata available only at training time can still improve a deployable visual model without inference-time dependencies or large architectural additions. The continuous-versus-discrete comparison, multi-protocol design (including explicit temporal separation), dual-space triplet and auxiliary supervision ablations, prompt-layout ensembling, and multi-seed reporting constitute a thorough experimental package. Public code and evaluation splits further raise the contribution's value. Parameter counts (approximately 73 M trainable versus 150 M+ for full CLIP fine-tuning or prior CLIP-ReID/IndivAID pipelines) while matching or exceeding full fine-tuning on Melops closed-set body crops are a concrete efficiency gain. The longitudinal Melops setting itself, with multi-year growth and seasonal change, is a useful stress test for the community.","major_comments":[{"comment":"Tables 4–5 and Table 14 (TACS/TAOS): the time-aware protocols rest on very small query sets (37 identities / 76 query images for TACS; 86 identities / 173 query images for TAOS) and short observed temporal spans (roughly 1–113 days). While the paper correctly notes that this sparsity is inherent to capture–mark–recapture ecology, the absolute mAP/Rank gains attributed to continuous metadata conditioning under temporal shift therefore rest on limited statistical power. A bootstrap or query-level uncertainty analysis (beyond five random seeds) would make the longitudinal robustness claim more secure.","section":null},{"comment":"Sec. 4.5.2 and Tables 6–7: the best-performing metadata attribute and encoding (Year+FiLM versus DaySeq+Sin, single versus multi-attribute) varies by protocol and crop type, and the main comparison tables report a single chosen configuration “for clarity.” The manuscript should state explicitly whether these configurations were selected a priori, by validation, or post hoc, and should report the full range of continuous-conditioning outcomes (or a pre-specified default such as capture year) so that the claimed advantage of continuous over discrete conditioning cannot be read as selective reporting.","section":null}],"minor_comments":[{"comment":"Sec. 3.7: the sentence ending “…to obtain the final descriptor. the Euclidean distance…” has a capitalization error and should be cleaned.","section":null},{"comment":"Figure 3 caption and Sec. 4.5.9: the logistic-regression slopes are reported as ranges; adding the exact fitted coefficients (or a small table) would aid reproducibility of the temporal-decay analysis.","section":null},{"comment":"Table 12: “Ours + Meta (All 3 +FiLM)” jumps to 98.20 M parameters; a one-sentence explanation of which FiLM components drive the increase would help readers judge the efficiency claim under the heaviest metadata setting.","section":null},{"comment":"Sec. 2.4 and related-work discussion of MetaWild [29]: the discrete textual baseline is a useful control, but a short clarification that the discrete bins were defined on the whole dataset (as stated later) versus training-set only would remove any residual leakage concern.","section":null},{"comment":"Throughout: “Re-ID” / “ReID” and “vision–language” / “vision-language” hyphenation is inconsistent; a single house style would improve polish.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Solid applied CV / ecological-monitoring paper with public code. The continuous-metadata idea is incremental but cleanly executed and well ablated; the sparse TACS/TAOS query counts are the main soft spot and are already partially acknowledged. Suitable for Expert Systems with Applications after the two major points are addressed in revision. No integrity or novelty-disclosure concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: keep numerical metadata continuous inside the CLIP prompt only during training (sinusoidal or FiLM), adapt the vision tower with LoRA plus dual-space triplet and symmetric cross-modal loss, then throw away text and metadata at test time. That yields a purely visual retrieval model that is more robust to multi-year appearance drift than CLIP-ReID or IndivAID on Melops and transfers cleanly to other wildlife sets.\n\nWhat is actually new is not LoRA or CoOp-style prompts—those are standard—but the refusal to discretize year/day-sequence/length into text bins the way MetaWild does, plus the explicit claim that the continuous signal shapes the visual geometry enough to survive after the text branch is discarded. The paper backs this with the right experiments: CS/OS plus time-aware closed- and open-set splits, head vs body crops, continuous-vs-discrete ablations (Tables 6–7), adapter/prompt/loss ablations (Tables 8–11), temporal-decay curves (Fig. 3), and SeaTurtleID2022 transfer with the same recipe. Code and splits are public. Parameter counts are honest (~73M trainable vs ~150M full FT).\n\nSoft spots are real but proportional. Time-aware query sets are small (Table 14) because long-term recaptures are sparse; that is ecology, not a design flaw, but it limits statistical power. Which attribute and which encoder (year+FiLM vs dayseq+sin) wins is somewhat free, and several free parameters (rank, M, loss weights) are fixed by convention rather than swept. None of that overturns the central claim: continuous conditioning beats discrete and beats no-meta under temporal shift, and the inference pipeline stays visual-only.\n\nThis is for people who actually ship wildlife ReID systems or who care about PEFT under longitudinal shift. It will not reorganize vision, but it is a clean, citable engineering result with honest evaluation. I would send it to peer review without hesitation; a serious referee will tighten the free-parameter discussion and the sparse-split caveats, not reject the contribution.","headline":"Solid, reproducible PEFT recipe for longitudinal animal ReID; continuous training-time metadata in the prompt is the real (if modest) novelty and the evidence holds.","tokens_in":33305,"tokens_out":543,"would_cite":true,"duration_ms":7835,"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":"Continuous numerical metadata, used only while training, makes animal re-identification more robust to years of growth and seasonal change without needing that metadata at test time.","keywords":["Animal Re-Identification","Vision-Language Models","Prompt Learning","Low-Rank Adaptation","Continuous Metadata Conditioning","Longitudinal Evaluation","CLIP"],"falsifier":"On a held-out multi-year recapture set, retrain the identical architecture but replace continuous metadata injection with random noise or with no metadata at all; if closed-set and time-aware open-set mAP remain statistically indistinguishable from continuous conditioning, the claim that continuous structure is what drives the improvement fails.","tokens_in":33317,"feed_emoji":"🐟","tokens_out":597,"duration_ms":7962,"temperature":0.7,"pith_summary":"Long-term animal re-identification must match the same individual across months or years even as size, coloration, and life stage drift. This paper argues that a frozen CLIP vision-language backbone can be adapted for that task with far fewer trainable parameters by combining low-rank visual updates, learnable prompt tokens, and cross-modal alignment. Its central claim is that continuous numerical metadata (capture time, body length, year) should be injected directly into the prompt as smooth embeddings during training rather than being discretized into text bins or fused through extra modules. Because the metadata never appears at inference, the deployed system remains a pure visual nearest-neighbor matcher. On a seven-year corkwing-wrasse dataset and several other wildlife benchmarks, the continuous-conditioning strategy improves closed-set, open-set, and time-aware accuracy over prior CLIP-based ReID methods while keeping the model lightweight.","feed_headline":"Metadata trains animal ReID without being needed at test","feed_subtitle":"Continuous year and size cues reshape embeddings so pure vision still works years later","key_machinery":"Continuous metadata-conditioned prompt learning: real-valued attributes are encoded by sinusoidal projection or FiLM modulation and appended as a token to identity-conditioned CLIP prompts during training, then completely removed at inference so that only the LoRA-adapted visual encoder is used for retrieval.","core_discovery":"Preserving the continuous structure of numerical metadata and injecting it into the prompt representation only during training produces a smoother geometry in the shared embedding space that remains useful for pure visual retrieval after every text and metadata component is discarded, improving robustness to longitudinal appearance change and temporal distribution shift.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Continuous metadata conditions animal ReID for pure vision later","Metadata shapes embeddings so animal ReID works years later visually","Train with continuous year-size cues then drop them for animal ReID","Continuous prompt conditioning boosts long-term animal re-identification","Metadata trains smoother animal ReID space then leaves pure vision"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The assumption that continuous metadata injected only into the frozen text prompt during training will permanently reshape the visual embedding space so that the same gains remain after metadata and all text components are thrown away at test time.","fun_headline_variants_meta":{"raw":{"variants":["Continuous metadata conditions animal ReID for pure vision later","Metadata shapes embeddings so animal ReID works years later visually","Train with continuous year-size cues then drop them for animal ReID","Continuous prompt conditioning boosts long-term animal re-identification","Metadata trains smoother animal ReID space then leaves pure vision"]},"model":"grok-4.5","effort":"low","cost_usd":0.003578,"raw_usage":{"total_tokens":1159,"prompt_tokens":748,"num_sources_used":0,"completion_tokens":86,"cost_in_usd_ticks":35780000,"prompt_tokens_details":{"text_tokens":748,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":325,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":748,"tokens_out":86,"duration_ms":5394,"temperature":1.0,"reasoning_tokens":325,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T02:57:21.283396+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out multi-year recapture set, retrain the identical architecture but replace continuous metadata injection with random noise or with no metadata at all; if closed-set and time-aware open-set mAP remain statistically indistinguishable from continuous conditioning, the claim that continuous structure is what drives the improvement fails.","supporting_citations":[],"review_version":1}