{"id":"90afc060-b2b6-4f93-8adf-86725a4144d2","arxiv_id":"2412.10707","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MambaPro combines CLIP with a parallel adapter, synergistic residual prompts, and Mamba aggregation to achieve state-of-the-art mAP on RGBNT201, RGBNT100, and MSVR310.","lead":"An object re-identification framework that adapts a frozen CLIP vision model for multi-modal person and vehicle retrieval using parallel adapters, synergistic prompts, and a Mamba-based fusion block. It reports top accuracy on three RGB/NIR/TIR benchmarks, but the gains over prior methods may largely reflect CLIP's pretrained features rather than the new components.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is not CLIP-controlled: CLIP initialization alone buys ~11.4 mAP over ViT, nearly twice the reported margin over TOP-ReID.","rationale":"The reader's weakest assumption is the same one I judge most load-bearing: the main performance comparison is confounded by CLIP initialization. The paper's own Table 9 demonstrates that the pretrained backbone choice contributes more to the RGBNT201 mAP difference than the claimed margin over TOP-ReID, so the headline 'surpasses TOP-ReID by 6.6%' is not an apples-to-apples comparison. The supplementary Table 10 is a partial control and deserves credit, but it covers only a single baseline and does not appear in the main tables. A concrete test with CLIP-initialized versions of several strong prior methods would settle whether the architecture or the backbone drives the result. I also considered the lower-complexity claim, which is weakened by Table 6 (MA: 1.42 G FLOPs vs Transformer: 1.36 G) and by parameter-count inconsistencies between Tables 3, 4, and 5; however, those issues affect the secondary efficiency claim rather than the primary performance claim, and they do not change the recommended verdict. The reader's CONDITIONAL verdict is appropriate: the internal ablations are well controlled, but the external SOTA statement needs additional experiments before acceptance.","tokens_in":18837,"tokens_out":14927,"duration_ms":129164,"concrete_test":"Train TOP-ReID, EDITOR, and RSCNet with CLIP ViT-B/16 initialization using the same full fine-tuning recipe, loss weights, batch size, and epochs as in the paper, and enter their mAP/Rank-1 into Tables 1 and 2 (or a new comparison table). Run each configuration with at least three seeds and report mean +/- std. If any CLIP-initialized baseline reaches or exceeds MambaPro's 78.9/83.4 on RGBNT201, the SOTA claim as stated is false; if all remain below and relative ordering is preserved, the architectural contribution is established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MambaPro surpasses all prior multi-modal ReID methods is not supported by a controlled comparison. In Tables 1 and 2, no baseline uses CLIP initialization; prior methods are ImageNet-pretrained or trained from scratch, while MambaPro is CLIP-based. Supplement Table 9 quantifies this confound: with identical trainable modules, replacing the frozen ImageNet ViT backbone with frozen CLIP raises RGBNT201 mAP from 67.5 to 78.9 (11.4 points), roughly twice the 6.6-point margin claimed over TOP-ReID in Table 1. The only CLIP-initialized baseline, TOP-ReID+CLIP in Table 10, trails MambaPro by 5.4 mAP, so the confound does not automatically erase the gain; however, it is one method, fully fine-tuned rather than PEFT-tuned, and is absent from the main comparison tables. Stronger non-CLIP baselines such as EDITOR (66.5) and RSCNet (68.2) could plausibly close or invert the 78.9 mAP result once initialized with CLIP and given comparable training budgets. The internal ablations control for backbone and support the modules, but the externally valid SOTA statement requires CLIP-controlled baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MambaPro, a parameter-efficient fine-tuning framework for multi-modal object re-identification. It uses a frozen CLIP image encoder with three proposed modules: a Parallel Feed-Forward Adapter (PFA) for knowledge transfer, a Synergistic Residual Prompt (SRP) for modality interaction, and a Mamba Aggregation (MA) block for intra- and inter-modality sequence modeling with claimed linear complexity. Experiments on RGBNT201, RGBNT100, and MSVR310 report state-of-the-art mAP, and ablations show each module contributes over a frozen CLIP baseline. The main empirical claim is that MambaPro surpasses prior multi-modal ReID methods, e.g., by 6.6 mAP over TOP-ReID on RGBNT201, while using fewer trainable parameters than full fine-tuning.","tokens_in":19109,"tokens_out":7509,"duration_ms":61208,"significance":"If validated, the contribution is significant for the multi-modal ReID community: it demonstrates that a frozen CLIP backbone can be adapted to multi-modal object ReID with lightweight modules, and the internal ablation design is careful, with each proposed component tested over the same frozen-backbone baseline. The released code is a practical asset. However, the headline state-of-the-art claim is currently confounded by the CLIP pretraining of MambaPro versus the non-CLIP baselines in the main tables; the internal ablations are convincing, but the externally valid comparison needs additional controlled baselines.","major_comments":[{"comment":"The state-of-the-art claim is not controlled for pretraining. Supplementary Table 9 shows that replacing the frozen ImageNet ViT backbone with frozen CLIP, while keeping all proposed modules fixed, raises RGBNT201 mAP from 67.5 to 78.9 (11.4 points), which is larger than the 6.6-point margin over TOP-ReID reported in Table 1. Since none of the baselines in Tables 1-2 uses CLIP initialization, the reported superiority cannot be attributed to the proposed modules alone. The only CLIP-based baseline, TOP-ReID+CLIP in Table 10, is compared against fully fine-tuned MambaPro (160.34M parameters, Model G in Table 3) rather than the proposed efficient MambaPro (74.20M, Model F), and this comparison is absent from the main tables. Please add CLIP-initialized versions of at least TOP-ReID, EDITOR, and RSCNet to Tables 1-2, or to a clearly labeled separate table, using comparable training recipes, and state explicitly which comparisons are CLIP-controlled.","section":"Performance Comparison (Tables 1-2) and Supplementary Tables 9-10"},{"comment":"The parameter counts for the same configuration are inconsistent across tables. CLIP(L)+PFA is reported as 28.34M trainable parameters in Table 3 but 2.36M in Table 5, and CLIP(L)+SRP(Fusion) is 28.48M in Table 3 but 2.37M in Table 4. This discrepancy directly undermines the parameter-efficiency claims, including the statement in Section 'Effect of Key Components' that Model C uses '32.9% of the trainable parameters compared with Model B'. Please clarify what each column counts (module-only vs total trainable parameters, including any unfrozen LayerNorm, position embeddings, or classification head) and make the numbers consistent across all tables.","section":"Ablation Study, Tables 3-5"},{"comment":"The abstract claims 'lower FLOPs' and 'linear complexity' for the full framework, but the module-level numbers in Table 6 show the full MA block at 1.42G FLOPs and 9.47M parameters versus 1.36G and 4.73M for the single Transformer baseline used in the same table. The fair matched-structure comparison is deferred to supplementary Table 11 (MA(Trans) 1.81G/14.19M vs MA(Mamba) 1.42G/9.47M). Please bring the matched-structure comparison into the main text and report total model FLOPs for the full MambaPro compared with full-fine-tuning baselines, so that 'lower complexity' is substantiated by the actual empirical setup rather than by the asymptotic scaling argument in Fig. 1.","section":"Mamba Aggregation and Complexity Claims (Tables 6, 11; Fig. 1)"}],"minor_comments":[{"comment":"There is a typo: 'RGBNT201/MVSR310' should be 'RGBNT201/MSVR310'.","section":"Implementation Details"},{"comment":"On RGBNT100, MambaPro's Rank-1 (94.7) is lower than several baselines (LRMM 96.7, CCNet 96.3, UniCat 96.2, TOP-ReID 96.4, EDITOR 96.4, RSCNet 96.6). The text only claims mAP superiority, but the abstract's 'surpassing TOP-ReID by 6.6% and 6.8%' refers to RGBNT201; please specify that the state-of-the-art claim is on mAP for RGBNT100 (and on both metrics for RGBNT201 and MSVR310), or provide additional evidence for a broader claim.","section":"Table 2"},{"comment":"The name 'MambaPro' is used for two different models: the default PEFT variant with 74.20M parameters (Model F, Table 3) and the fully fine-tuned variant with 160.34M parameters (Model G, Table 10). Please disambiguate these variants in the text and tables to avoid confusion about which model is being compared.","section":"Tables 3, 7, 10"},{"comment":"The notation for transferred prompts, e.g., \\p^{r→n}_1], is used in the equations but not explicitly defined in the method text; define all prompt tokens and the transfer block T before using them in the formulas.","section":"Equations (8)-(12)"},{"comment":"Figure 1(b) labels the comparison as 'SSM vs SA' but the actual empirical FLOPs comparison in Table 6 uses a Transformer aggregation block, not self-attention in the same dual intra/inter structure; consider aligning the figure's claim with the matched-structure experiment in Table 11.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The core modules appear well-ablated over a frozen CLIP baseline, and the paper addresses an interesting question. The main barrier to acceptance is the uncontrolled SOTA comparison and the inconsistent parameter reporting. I would encourage the editor to ask for CLIP-initialized baselines in the main tables and for a single consistent definition of 'trainable parameters' across all tables. The paper is otherwise within scope for a computer vision venue, but the current presentation overstates the external claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on MambaPro. The paper does something concrete: it takes a frozen CLIP image encoder, adds a parallel adapter, a synergistic residual prompt, and a Mamba-based aggregation module, and evaluates on three multi-modal ReID benchmarks. The internal ablations are the strong part. Each module earns its keep over a frozen CLIP baseline, and the full model beats fully fine-tuned CLIP while training far fewer parameters. The prompt mechanism ablation (independent vs synergistic vs residual) is clean, and the Mamba vs Transformer aggregation comparison is properly done. That is real engineering work, reported honestly.\n\nThe soft spots are real but not fatal. The headline SOTA comparison is confounded. In Tables 1 and 2, every prior method is trained without CLIP initialization, and MambaPro's margin over TOP-ReID is 6.6 mAP. The supplement's Table 9 shows that swapping the frozen backbone from ImageNet ViT to CLIP is worth 11.4 mAP by itself, with the same trainable modules. The paper does include a CLIP-initialized TOP-ReID in Table 10, where MambaPro still wins by 5.4 mAP, so the gain does not evaporate entirely. But one fully fine-tuned baseline is thinner than the main tables suggest; the main comparison should have included that or an equivalently PEFT-tuned CLIP baseline.\n\nThere are also presentation problems. Parameter counts are inconsistent: Table 3 lists the PFA row at 28.34M total trainable parameters, while Table 5 lists PFA at 2.36M, presumably module-only, without saying so. The \"lower complexity\" claim is not supported by Table 6, where MA has 1.42 GFLOPs versus 1.36 GFLOPs for a standard Transformer block at these sequence lengths; the linear-complexity argument is asymptotic and does not hold at this operating point.\n\nNo circularity problem: hyperparameters are tuned on validation, and the modules are ablated cleanly. The code link also helps reproducibility.\n\nIf I were refereeing, I would ask for CLIP-controlled baselines in the main comparison, a consistent parameter accounting, and a softened complexity claim. The core contribution — that a PEFT-tuned CLIP with Mamba aggregation works well for multi-modal ReID — survives those fixes.\n\nBottom line: it's a solid empirical paper for the ReID subfield, worth a serious referee, but the SOTA claim needs reconditioning before it should be accepted as stated.","headline":"Solid PEFT+CLIP+Mamba ReID paper with clean ablations; SOTA claim is confounded by CLIP pretraining and needs reconditioning before acceptance.","tokens_in":19642,"tokens_out":2228,"would_cite":false,"duration_ms":19685,"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":"MambaPro freezes CLIP, adds parallel adapters, synergistic prompts, and Mamba fusion, and reports state-of-the-art mAP on three multi-modal ReID benchmarks.","keywords":["multi-modal object re-identification","parameter-efficient fine-tuning","CLIP adaptation","Mamba state space model","prompt tuning","adapter tuning","RGB-NIR-TIR fusion","re-identification benchmarks"],"falsifier":"Train the strongest prior architecture (for example TOP-ReID) with the same frozen CLIP image encoder and training recipe but without PFA/SRP/MA, and measure RGBNT201 mAP; if it reaches the 78.9% that MambaPro reports, the claimed module-level gains are not responsible for the headline result. The paper's own supplementary runs this control for TOP-ReID under full fine-tuning and records 73.3% mAP, so a frozen-backbone control is the missing experiment.","tokens_in":18653,"feed_emoji":"🔍","tokens_out":11008,"duration_ms":89860,"temperature":0.7,"pith_summary":"Multi-modal object re-identification (ReID) tries to find the same person or vehicle across cameras when images come from RGB, near-infrared, and thermal sensors. MambaPro's central claim is that a frozen CLIP image encoder, left untouched, can beat full fine-tuning on this task once three small trainable modules are attached: a Parallel Feed-Forward Adapter (PFA), a Synergistic Residual Prompt (SRP) that exchanges prompt tokens between modalities, and a Mamba Aggregation (MA) that models within- and across-modality token sequences with linear-complexity state space blocks. On the three benchmarks the paper reports the best mean average precision (mAP) so far: 78.9% on RGBNT201, 83.9% on RGBNT100, and 47.0% on MSVR310, with a trainable-parameter count well below full fine-tuning. The reason a sympathetic reader should care is that this points toward cheap reuse of general-purpose vision-language models for multi-modal retrieval, and toward state space models as a substitute for quadratic self-attention when fusing long multi-modal sequences.","feed_headline":"MambaPro tops multi-modal ReID with a frozen CLIP","feed_subtitle":"Parallel adapters, synergistic prompts, and Mamba aggregation beat full fine-tuning at lower cost.","key_machinery":"The load-bearing mechanism is the interaction of the frozen CLIP image encoder with three trainable side modules. The Parallel Feed-Forward Adapter (PFA) is a linear-GELU-linear branch connected in parallel with each frozen feed-forward block; it preserves the original feature flow while transferring CLIP knowledge to the ReID task. The Synergistic Residual Prompt (SRP) concatenates prompt tokens into each layer's input, transforms each modality's prompts into the other modalities via a transfer block, and adds the average of previous prompts to the next layer's prompts, so multi-modal information is exchanged and carried forward. The Mamba Aggregation (MA) uses selective state space models (SSMs), the mechanism behind Mamba that processes sequences in linear rather than quadratic time, to model intra-modality token interactions and then inter-modality interactions over the concatenated three-modality token sequence.","core_discovery":"On the paper's own terms, the discovery is that a strong multi-modal ReID model does not need to fine-tune a large pre-trained transformer; it needs to preserve the pre-trained representations while adding modality-aware information exchange around them. The PFA sits parallel to each feed-forward block so original features flow through unchanged, while an ascending-then-descending linear adapter injects ReID-specific refinements. The SRP makes each modality's prompt tokens learnable, transfers them into the other modalities through a linear-GELU-linear transfer block, and carries fused prompts forward with residual connections so no layer's prompt information is discarded. The MA then processes patch tokens of each modality separately through a selective state space block, concatenates tokens from all modalities into one long sequence, and processes that sequence through another state space block before splitting it back. With these parts, MambaPro reports 78.9% mAP and 83.4% Rank-1 on RGBNT201, 83.9% mAP on RGBNT100, and 47.0% mAP on MSVR310.","pith_inferences":["Editorial inference: because all main-table comparisons train prior methods without CLIP initialization, the size of the reported margin is partly a statement about CLIP pretraining, not only about the new modules; the supplement controls this for TOP-ReID only, where the CLIP-based comparison is 73.3 vs 78.7 mAP.","Editorial inference: the linear-complexity claim applies to the Mamba aggregation stage; the frozen CLIP backbone itself still runs quadratic self-attention over its patch tokens, so the whole pipeline is not linear end to end.","Editorial inference: the synergistic prompt design could be lifted into other multi-modal retrieval settings, such as text-based person search or audio-visual matching, wherever each modality has its own set of prompts."],"forward_implications":["Full fine-tuning of a large pre-trained CLIP encoder is not necessary for multi-modal ReID: the frozen-backbone model (78.9% mAP) matches the fully fine-tuned variant (78.7%) while training far fewer parameters.","Mamba blocks can replace transformer attention in the fusion stage: the Mamba aggregator reaches 78.9% mAP versus 76.5% for a transformer aggregator at roughly comparable FLOPs.","Cross-modal prompt transfer works as a fusion mechanism: transferring prompt tokens between modalities beats independent per-modality prompts (68.9 vs 64.7 mAP), and adding the residual prompt fusion pushes this further (71.3 mAP).","Where the adapter is placed matters: a parallel branch to the feed-forward network gives 73.2 mAP, whereas inserting it after the feed-forward block drops to 46.9 mAP.","CLIP pre-training itself matters in this setup: the frozen CLIP backbone reaches 78.9 mAP while a frozen ImageNet-pretrained ViT reaches 67.5 mAP."],"supporting_citations":[{"why":"Supplies the frozen CLIP image encoder that all three MambaPro modules adapt; the pre-trained knowledge the paper claims to transfer.","marker":"Radford et al. 2021"},{"why":"Defines the selective state space model (Mamba) used in Mamba Aggregation; the reason MA can handle long multi-modal sequences with linear complexity.","marker":"Gu and Dao 2023"},{"why":"TOP-ReID, the main multi-modal ReID baseline MambaPro compares against on RGBNT201; also re-run with CLIP in the supplement.","marker":"Wang et al. 2023"},{"why":"Defines the ViT transformer layer structure that CLIP's frozen encoder and the PFA/SRP modules are inserted into.","marker":"Dosovitskiy et al. 2020"},{"why":"Introduces adapter tuning, the parameter-efficient paradigm from which PFA is derived.","marker":"Houlsby et al. 2019"},{"why":"Introduces prompt tuning, the paradigm the Synergistic Residual Prompt builds on.","marker":"Lester, Al-Rfou, and Constant 2021"},{"why":"Introduces the RGBNT201 multi-modal person ReID benchmark and the PFNet progressive fusion baseline.","marker":"Zheng et al. 2021"},{"why":"Introduces the RGBNT100 multi-spectral vehicle ReID benchmark and the HAMNet baseline.","marker":"Li et al. 2020"},{"why":"Introduces the MSVR310 vehicle ReID benchmark used for generalization testing.","marker":"Zheng et al. 2022"}],"fun_headline_variants":["Frozen CLIP plus adapters and Mamba for multi-modal ReID","MambaPro: no fine-tuning, just adapters, prompts, and Mamba","MambaPro: keep CLIP frozen, add adapters and Mamba","Adapter, prompt, Mamba: MambaPro dethrones fine-tuning","MambaPro: frozen CLIP + adapters + Mamba for ReID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison assumes MambaPro's gains come from its modules, but the prior methods in the main tables were not trained with CLIP initialization; if CLIP's pre-trained features alone explain most of the gap, the architectural advantage would shrink.","fun_headline_variants_meta":{"raw":{"variants":["Frozen CLIP plus adapters and Mamba for multi-modal ReID","MambaPro: no fine-tuning, just adapters, prompts, and Mamba","MambaPro: keep CLIP frozen, add adapters and Mamba","Adapter, prompt, Mamba: MambaPro dethrones fine-tuning","MambaPro: frozen CLIP + adapters + Mamba for ReID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000884,"raw_usage":{"total_tokens":3842,"prompt_tokens":996,"completion_tokens":2846,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2742}},"tokens_in":612,"tokens_out":2846,"duration_ms":18786,"temperature":1.0,"reasoning_tokens":2742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:41:25.712215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the strongest prior architecture (for example TOP-ReID) with the same frozen CLIP image encoder and training recipe but without PFA/SRP/MA, and measure RGBNT201 mAP; if it reaches the 78.9% that MambaPro reports, the claimed module-level gains are not responsible for the headline result. The paper's own supplementary runs this control for TOP-ReID under full fine-tuning and records 73.3% mAP, so a frozen-backbone control is the missing experiment.","supporting_citations":[{"cited_title":"TOP-ReID: Multi-spectral Object Re-Identification with Token Permutation","cited_arxiv_id":"2312.09612","evidence_quote":"TOP-ReID, the main multi-modal ReID baseline MambaPro compares against on RGBNT201; also re-run with CLIP in the supplement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces adapter tuning, the parameter-efficient paradigm from which PFA is derived."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the RGBNT201 multi-modal person ReID benchmark and the PFNet progressive fusion baseline."}],"review_version":1}