{"id":"842cbe08-0ebc-4974-82ce-0bdd46d44284","arxiv_id":"2510.13808","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding layer-wise learnable visual probes to a frozen VLM vision encoder improves domain adaptation across egocentric, depth, and robot-control domains while largely retaining source-domain performance.","lead":"VisCoP adapts vision-language models to new visual domains by adding small trainable 'visual probes' that read intermediate layers of a frozen vision encoder. The paper claims better target-domain performance with less forgetting than retraining the vision encoder.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ego/depth target evaluations likely overlap training data: no split reported between EgoExo4D training subset and Ego-in-Exo PerceptionMCQ, so reported Δtarget may reflect memorization rather than adaptation.","rationale":"The reader's weakest assumption identifies the same concern. This is the single most load-bearing issue because it directly undermines the ego and depth adaptation results, which are two of the three claimed scenario wins. The remaining robot-control results are less affected, but the paper's headline claim covers all three. No split is reported, and the manuscript text itself highlights that both training and evaluation derive from EgoExo4D/[42]; this is an omitted support that should have been disclosed. Although other issues exist (no error bars, small real-robot trials), they are secondary. If the overlap test shows disjoint sets and no question leakage, the verdict could be revisited; absent that evidence, rejection is appropriate.","tokens_in":20174,"tokens_out":3346,"duration_ms":28111,"concrete_test":"Request or reconstruct the exact split: obtain EgoExo4D video IDs for the training keystep subset (§5.2) and for all Ego-in-Exo PerceptionMCQ evaluation videos (§5.2.1). Compute the intersection. If non-empty, rerun Table 1 and Table 2 with a strictly disjoint evaluation set (e.g., a held-out EgoExo4D split or EgoSchema as the only target benchmark) and recompute Δtarget and Δsource. Also compare the training instruction pairs from [42] against the PerceptionMCQ QA pairs for exact or near-duplicate text; if duplicates exist, the evaluation protocol is confounded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of superior target-domain adaptation rests on ego and depth results (Tables 1–2). Training uses 24,688 EgoExo4D keystep videos, with 74,064 instruction pairs 'recaptioned from [42]' (§5.2). Evaluation uses Ego-in-Exo PerceptionMCQ, which is also derived from EgoExo4D and comprises 3,991 video-QA pairs from [42] (§5.2.1). The paper reports no train/test split or video-ID disjunction; nothing prevents the evaluation videos or QA pairs from being drawn from the same keystep subset used for training. Because the training objective is next-token prediction on exactly these QA formats, overlap would let VisCoP (and the baselines) memorize answers rather than learn transferable domain features. The magnitude of the ego gain (+3.5 Δtarget) is small enough to be explained by memorizing a fraction of the 3,991 benchmark items, and the depth gain (+19.27) is even more suspect if the depth maps are derived from the same RGB videos. Without a disclosed split, the cross-view and cross-modal superiority claims are unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VisCoP, a parameter-efficient domain-adaptation method for video VLMs. VisCoP augments a frozen vision encoder with a compact set of learnable visual probes and per-layer interaction modules that cross-attend to intermediate visual features. The resulting probe embeddings are fed to the LLM through a dedicated connector while the LLM is updated with LoRA (or fully, in the robot-control setting). The method is evaluated on three adaptation scenarios: exocentric-to-egocentric video understanding, RGB-to-depth video understanding (depth maps generated by DepthAnythingV2), and human-action-to-robot-control. Performance is reported as Δtarget and Δsource, the accuracy differences between the adapted expert and the base VLM on target and source benchmarks, respectively. The paper claims that VisCoP achieves superior target-domain performance while better retaining source-domain knowledge than alternative component-wise finetuning strategies.","tokens_in":20444,"tokens_out":8247,"duration_ms":64551,"significance":"The conceptual idea of extracting domain-specific features through layer-wise probing of a frozen vision encoder is interesting and potentially useful as a PEFT recipe, and the robot-control experiments on VIMA-Bench and a new real-robot dataset provide a partially external evaluation. However, the paper's headline cross-view and cross-modal claims rest almost entirely on Ego-in-Exo PerceptionMCQ, a benchmark created by the same authors from the same EgoExo4D data used for training, with no disclosed train/test split and with training instruction pairs recaptioned from the same source. This creates a serious risk that the reported target-domain gains are memorization rather than adaptation. The central claim is therefore not currently substantiated, although the issue is fixable by reporting a clean split or by re-evaluating on external benchmarks.","major_comments":[{"comment":"The ego target evaluation uses Ego-in-Exo PerceptionMCQ, derived from EgoExo4D and comprising QA pairs from [42]. Training uses 24,688 EgoExo4D keystep videos and 74,064 instruction pairs 'recaptioned from the instruction pairs provided in [42]'. No video-ID disjunction or QA-pair deduplication between training and evaluation is reported. Because the training objective is next-token prediction on the same QA format, overlap would allow memorization of answers. The reported Δtarget=+3.53 on this benchmark is therefore not interpretable as evidence of adaptation. Please report the exact disjointness of videos and QA pairs, and provide results on a held-out EgoExo4D split or an external ego-video benchmark.","section":"§5.2, §5.2.1, Table 1"},{"comment":"The depth target is evaluated only on Ego-in-Exo PerceptionMCQ (Exo Depth), generated by running DepthAnythingV2 on the same EgoExo4D-derived videos used for training, with no split reported. The +19.27 Δtarget could thus reflect memorization of the 3,991-item benchmark rather than cross-modal adaptation. Furthermore, because both training and evaluation depth maps come from the same monocular depth estimator, the setting is a synthetic depth proxy rather than a genuine RGB-to-depth sensor shift. A held-out split or an independent depth benchmark is needed to support the cross-modal claim.","section":"§5.2.2, Table 2"},{"comment":"The robot-control experiments on VIMA-Bench and xArm-Det are more credible and provide some external validation. However, they only cover the cross-task scenario; they do not substantiate the paper's central claims about cross-view and cross-modal adaptation, which rely on the contaminated ego/depth benchmarks. The authors should either provide clean ego/depth evaluations or substantially qualify the central claim.","section":"§5.2.3, Tables 3–4"}],"minor_comments":[{"comment":"The probe update equation writes 'Pℓ = softmax(...)' on the left-hand side, but from the preceding definition Pℓ+1 = Φℓ(Pℓ,Xℓ), the updated probes should be Pℓ+1. Please fix the subscript.","section":"§4.2, Eq. (3)"},{"comment":"The sentence 'in some settings, such as robotic control, we restrict vision-probe cross-attention to spatial tokens only' is underspecified. Which settings and why? Please clarify for each experiment.","section":"§4.2"},{"comment":"Ablation results are reported without variance or multiple seeds. Differences such as QFormer-Style (+0.56) versus VisCoP (+3.53) could be within run-to-run noise. Please report the number of seeds and variance, or at least state the experimental protocol.","section":"§5.3, Table 5 and Figures 3–4"},{"comment":"Reference [55] lists the authors in the wrong order and incorrectly capitalizes the title. Please check the citation against the original paper.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The train/eval overlap concern is serious and directly affects the paper's two headline experiments. The robot-control results give some confidence that the method has merit, and the issue is fixable by disclosing splits or re-running on external benchmarks. I therefore recommend major revision rather than rejection, but the authors must address the overlap explicitly before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. VisCoP is a genuine architectural idea — learnable probes that cross-attend to intermediate layers of a frozen vision encoder — and the ablations show that intermediate access matters. But the paper's headline claim on ego and depth adaptation rests on a benchmark that probably overlaps the training data, and the paper never addresses it.\n\nWhat's new: the layer-wise probing design is different from Q-Former's single-layer compression, and the comparison in Table 5 makes that point fairly. The probe-count and placement studies are useful. The method does preserve source performance on external benchmarks like NeXTQA, VideoMME, and ADL-X, which is not trivial.\n\nThe problem: training uses 24,688 EgoExo4D keystep videos with instructions recaptioned from [42]; evaluation uses Ego-in-Exo PerceptionMCQ, which is also from [42] and derived from the same EgoExo4D pool. No split is reported. Nothing in the paper rules out that the same videos or even the same QA pairs appear in both. If so, the +3.5 ego gain and the +19.3 depth gain are partly memorization. This is not a far-fetched worry — the burden is on the authors to show disjointness, and they don't. EgoSchema, which comes from Ego4D, is independent, and VisCoP improves there slightly, so the contamination is specific to the in-house benchmark. The depth case is worse because the depth maps are derived from the same RGB videos.\n\nThe robot-control experiments are more credible: VIMA has fixed splits, and xArm-Det is fresh data. But the real-world numbers come from 20 trials per task with no error bars, and no significance tests appear anywhere in the paper. That is a general weakness.\n\nFor whom: people working on parameter-efficient VLM adaptation and egocentric video. The method is sensible and the failure mode is instructive. It deserves a serious referee, but the paper should not be accepted without a disclosed train/test split (ideally video-ID disjoint), error bars, and preferably an external ego benchmark as primary evidence. If the ego/depth gains collapse after decontamination, the contribution reduces to the robot results plus the architecture ablation — still a useful incremental step, but not the broad claim in the abstract.","headline":"VisCoP is a sensible parameter-efficient adaptation idea with good ablations, but the headline ego/depth gains are compromised by a likely train/eval overlap that the paper never rules out.","tokens_in":20955,"tokens_out":5498,"would_cite":false,"duration_ms":44312,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"VisCoP adapts video-language models to shifted domains by attaching learnable probes to a frozen vision encoder, reporting gains of +3.5 to +19.3 on target benchmarks while retaining source skills.","keywords":["video domain adaptation","vision-language models","visual probing","parameter-efficient finetuning","catastrophic forgetting","egocentric video","depth perception","robot control"],"falsifier":"Compare the 45,888 ego and 28,176 depth training instruction pairs against the 3,991 Ego-in-Exo PerceptionMCQ evaluation samples; if any overlap exists, rerun adaptation on a non-overlapping split and check whether VisCoP's target gains survive.","tokens_in":20029,"feed_emoji":"🎥","tokens_out":5863,"duration_ms":47817,"temperature":0.7,"pith_summary":"This paper argues that a vision-language model can be adapted to a new visual domain without retraining its vision encoder, by attaching a compact set of learnable visual probes that read from intermediate encoder layers. The authors claim these probes capture domain-specific cues — first-person viewpoint, depth imagery, robot scenes — that the frozen encoder's final output loses, so the model learns the target domain without catastrophic forgetting. The proposal, called VisCoP, is evaluated on three shifts: exocentric-to-egocentric video, RGB-to-depth video, and human-action-understanding-to-robot-control. In each setting VisCoP is reported to beat connector-only, vision-encoder, and full-model finetuning on target benchmarks while retaining more source-domain knowledge. If correct, it offers a parameter-efficient adaptation recipe that avoids the usual trade-off between specialization and forgetting.","feed_headline":"Probing beats finetuning for adapting video AI to new domains","feed_subtitle":"VisCoP gains up to +19.3 on targeted video shifts while holding onto source-domain skills.","key_machinery":"The central mechanism is the visual probe set (16 learnable tokens by default) together with per-layer interaction modules. Each interaction module is a multi-head cross-attention in which the probes query the frozen encoder's intermediate visual features, letting them aggregate cues across all spatio-temporal tokens at multiple abstraction levels. The final probe embeddings go through a dedicated connector and join the standard VLM embeddings as conditioning for the LLM. During adaptation, only the probes, interaction modules, connectors, and LLM low-rank adapters are trained; the vision encoder stays frozen. The probes act as a gradient bypass, steering learning toward domain-relevant stru","core_discovery":"The paper's central claim is that a small set of learnable visual probes, inserted at every layer of a frozen vision encoder and updated through cross-attention, forms an alternative pathway for learning domain-specific visual representations. At each layer, probes attend to intermediate spatio-temporal features; after the final layer, probe embeddings are projected into the language model and added to the standard visual conditioning. Across cross-view, cross-modal, and cross-task adaptation, VisCoP reportedly yields the largest target-domain improvement among the compared strategies while maintaining or improving source-domain performance, where full finetuning of the vision encoder loses","pith_inferences":["Editorial inference: the large gap between VisCoP's depth gain (+19.27) and its ego gain (+3.53) suggests the benefit scales with the size of the visual shift; this predicts even larger gains on more extreme modality changes, which could be tested directly.","Editorial inference: the mechanism implies the frozen encoder preserves useful domain information in middle layers that the final token aggregation discards; the same probe design may transfer to other shifts such as extreme illumination, sensor artifacts, or synthetic-to-real data without modification.","Editorial inference: because the ego and depth evaluations draw on data derived from the same source as the adaptation training and no train/test split is reported, the magnitude of those two gains should be re-measured on disjoint splits before being used as a benchmark comparison."],"forward_implications":["VisCoP is reported to improve target-domain performance over other adaptation strategies in all three settings: +3.53 on egocentric video, +19.27 on depth video, and +67.82 on simulated robot control.","It retains more source-domain knowledge than vision-encoder finetuning, with source deltas of +1.77 (ego), +1.84 (depth), and -4.58 (robot) versus -8.87 for full finetuning.","Layer-wise placement of the probes is load-bearing: using probes only at the final layer, or training only the last four encoder layers, gives much smaller or negative target gains.","A moderate probe count (16) is the sweet spot; 128 probes give no additional gain and can hurt, suggesting a compact representation is what matters.","In real-world robot manipulation, VisCoP improves success rates over vision-encoder finetuning in both zero-shot transfer from simulation and joint training with real-robot data."],"fun_headline_variants":["Visual probes adapt video AI to new domains without forgetting","Lightweight probes beat fine-tuning for video domain adaptation","VisCoP: tiny probes preserve VLM skills while adapting to new domains","Probes only: parameter-efficient adaptation for video VLMs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The ego and depth results assume that the Ego-in-Exo PerceptionMCQ evaluation videos and question-answer pairs are disjoint from the EgoExo4D-derived instruction pairs used for adaptation training, but no train/test split is reported.","fun_headline_variants_meta":{"raw":{"variants":["Visual probes adapt video AI to new domains without forgetting","Lightweight probes beat fine-tuning for video domain adaptation","VisCoP: tiny probes preserve VLM skills while adapting to new domains","Probes only: parameter-efficient adaptation for video VLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1062,"prompt_tokens":746,"completion_tokens":316,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":247}},"tokens_in":490,"tokens_out":316,"duration_ms":3441,"temperature":1.0,"reasoning_tokens":247,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:40:02.483224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the 45,888 ego and 28,176 depth training instruction pairs against the 3,991 Ego-in-Exo PerceptionMCQ evaluation samples; if any overlap exists, rerun adaptation on a non-overlapping split and check whether VisCoP's target gains survive.","supporting_citations":[],"review_version":1}