{"id":"7d3f1ec3-93d7-4e1f-aa19-82af84f77f45","arxiv_id":"2608.10864","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Multi-view relational distillation transfers geometric knowledge to vision-language models by matching cross-view patch similarity matrices, improving spatial reasoning with minimal overhead.","lead":"This paper introduces a training method that makes vision-language models spatially aware by distilling cross-view similarity patterns from a 3D vision teacher instead of copying its features. The method improves spatial reasoning benchmarks while adding far fewer parameters and less latency than feature-fusion alternatives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Alignment-retention evidence is selection-biased: the 42-question probe in Appendix A.1 conditions on SFT beating distillation, so Figure 4b's precision gap cannot support a general claim that MVRD preserves vision-language alignment.","rationale":"The paper is clear and the main VSI-Bench results are internally consistent: MVRD improves over SFT on LLaVA by 2.2 points and over LoRA-based feature distillation by 1.4, with similar or larger margins on InternVL3; the dual-pathway variant further closes the gap to VLM-3R at low overhead; and transfer to scene understanding improves over 3DRS on all reported metrics. The relational loss design is sensible, the ablations in Table 13 show that both the cosine relation construction and the Pearson loss are load-bearing, and the reproducibility request is a standard conditional condition. The identified weakest assumption is the alignment-retention premise. Section 4's invariance argument establishes that solutions with low relational loss exist near any given feature configuration under rotation and per-vector scaling, but it does not prove that gradient training selects one close to the pretrained vision-language space. The quantitative probe meant to demonstrate this is selection-biased, as the appendix itself discloses. Because this probe is the paper's main representation-level evidence for the 'retaining language alignment' claim, that part of the central claim is not yet substantiated. The task-level Appr. Order and Obj. Size results provide indirect support for the alignment advantage and keep the overall method credible, so the appropriate response is CONDITIONAL, as the reader concluded, not rejection. The concrete test above would settle the probe's validity: if an unbiased probe confirms a large MVRD advantage, the alignment claim is solid; if not, the paper should de-emphasize the representation-level claim and rely on benchmark results.","tokens_in":22073,"tokens_out":10609,"duration_ms":96593,"concrete_test":"Re-run the Appendix A.1 text-grounded detection precision pipeline on all ScanNet-source Object Size questions (or a random subset) without the 'SFT > distillation' filter, using identical forward passes and adaptive thresholds, and report per-method precision for SFT, feature distillation, MVRD, and MVRD+DP. If the MVRD-vs-feature-distillation gap falls below about 0.04 (or reverses), the claimed representation-level alignment advantage is an artifact of question selection; the alignment story would then rest solely on task-level metrics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central differentiator over feature distillation is that MVRD retains vision-language alignment. The strongest representation-level evidence, the text-grounded detection precision in Figure 4b (0.16 vs 0.08), is computed on only 42 Object Size questions from ScanNet, selected, as Appendix A.1 states, 'where the SFT model performs better than distillation methods.' Conditioning on this event selects cases in which distillation has already failed relative to SFT, so the subset is not representative of the general distribution. If the selection criterion uses the feature-distillation baseline specifically, the MVRD-vs-feature-distillation comparison is stacked toward whichever method degrades less on precisely those failure cases, inflating the apparent gap. If it selects on SFT beating both distillation variants, the absolute precision numbers are still unrepresentative, and no random or full-sample comparison is reported. The benchmark task results (e.g., Appr. Order in Table 2) are not selection-biased and do show MVRD and MVRD+DP well above feature distillation, but they are only three tasks and depend on the specific QA format. The representation-level claim that relational targets keep the student 'close to the pretrained vision-language space' therefore lacks unbiased quantitative support. A second, independent weakness is that the headline VSI-Bench differences over SFT and feature distillation (1.4-2.2 points) have no error bars or significance tests; on Qwen2.5-VL (Table 6) MVRD ties SFT, so the 'outperforming SFT' statement is not universally supported. Neither issue disproves the method, but both must be addressed before full acceptance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes multi-view relational distillation (MVRD), which distills cross-view patch-wise cosine-similarity matrices from a frozen geometry-grounded teacher (VGGT) into a VLM, rather than distilling the teacher features directly. On VSI-Bench, the authors report that MVRD outperforms supervised fine-tuning and feature distillation across several backbones, approaches the feature-fusion method VLM-3R with far lower inference overhead, and transfers to 3D scene understanding tasks. The paper also presents a diagnosis of feature distillation's tendency to disrupt vision-language alignment, supported by a text-grounded object-detection precision probe.","tokens_in":22373,"tokens_out":3137,"duration_ms":30867,"significance":"If the central claim holds, MVRD offers a practical way to inject geometric structure into VLMs while preserving their pretrained vision-language alignment, at a fraction of the inference cost of feature fusion. The main benchmark comparison is well controlled: the same backbone, the same training data for MVRD and its direct baselines, and a reproduced 3DRS baseline that matches the original paper's trend. The paper also includes backbone generality experiments, a transfer study to 3D scene understanding, and detailed ablations of the distillation target and loss, which strengthens the empirical case. However, the representation-level evidence for alignment retention rests on a selection-biased probe, and the headline VSI-Bench gaps lack error bars, so the empirical support for the central mechanism is currently incomplete.","major_comments":[{"comment":"The text-grounded detection precision probe is computed on only 42 Object Size questions selected 'where the SFT model performs better than distillation methods.' This conditions on the outcome being compared: the subset is defined by the distillation methods being worse than SFT, so the precision gap (0.16 vs 0.08) is not representative of the general distribution and may be inflated. Because this probe is the primary representation-level evidence for the claim that MVRD retains vision-language alignment, please report the same metric on the full ScanNet subset of Object Size questions or on a random subset chosen independently of model performance, with per-question standard errors. This is load-bearing, since alignment retention is the central differentiator of MVRD over feature distillation.","section":"Appendix A.1, Figure 4(b)"},{"comment":"The headline claims that MVRD outperforms SFT and feature distillation by 1.4-2.2 points on average VSI-Bench accuracy are reported without error bars, significance tests, or multiple-seed variation. Given that VSI-Bench is a multi-task benchmark and per-task numbers fluctuate, the reported gaps may be within run-to-run noise. Please provide at least three seeds for the main comparisons (Tables 2, 3, and 6) or bootstrap confidence intervals over the benchmark questions, so the reader can judge the robustness of the 59.9-60.4 averages against the 57.7 and 58.5 baselines.","section":"Table 2, Table 6"},{"comment":"The underdetermination argument shows that if v* is optimal for the relational loss, then every shared rotation and independent scaling of v* is also optimal; hence a relational-loss-optimal solution near the pretrained vision-language space exists. This is an existence claim, not a statement that gradient-based minimization of LSFT + gamma*LRD will select such a solution. The unbiased alignment probe requested in the first major comment is therefore the necessary empirical support for 'allowing it to remain close.' Please either provide unbiased evidence that the trained student indeed stays near the pretrained space, or state explicitly what conditions on the loss landscape would guarantee convergence to such a solution.","section":"Section 4, Eq. (2)"}],"minor_comments":[{"comment":"The text says 'negative Person correlation'; this should be 'Pearson correlation.'","section":"Section 4"},{"comment":"The method name is inconsistent: Figure 1 uses 'MVRD + Dual Pathway' while Table 2 uses 'MVRD + Dual Pathways.' Please use a single consistent name throughout.","section":"Figure 1 caption and Table 2"},{"comment":"The phrase 'with considerably less added parameters' should be 'with considerably fewer added parameters.'","section":"Abstract"},{"comment":"Table 8 lists 'Scheduled training epochs 5 (stopped after epoch 1)' without describing the stopping criterion. Please state what validation signal triggers early stopping, or remove the scheduling of 5 epochs if only 1 epoch is ever run.","section":"Appendix C.1"},{"comment":"The question-selection sentence is ambiguous: does 'distillation methods' refer to feature distillation only, or both feature distillation and MVRD? Clarifying this matters because the selection criterion directly affects the interpretation of the probe results.","section":"Appendix A.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for the journal and the central idea is credible, provided the alignment-retention evidence is made unbiased. I would not require re-deriving the theory, but the selection-biased probe and the lack of error bars need to be addressed before publication. The 3D scene understanding transfer experiments are a nice addition and should be kept."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central idea here is worth taking seriously: instead of distilling 3D teacher features into a VLM, which the paper shows can break vision-language alignment, distill the cross-view cosine-similarity structure. The main benchmark comparisons are well controlled—same backbone, same training data, and a reproduced 3DRS baseline that tracks the original. MVRD beats SFT and feature distillation on VSI-Bench by a few points, and the transfer to 3D scene understanding is consistent. That is a genuine, useful result, and the method is cheap at inference compared to feature fusion.\n\nThe paper also does a good job of diagnosing a real failure mode: feature distillation from VGGT degrades linguistic-grounding tasks like object appearance order. The evidence that this happens is credible. The relational objective, with cosine similarity, row-wise Pearson loss, and causal masking, is well motivated, and the ablations show both choices matter. The Dual Pathway variant is a sensible way to keep the semantic stream clean.\n\nThe soft spots are mostly about the strength of the alignment-retention claim. The main representation-level evidence, text-grounded detection precision in Figure 4b, is computed on only 42 Object Size questions selected because the SFT model beats distillation methods. That selection stacks the deck and cannot support a general claim about MVRD preserving vision-language alignment. The benchmark tasks (Appr. Order, Obj. Size) are not selection-biased and do point in the right direction, but they are a handful of tasks and the gaps are small. Second, there are no error bars or significance tests on the 1.4–2.2 point gains; on Qwen2.5-VL MVRD ties SFT, so the 'outperforms SFT' statement is not universal. Neither issue disproves the method, but both need addressing. Code and artifacts would also help.\n\nWho is this for? Anyone working on VLMs for embodied AI, or on knowledge distillation for geometric grounding. It deserves a serious referee; the contribution is real, even if one piece of evidence is currently overstated. I would send it out and ask for a fix to the probe or a softened claim, plus error bars and code release.","headline":"A practical relational-distillation method for geometry grounding that beats feature distillation in controlled benchmarks, though its alignment-retention evidence leans on a selection-biased probe and needs corroboration.","tokens_in":22934,"tokens_out":2286,"would_cite":true,"duration_ms":20090,"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":"Distilling how patches relate across views, rather than the features themselves, makes vision-language models spatially smarter while keeping them aligned to language.","keywords":["multi-view relational distillation","vision-language models","spatial reasoning","geometric grounding","knowledge distillation","vision-language alignment","VSI-Bench","LoRA"],"falsifier":"Compute the distance (for example, mean cosine distance) between the student's vision features at the distillation layer before and after MVRD training, and measure text-based object-detection precision on a randomly sampled, unbiased set of VSI-Bench Object Size and Appearance Order questions. If the student drifts as far from the pretrained features as feature distillation does, or if the unselected-question precision drops to feature-distillation levels while the cross-view relation correlation remains high, then the underdetermination premise does not protect vision-language alignment in practice.","tokens_in":21858,"feed_emoji":"🧭","tokens_out":9688,"duration_ms":78561,"temperature":0.7,"pith_summary":"Vision-language models have strong image and video understanding, but their visual-spatial representations are geometrically fragile, which shows up as failures on spatial reasoning tasks. The paper identifies a previously undiagnosed failure of the usual remedy, feature distillation from a geometry-grounded model: matching multi-view teacher features grounds geometry but pulls the student's visual features away from the pretrained vision-language space, degrading linguistic grounding such as tracking object appearance order. Its proposal, multi-view relational distillation (MVRD), distills instead the pairwise cosine similarities among patch tokens across views, which encode geometric correspondence while leaving the student underdetermined, so the representation can stay close to the pretrained space. On VSI-Bench, MVRD reaches 59.9% average accuracy versus 57.7% for supervised fine-tuning and 57.0% for feature distillation; a dual-pathway variant reaches 60.4%, within 0.5 points of the feature-fusion method VLM-3R while adding roughly 25% of its parameters and 16% of its latency. If correct, relational distillation is a low-cost way to make VLM representations more geometric without giving up language alignment.","feed_headline":"Distill patch relations, not features, for VLM spatial skills","feed_subtitle":"MVRD hits 60.4% on VSI-Bench, within 0.5 points of feature fusion at a fraction of inference cost.","key_machinery":"The load-bearing object is the multi-view relational representation $R(\\mathbf{h})[i,j]=r(\\mathbf{h}[i],\\mathbf{h}[j])$, an $N\\times N$ matrix of pairwise similarities among all patch tokens from all input views, with $r$ chosen as cosine similarity and $N=T H_p W_p$. The student's projected relation matrix is aligned to the frozen teacher's relation matrix by a row-wise Pearson correlation loss (with causal masking and zeroed diagonal), trained jointly with the SFT loss. The property that does the work is underdetermination: because cosine similarity is invariant to global rotation and independent scaling of the feature set, any rotation or scaling of an optimal student representation is also optimal, leaving room for the optimizer to stay near the pretrained vision-language region. A dual-pathway variant strengthens this by using two LoRA adapters at layers 13-20, a geometric adapter supervised by the composite loss and a semantic adapter supervised only by SFT, whose outputs are summed before the remaining layers.","core_discovery":"The paper's central claim is that the bottleneck in grounding VLMs to geometry is not the amount of geometric signal but the choice of distillation target. Matching the teacher's per-patch features directly (feature distillation) pulls the student out of its pretrained vision-language space; matching the teacher's multi-view pairwise cosine similarities (relational distillation) transfers 3D correspondence information while leaving the student representation underdetermined, because cosine similarities are invariant to global rotations and independent per-feature scalings. The paper verifies that the geometry-grounded teacher VGGT's cross-view similarity is a meaningful proxy for 3D proximity (mean row-wise Pearson correlation 0.79 with Euclidean distance on ScanNet), shows that MVRD beats supervised fine-tuning and feature distillation on VSI-Bench, and demonstrates transfer to 3D object grounding, dense captioning, and question answering. With the dual-pathway variant, the geometric and semantic supervision are decoupled into separate LoRA adapters, and the result approaches the feature-fusion state of the art at a fraction of its inference cost.","pith_inferences":["If the underdetermination mechanism is the real source of the gain, the same relational recipe should transfer to other teachers whose cross-view similarities encode useful structure, such as depth, motion, or object tracks, not just VGGT-style geometry; the paper does not test this, but it follows directly from its argument.","Because the loss is invariant to global rotation and per-feature scaling, MVRD may be unusually robust to shifts in teacher feature statistics across datasets or pretrained checkpoints; that is a plausible but untested consequence.","The paper's alignment probe uses 42 questions chosen because SFT beats distillation, so the claimed alignment preservation should also be checked on a random, unbiased question sample; the paper's own protocol leaves this open.","The underdetermination idea suggests an explicit extension: add a small anchor regularizer that pulls the student toward its own pretrained features while matching relations, potentially improving alignment further than the implicit freedom alone; the paper does not explore this."],"forward_implications":["MVRD outperforms supervised fine-tuning and feature distillation on VSI-Bench (59.9 vs 57.7 and 57.0), and wins every linguistic-grounding task where feature distillation (3DRS*) falls below the SFT baseline.","MVRD + Dual Pathway reaches 60.4%, within 0.5 points of VLM-3R, with 0.19B added parameters versus 0.79B and 0.14s added latency versus 0.85s, about 25% of the parameters and 16% of the latency.","The gain is consistent across three VLM backbones (LLaVA-Video-7B, InternVL3-8B, Qwen2.5-VL-7B), and the dual pathway gives the largest marginal gain on Qwen2.5-VL-7B (+3.7 points over MVRD alone).","MVRD transfers to 3D scene understanding: trained under the 3DRS protocol, it improves on every ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, and SQA3D metric over the feature-distillation baseline.","Ablations show both design choices are load-bearing: cosine relations plus row-wise Pearson correlation reach 59.9, while Euclidean-distance relations with MSE fall to 48.6."],"supporting_citations":[{"why":"Supplies the spatial question-answering training dataset and the feature-fusion baseline VLM-3R that MVRD approaches.","marker":"[1]"},{"why":"The feature-distillation baseline (3DRS) whose failure mode MVRD diagnoses, and the scene-understanding training protocol MVRD adapts.","marker":"[2]"},{"why":"The LLaVA-Video-7B-Qwen2 backbone used for the main experiments.","marker":"[16]"},{"why":"The VSI-Bench spatial-reasoning benchmark and its task taxonomy used for evaluation.","marker":"[19]"},{"why":"The frozen VGGT teacher whose multi-view relations are distilled.","marker":"[44]"},{"why":"The CUT3R encoder behind the VLM-3R fusion baseline, used for the added-parameter and added-latency comparison.","marker":"[45]"},{"why":"Prior relational knowledge distillation, extended by MVRD to multi-view patch relations.","marker":"[52]"},{"why":"The LoRA adapters used to keep training close to the pretrained space and to implement the dual-pathway separation.","marker":"[59]"}],"fun_headline_variants":["Distill cross-view patch relations, not features, for VLM spatial skills","MVRD: relational distillation grounds VLMs without feature fusion","Multi-view patch similarities improve VLM 3D reasoning cheaply","Copying relations, not features, boosts VLM spatial reasoning","MVRD: geometry from underdetermined relations, not features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benefit depends on the premise that matching only cross-view angle relations leaves the student's features free to remain near the pretrained vision-language space; the paper shows the loss is rotation- and scale-invariant but does not prove that gradient training actually picks such a solution, and its main quantitative evidence for preserved alignment comes from 42 questions selected because SFT outperforms distillation.","fun_headline_variants_meta":{"raw":{"variants":["Distill cross-view patch relations, not features, for VLM spatial skills","MVRD: relational distillation grounds VLMs without feature fusion","Multi-view patch similarities improve VLM 3D reasoning cheaply","Copying relations, not features, boosts VLM spatial reasoning","MVRD: geometry from underdetermined relations, not features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000806,"raw_usage":{"total_tokens":3570,"prompt_tokens":1005,"completion_tokens":2565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2474}},"tokens_in":621,"tokens_out":2565,"duration_ms":16146,"temperature":1.0,"reasoning_tokens":2474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:28:39.203170+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the distance (for example, mean cosine distance) between the student's vision features at the distillation layer before and after MVRD training, and measure text-based object-detection precision on a randomly sampled, unbiased set of VSI-Bench Object Size and Appearance Order questions. If the student drifts as far from the pretrained features as feature distillation does, or if the unselected-question precision drops to feature-distillation levels while the cross-view relation correlation remains high, then the underdetermination premise does not protect vision-language alignment in practice.","supporting_citations":[{"cited_title":"LLaV A- video: Video instruction tuning with synthetic data.Transactions on Machine Learning Research, 2025","cited_arxiv_id":null,"evidence_quote":"The LLaVA-Video-7B-Qwen2 backbone used for the main experiments."},{"cited_title":"Continuous 3d perception model with persistent state","cited_arxiv_id":null,"evidence_quote":"The CUT3R encoder behind the VLM-3R fusion baseline, used for the added-parameter and added-latency comparison."}],"review_version":1}