{"id":"8a5d0a10-8b9b-4ee6-932d-ecd8f95c0dff","arxiv_id":"2607.26283","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A prompt-based, privacy-preserving heterogeneity-alignment framework that uses low-rank FiLM to adapt BEV features and an autoencoder-based classifier for metadata-free modality routing.","lead":"HeteroPROMPT lets vehicles with different sensors and perception software share their view by inserting small adjustable 'prompt' modules that align each vehicle's feature map to the receiver's unified space. It also identifies a partner's sensor type automatically from the shared features alone, so no proprietary hardware or model details are exposed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is unsupported because Faster-HEAL and PEARL, the two closest prompt-based heterogeneous CP baselines, are described but never compared in Table II.","rationale":"The reader's weakest_assumption concerns closed-set deployment, which is a limitation the paper acknowledges and defers to future work. That is important but does not directly invalidate the central detection claim for the tested modality set. The reader's rationale also mentions the missing Faster-HEAL/PEARL comparison; this is the most load-bearing gap because the headline 'state-of-the-art' and 'orders of magnitude fewer parameters' claims cannot be evaluated without the closest prior methods. A concrete re-evaluation with those baselines would settle whether the paper's primary claim is valid. I therefore keep the verdict at CONDITIONAL/UNCHANGED: the paper should not be rejected outright, but acceptance should require the missing baseline comparison and correction of the classifier accuracy claim.","tokens_in":12949,"tokens_out":8804,"duration_ms":89153,"concrete_test":"Run Faster-HEAL and PEARL on OPV2V-H and V2XSet using the exact Table II protocol: m1 as ego base, progressively adding m2, m3, m4, and report AP50, AP70, and number of trainable parameters under identical splits and detection settings. If either baseline matches or exceeds HETEROPROMPT's AP at comparable or lower parameter count, the claim to state-of-the-art and to order-of-magnitude parameter efficiency collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—HETEROPROMPT improves AP over state-of-the-art heterogeneous CP methods with orders-of-magnitude fewer parameters—rests on Table II, which compares only against HEAL, CoBEVT, and HM-ViT. But Related Work (Section II-B) identifies Faster-HEAL [16] and PEARL [17] as the most directly comparable methods: both are prompt-based, parameter-efficient, privacy-preserving heterogeneous CP frameworks. Omitting them from the quantitative evaluation means the comparison set excludes precisely the baselines HETEROPROMPT must outperform to justify 'state-of-the-art' and 'orders of magnitude fewer parameters.' If either of those methods achieves comparable or better AP at similar parameter counts, the paper's main contribution is significantly weakened. This is not merely a missing citation; it is missing critical control experiments for the paper's own claimed advance. The closed-set limitation noted by the reader is real but explicitly acknowledged; the >99.99% accuracy inconsistency is a separate corrigible claim. The omission of the two closest baselines is the more load-bearing issue for the central SOTA assertion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HeteroPROMPT, a framework for heterogeneous collaborative perception. It aligns BEV features from newly joining agents to an ego-centric unified space using two lightweight prompt-conditioned FiLM modules (DRIFT) with PARAFAC low-rank prompts, while freezing agent encoders and the ego fusion/detection stack. A second contribution is metadata-free modality routing: an autoencoder compresses communicated BEV features to a compact latent space, and a lightweight MLP classifier predicts the joining agent's modality to select the appropriate DRIFT pair. Experiments on OPV2V-H and V2XSet compare against HEAL, CoBEVT, and HM-ViT, reporting higher AP50/AP70 with far fewer trainable parameters, plus a claimed >99.99% modality-classification accuracy. The central idea is plausible and the component-level ablations support the design, but the evaluation omits the two most closely related prompt-based baselines, and the stated classification accuracy is quantitatively inconsistent with the reported confusion matrix.","tokens_in":13231,"tokens_out":3697,"duration_ms":36100,"significance":"If the claims hold, HeteroPROMPT would be a useful step toward practical heterogeneous collaborative perception: parameter-efficient adaptation that avoids retraining large perception stacks, preserves agent encoder privacy, and enables metadata-free routing. The PARAFAC decomposition of visual prompts is a sensible way to reduce adaptation cost, and the two-point FiLM placement with ablations is a reasonable engineering contribution. The use of public datasets (OPV2V-H, V2XSet) and the inclusion of training-throughput/memory measurements are strengths. However, the paper's 'state-of-the-art' claim currently rests on a comparison set that excludes the two prior prompt-based heterogeneous-CP methods it identifies as most related, and the >99.99% classification accuracy cannot be reconciled with the one-error confusion matrix under the stated test-set sizes. The core approach is defensible, but these issues require substantive experimental or textual revision.","major_comments":[{"comment":"The related-work section identifies Faster-HEAL [16] and PEARL [17] as prompt-based, parameter-efficient, privacy-preserving heterogeneous CP methods that are directly comparable to HeteroPROMPT, yet neither appears in Table II. As a result, the central claim of 'state-of-the-art' AP and 'orders-of-magnitude fewer parameters' is not tested against the closest baselines. Please add quantitative comparisons to [16] and [17] on the same settings, or justify their exclusion with concrete numbers; otherwise the SOTA assertion is unsupported by the presented evidence.","section":"II-B, Table II"},{"comment":"The paper states that the modality classifier achieves '>99.99%' accuracy, with 'one confusion between m5 and m2.' Section IV-C reports per-modality test samples of 30 (or 140/30/30 splits) for each modality. With 8 modalities and 30 test samples each, one error yields 99.58%; even with 140 test samples per modality, one error yields 99.11%. Both are far below 99.99%. The abstract and conclusion repeat the >99.99% claim. Please report the exact total number of test samples and the true accuracy, and adjust all occurrences of the claim accordingly.","section":"IV-C, Fig. 3 vs Abstract/V"},{"comment":"The reported AP improvements over HEAL are small (+1.0 to +1.8 points AP50), and no error bars, confidence intervals, or multiple seeds are reported. Without such variability information, the claim of 'consistently improves' is not statistically grounded, especially when the differences are comparable to run-to-run noise. Please add results over at least three seeds (mean and std), or otherwise temper the claim to a single-run observation.","section":"Table II, IV-B"}],"minor_comments":[{"comment":"The abstract and introduction say 'orders of magnitude fewer trainable parameters,' citing 0.064M vs 15M. However, Table II lists HEAL parameters as 15.0M for +m2, 1.1M for +m3, and 1.9M for +m4. For +m3 and +m4 the reduction is a factor of 17–30, i.e., about one order of magnitude, not 'orders of magnitude.' Please either explain what the 15M/1.1M/1.9M counts represent or adjust the wording.","section":"Table II, Abstract"},{"comment":"The contribution bullet claims 'reducing average training overhead by 96%.' Table III reports 2.5x higher throughput, 2.27x higher TFLOPs/s, and 45.7% lower peak GPU memory — none of which directly translates to a 96% training-overhead reduction. Presumably this refers to trainable parameter reduction, not wall-clock or memory overhead. Please clarify the metric used for the 96% claim.","section":"Contributions, Table III"},{"comment":"The paper correctly acknowledges in Section III-C that Stage 2 adaptation is performed offline for a closed set of common sensor/model configurations, and Section V defers open-set generalization to future work. This is an important limitation; please state it explicitly in the abstract or introduction so that the 'real-time deployment' framing is not overstated.","section":"III-C, V"},{"comment":"The t-SNE visualization is qualitative. Consider reporting a quantitative separation metric (e.g., k-NN accuracy or adjusted Rand index on the latent space) to support the claim that the compact space 'forms distinct modality clusters.'","section":"IV-C, Fig. 2"},{"comment":"The 'privacy-preserving' claim is tied to not sharing sensor/model metadata, but intermediate BEV features themselves may expose scene information. A sentence acknowledging that privacy is defined relative to metadata disclosure, not feature content, would improve clarity.","section":"I, IV-B"}],"recommendation":"major_revision","confidential_remarks":"The two missing baselines, Faster-HEAL [16] and PEARL [17], are the authors' own prior works, so adding fair comparisons should be feasible and not require reimplementation from scratch. Their omission is the main obstacle to accepting the SOTA claim. The >99.99% accuracy claim is internally inconsistent and must be corrected. The lack of multiple seeds is a secondary but real concern given the small AP margins."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering paper with a real parameter-efficiency win, but the main comparative claim is under-supported because the two closest prior methods are described and never benchmarked.\n\nThe core contribution is a modular FiLM alignment with PARAFAC-decomposed prompts plus an autoencoder-based classifier for metadata-free routing. That combination is genuinely new, and the parameter counts are striking: 0.064M trainable parameters versus 15M for HEAL, with consistent AP improvements on OPV2V-H and V2XSet. The privacy motivation is sound — freezing the encoder and communicating only intermediate features is a reasonable threat model, and the routing-by-latent-classification idea is practical.\n\nThe biggest soft spot is the missing comparison to Faster-HEAL and PEARL, both cited as the most directly related prompt-based, parameter-efficient, privacy-preserving heterogeneous CP frameworks. Without those controls, 'state-of-the-art' is a claim about a tournament the paper only partially entered. This is not a citation nit but a missing control experiment; if either of those methods achieves comparable or better AP at similar parameter counts, the central contribution is much weaker.\n\nSecond, the >99.99% modality-classification claim contradicts the reported confusion matrix with one error. If that is one error out of a small test set, the accuracy is likely 98–99%, not 99.99. This needs a corrected statistic and sample size. Also, the AE latent size is selected on test loss, which is a mild model-selection-on-test issue; it should be presented as validation-based selection. No error bars or multiple seeds are reported for detection results, which matters because the AP gaps over HEAL are 1–2 points.\n\nThe closed-set routing assumption—unseen modalities are not handled—is explicitly acknowledged as future work, so it is a limitation rather than a hidden flaw.\n\nWho should read this? Anyone working on V2X deployment or parameter-efficient domain adaptation in collaborative perception. The central idea is defensible, and the flaws are addressable. I would send it to peer review rather than desk-reject, but the missing baselines and the classification statistic need to be fixed before acceptance.","headline":"HeteroPROMPT is a genuinely parameter-efficient prompt-based heterogeneous CP method, but its SOTA claim is undermined by omitting the two closest baselines: the authors' own Faster-HEAL and PEARL.","tokens_in":13691,"tokens_out":1987,"would_cite":true,"duration_ms":18648,"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":"HeteroPROMPT shows that heterogeneous collaborative perception reduces to a per-modality prompt-tuning problem, alignable with 0.064M parameters and no metadata sharing.","keywords":["collaborative perception","heterogeneous agents","feature alignment","visual prompts","FiLM","low-rank decomposition","privacy preservation","modality classification"],"falsifier":"Take any perception encoder not in the paper's modality set (for instance, a transformer-based LiDAR detector or a camera model trained on a different city), feed its BEV features to the classifier, and check whether routing picks the correct DRIFT and whether AP50 stays within the reported range. If an unseen-but-similar modality is misrouted with high confidence, the metadata-free routing claim does not generalize outside the closed set.","tokens_in":12846,"feed_emoji":"🚗","tokens_out":4251,"duration_ms":39358,"temperature":0.7,"pith_summary":"The paper claims that heterogeneous collaborative perception—where vehicles with different sensors, backbones, and training data share intermediate features—can be made practical and privacy-preserving without retraining the heavy fusion and detection stacks. Its central proposal, HeteroPROMPT, adapts each newly joining agent by training only two tiny 'prompt' modules (DRIFT) that modulate received Bird's Eye View features through rank-reduced FiLM affine transforms. With only 0.064M trainable parameters, it reports higher detection Average Precision than state-of-the-art retraining-based and interpreter-based methods on two standard datasets. To avoid exposing proprietary sensor/model metadata, an autoencoder compresses the received features into a compact latent space and a lightweight classifier routes each agent to the correct DRIFT pair, achieving >99.99% modality classification accuracy on the evaluated closed set. A sympathetic reader should care because this is the first time the heterogeneous-collaboration problem is reduced to a per-modality prompt-tuning task with near-zero adaptation cost.","feed_headline":"0.064M-parameter prompts unify mixed-sensor vehicle collaboration","feed_subtitle":"No heavy retraining, no shared metadata: low-rank prompts align every new vehicle to one unified space in real time.","key_machinery":"The load-bearing object is DRIFT (Decomposed-Rank Interpreter for FiLM Tuning), a per-modality module that computes channel-wise affine parameters (gamma, beta) from a learnable visual prompt and applies them to BEV features. The prompt itself is stored as a PARAFAC low-rank decomposition (three factor matrices over channel, height, width), which is what collapses parameter count. Two DRIFT modules are inserted at different pipeline depths: one on the raw communicated feature (matching the sender's resolution) and one after the resizer (matching the ego's unified resolution), so alignment happens both in the sender's domain and in the unified domain. The second key mechanism is the autoencod","core_discovery":"HeteroPROMPT establishes that feature-space shifts between heterogeneous agents can be closed by two low-rank visual-prompt-conditioned FiLM modules placed before and after the feature resizer, while keeping the ego agent's fusion/head and the joining agent's encoder frozen. The prompts are decomposed via PARAFAC into factor matrices of rank R, cutting trainable parameters from C×H×W to R(C+H+W), which is what makes the whole system fit in 0.064M parameters. A second, deployment-time discovery is that modality identity is recoverable from compact features: after Stage-1 training, an autoencoder learned on ego-domain features, together with a three-layer MLP on the pooled latent, classifies t","pith_inferences":["The closed-set assumption is the main boundary: the paper tests only modalities m2–m9; an open-set variant could treat the classifier's softmax confidence as a rejection rule and fall back to no collaboration or a generic aligner for unknown modalities.","The compact AE latent space might serve as a general domain fingerprint for other tasks (e.g., predicting pose noise, latency, or best fusion weight) not just modality routing.","One could test whether the DRIFT alignment transfers to entirely different downstream heads (e.g., segmentation) without retraining, since alignment happens in a task-agnostic unified feature space.","The method implicitly assumes feature distributions are stable for a given modality; a natural extension is to learn a DRIFT per training-domain per sensor rather than per sensor+model, which would shrink storage further."],"forward_implications":["New agents can be added by training only two low-rank prompt modules per modality, so fleet expansion becomes cheap and does not require access to the agent's encoder weights.","Because routing uses features, not metadata, collaboration can proceed without partners revealing their sensor/model configuration, reducing privacy exposure and attack surface.","The >99.99% modality classification means reliable module selection is feasible at deployment latencies, not just in principle.","The 94–99.6% parameter reduction relative to retraining baselines suggests the approach fits on onboard compute for real-time inference.","If the frozen single-agent stack is preserved, agents can fall back to their own perception when collaboration is interrupted without any retraining."],"fun_headline_variants":["Tiny prompts align mixed-sensor vehicles in real time","0.064M-parameter prompts bridge heterogeneous perception gaps","Privacy-preserving prompts unify vehicle sensor fusion","Frozen models, tiny prompts: real-time heterogeneous CP","Low-cost prompts enable privacy-friendly vehicle collaboration"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The deployment pipeline assumes the joining agent's sensor/encoder/training combination is one of the closed set of modalities used to train DRIFT and the classifier (m2–m9); an unseen combination gets routed to a wrong interpreter and alignment fails.","fun_headline_variants_meta":{"raw":{"variants":["Tiny prompts align mixed-sensor vehicles in real time","0.064M-parameter prompts bridge heterogeneous perception gaps","Privacy-preserving prompts unify vehicle sensor fusion","Frozen models, tiny prompts: real-time heterogeneous CP","Low-cost prompts enable privacy-friendly vehicle collaboration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1171,"prompt_tokens":847,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":248}},"tokens_in":591,"tokens_out":324,"duration_ms":4098,"temperature":1.0,"reasoning_tokens":248,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:15:17.159415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any perception encoder not in the paper's modality set (for instance, a transformer-based LiDAR detector or a camera model trained on a different city), feed its BEV features to the classifier, and check whether routing picks the correct DRIFT and whether AP50 stays within the reported range. If an unseen-but-similar modality is misrouted with high confidence, the metadata-free routing claim does not generalize outside the closed set.","supporting_citations":[],"review_version":1}