{"id":"ac612825-9e0c-4119-8384-9909453bf16c","arxiv_id":"2507.20738","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A unimodal student model, taught by reinforcement-selected combinations of multimodal teachers via neighbor-decoupled knowledge distillation, sets new state-of-the-art results on five multimodal knowledge graph reasoning datasets.","lead":"This paper trains a single knowledge-graph model to imitate the combined predictions of separate structure, text, and image models, and uses a small reinforcement-learning agent to choose which of those 'teacher' models to imitate for each query. The resulting student outperforms existing multimodal knowledge-graph reasoning methods on five datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (8)'s reward is non-stationary and binary, so the reported superiority over BestStrategy in Table 2 is not explained by the reinforced-selection mechanism and may be a co-training artifact.","rationale":"The reader's weakest assumption identifies the non-stationarity of the Eq. (8) reward, which is indeed the core issue. I agree with that diagnosis, but I would sharpen it: even if the reward were stationary, its binary form makes it unable to prefer the most informative teacher combination over a merely acceptable one, so the paper's explanation for why DSoM beats BestStrategy is internally weak. This is a genuine concern about the RC mechanism's contribution, and the lack of error bars and promised-but-unreleased code makes it hard to rule out empirically. However, the concern is not fatal to the empirical headline: the NDKD component alone already reaches 41.91 MRR on DB15K versus 37.72 for MyGO, so the central SOTA claim does not collapse if RC is disregarded. The right verdict remains conditional: the authors should release code, report seeds/error bars, and run the fixed-policy ablation to show that the RC gain is not an artifact of the co-training loop. I do not see an internally inconsistent derivation or a clearly unfair baseline comparison that would force rejection.","tokens_in":21808,"tokens_out":10554,"duration_ms":146189,"concrete_test":"On DB15K, train the RC policy with rewards computed against a fixed reference instead of the live student—for example, use CE(Teacher Avg.) as the threshold, or freeze a copy of the student at epoch 1—then freeze that policy and train the student with NDKD using the selected teacher subsets. Compare final MRR with the reported 42.68 and with the BestStrategy row of 41.65. If the fixed-policy MRR falls to roughly 41.6 or below, the RC gain depends on the non-stationary co-training reward. As a second check, record the cross-entropy of the teacher subsets selected by the trained policy on training triples and compare it with CE(BestStrategy); if the policy's selected subsets are not lower-cross-entropy than BestStrategy, the reward-maximization explanation is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The RC module's only supervision is the reward in Eq. (8): a teacher subset is labeled beneficial if its cross-entropy on the current training triple is lower than the student's current cross-entropy, with a +1/-10 binary payoff. This creates three linked problems. First, the reward is non-stationary: the student is updated in the same loop, so the comparison target moves, and the policy is trained against a criterion that depends on the very model it is supposed to improve. Second, the reward is myopic: it measures fit to the same training triple used for the student's gradient step, not whether the selected soft labels improve held-out ranking, so the policy can be driven by the student's memorization rather than by genuine modality quality. Third, the reward is binary: any action with cross-entropy below the student receives the same +1, so maximizing expected reward does not prefer the most informative soft labels over a barely-better confident teacher. The deterministic greedy BestStrategy (lowest cross-entropy) should therefore be at least as good under this reward, yet Table 2 reports DSoM beating BestStrategy by about 1 MRR on DB15K (42.68 vs 41.65). The paper attributes this to the agent finding the set 'best for the student,' but the reward cannot distinguish among equally rewarded strategies, so the advantage must come from stochasticity, exploration, or student co-evolution, none of which is analyzed. Since Table 4 shows RC adds only ~0.8 MRR over NDKD alone, the headline SOTA claim is not solely dependent on RC, but the claimed reinforced-selection contribution is not robustly supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DSoM, a framework for multimodal knowledge graph reasoning (MKGR). DSoM pre-trains three unimodal teachers (structural, visual, textual) and then trains a unimodal student KGR model by knowledge distillation from the teachers' soft labels. A novel reinforced teacher combination module selects, per triple, a subset of teachers via a policy network trained with a binary reward based on cross-entropy comparison with the student. The distillation loss is decoupled into neighbor and non-neighbor entity terms, extending Decoupled KD to MKGR's multiple true answers. Experiments on DB15K, MKG-W, MKG-Y, FB15K-237, and WN18 (plus WN9 in an appendix) report consistent gains over prior MKGR baselines, with the unimodal ComplEx student reaching 42.68 MRR on DB15K versus 37.72 for the previous best MyGO.","tokens_in":22098,"tokens_out":6307,"duration_ms":61938,"significance":"If the empirical results hold, DSoM makes a valuable conceptual contribution: instead of fusing modalities into a multimodal model, it trains a unimodal student on multimodal soft labels, which can exploit inter-entity label correlations and avoid negative transfer from unhelpful modalities. The gains are large and consistent on three medium-scale datasets, and the efficiency experiments (robustness to missing modalities and reduced embedding dimensions) are a genuine strength. The neighbor-decoupled distillation is a sensible adaptation of DKD to KGs. However, the paper is currently weakened by the absence of statistical validation, the incomplete large-scale comparisons, and a reinforcement reward design that is not analyzed well enough to explain the claimed advantage over greedy selection. These issues are fixable and do not, in my view, invalidate the central distillation idea, but they are load-bearing for the SOTA claim.","major_comments":[{"comment":"The central claim of state-of-the-art performance rests entirely on single-run results: no standard deviations, no number of seeds, and no significance tests are reported. On several metrics the advantage is small (e.g., WN18 Hits@10 is 97.4 vs. SGMPT's 97.8, and FB15K-237 Hits@1 is 27.5 vs. LAFA's 26.9). Without error bars or paired tests, the superiority over strong baselines is not statistically established. Please report means and standard deviations over at least three random seeds, and perform paired significance tests for the key comparisons in Tables 1 and 3.","section":"Section 5.1, Tables 1–3"},{"comment":"The reward in Eq. (8) is non-stationary (the student is updated in the same loop as the policy) and binary (any teacher subset with lower cross-entropy than the student receives the same +1). Under this reward, a greedy BestStrategy that always selects the lowest-cross-entropy subset should be at least as good, yet DSoM beats BestStrategy by about 1 MRR on DB15K (42.68 vs. 41.65). The paper does not analyze the learned policy—for example, how the selected subsets differ from BestStrategy, or whether the advantage comes from exploration, stochasticity, or student co-evolution. Please add such an analysis, or an experiment with a stationary reward from a frozen student, to substantiate the claim that the reinforced module (rather than side effects of co-training) drives the improvement.","section":"Section 4.4, Eq. (8), and Table 2"},{"comment":"The ablation 'w/ RC on val.' trains the reinforced teacher combination on the validation set and then transfers it to the test set, whereas the full DSoM trains RC jointly during student training. In addition, the full model combines RC and NDKD, so the protocol mismatch makes it difficult to attribute the improvement to the reinforcement mechanism versus the co-training dynamics. Please run a joint-training ablation of RC alone (without NDKD) and an NDKD-only ablation under the same protocol, or explain why the validation-based protocol is equivalent to joint training.","section":"Section 5.3, Table 4"},{"comment":"The large-scale comparison is incomplete: several baselines have missing values (e.g., MyGO's MR and Hits@10, AdaMF-MAT's MR), and the numbers are taken from different publications that may use different evaluation protocols or filtered/unfiltered settings. The claim of superiority on FB15K-237 and WN18 is therefore not fully supported. Please provide a complete comparison under a single evaluation protocol, or clearly mark the source and condition of each baseline number and restrict the SOTA claim to metrics with complete and comparable entries.","section":"Section 5.2, Table 3"},{"comment":"The abstract states that 'Codes are available at github.com/OreOZhao/DSoM', but Appendix A says 'We will open-source our code on GitHub after the paper is accepted.' This is a direct contradiction and undermines the reproducibility claim. Please reconcile the availability statement, and in the meantime provide the exact per-dataset hyperparameters (learning rate, batch size, epochs, optimizer) and the settings for gamma, tau, alpha, and beta so that the experiments can be reproduced.","section":"Abstract and Appendix A"}],"minor_comments":[{"comment":"The paper says '5 datasets' in the Abstract and Section 5.1, but Appendix C states that experiments are conducted on 6 datasets and Table 5 includes WN9. Please make the count consistent throughout.","section":"Abstract and Section 5.1 vs. Appendix C"},{"comment":"The appendix header reads 'Pipeling of DSoM'; this should be 'Pipeline of DSoM'.","section":"Appendix A"},{"comment":"The sentence 'This strategy may not be the best strategy itself, but it is the best for the student' is vague; please define operationally what 'best for the student' means, for example in terms of final held-out MRR or a separate reward computed on validation triples.","section":"Section 5.4"},{"comment":"The asymmetric +1/-10 reward scale is not discussed; since the magnitude of the penalty can affect the entropy of the policy and the training dynamics, please justify this choice or provide a sensitivity analysis.","section":"Section 4.4, Eq. (8)"},{"comment":"The label 'SOTA(MyGO)' is imprecise because MyGO is not the best model on every metric or dataset; please use a more neutral label such as 'MyGO (best published baseline)'.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising central idea and the empirical trend is consistent, but the two most important weaknesses are the missing statistical validation and the unsubstantiated advantage of the reinforced selection over greedy heuristics. Both are fixable with additional experiments. The inconsistency between the abstract's claim of available code and the appendix's 'will open source after acceptance' should be corrected before publication; this is the kind of discrepancy that can cause post-publication complaints. Given that the reader's stress-test concern about Eq. (8) largely lands, I recommend major revision rather than rejection: the distillation and neighbor-decoupling contributions appear sound, and the RC module's role can be clarified with focused analyses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2507.20738. The neighbor-decoupled KD is a genuine and apparently effective idea: training a unimodal student on multimodal soft labels, split into neighbor and non-neighbor terms, gives consistent gains across five datasets. The reinforced teacher-selection module is the soft spot. Its reward is non-stationary and myopic, and the claimed advantage over a greedy baseline is not convincingly explained.\n\nWhat's new: the paper is the first to distill multimodal logits into a unimodal student for MKGR, and the first to frame teacher selection as discrete RL rather than continuous weighting. The NDKD extension of Decoupled KD to multiple true neighbors is reasonable, and the ablations show it beats vanilla KD and target-only DKD. The experimental coverage is broad: five datasets, ablations, efficiency analysis, hyperparameter study.\n\nThe problems are all around the RC module. Equation (8) labels a teacher subset as beneficial if its cross-entropy on the current training triple is below the student's own cross-entropy, with a +1/-10 payoff. The student is being updated in the same loop, so the comparison target moves; the reward is also myopic because it measures fit to the training triple, not held-out ranking. And binary rewards cannot distinguish among equally rewarded teacher sets. Under this reward, the deterministic greedy BestStrategy—lowest cross-entropy—should be at least as good as any learned policy. The paper reports DSoM beating BestStrategy by about 1 MRR on DB15K, but never explains why. Table 4 shows RC adds only ~0.8 MRR over NDKD alone, so the headline SOTA claim is not dependent on RC, but the RC contribution as a novel mechanism is not robustly supported. I'd also like to see error bars or seeds; the abstract promises code, but the appendix says it will be released after acceptance. These are addressable issues, not fatal ones.\n\nWho is this for: people working on multimodal knowledge graphs or knowledge distillation. The NDKD technique is worth citing; the RL selection needs a stronger analysis. Send it to a serious referee. The reviewers should push on the reward design and ask for reproducible numbers.","headline":"The neighbor-decoupled distillation is the real contribution; the reinforced teacher selection is under-justified and adds only a small margin, but the paper deserves a serious read.","tokens_in":22725,"tokens_out":3124,"would_cite":true,"duration_ms":33920,"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":"A unimodal student trained on soft labels from a reinforcement-selected subset of multimodal teachers beats adaptive-fusion MKGR baselines on five datasets.","keywords":["multimodal knowledge graph reasoning","knowledge distillation","reinforcement learning","teacher selection","soft labels","entity label correlations","knowledge graph completion"],"falsifier":"Compare the learned policy's teacher choices on held-out triples with the teacher subset that actually minimizes cross-entropy on those triples; if agreement is near chance, or if replacing the learned agent with a static per-triple best-teacher-by-validation-CE heuristic reproduces the same MRR, the reinforcement component is not doing the claimed work.","tokens_in":21558,"feed_emoji":"🧩","tokens_out":6237,"duration_ms":71830,"temperature":0.7,"pith_summary":"This paper claims that multimodal knowledge graph reasoning is best learned not by fusing all modalities into one model, but by training a plain unimodal student to imitate soft label distributions from pre-trained structural, visual, and textual teachers. The soft labels carry 'dark knowledge': correlations between the target entity and other plausible answers, and between distractors, which one-hot cross-entropy discards. The paper further claims that, for each training triple, some modalities are actively misleading, so a reinforcement-trained agent should discretely choose which teachers to keep rather than down-weight the bad ones. On five MKGR benchmarks (DB15K, MKG-W, MKG-Y, FB15K-237, WN18) the resulting DSoM student reports state-of-the-art MRR, Hits@{1,3,10}, and Mean Rank.","feed_headline":"Deleting unhelpful modalities beats down-weighting them","feed_subtitle":"A reinforced-distillation student learns from chosen teachers' soft labels and beats adaptive fusion on five benchmarks.","key_machinery":"Two coupled mechanisms carry the argument. The first is a reinforced teacher-combination policy: an MLP agent that sees the concatenated score distributions of structural, visual, and textual teachers, samples a subset from the seven possible teacher combinations, and is trained by a REINFORCE-style policy gradient whose reward compares the selected teachers' cross-entropy against the student's. The second is neighbor-decoupled knowledge distillation, an extension of decoupled KD that splits temperature-scaled teacher and student logits into an average binary probability over all neighbor entities (true answers) and a distribution over non-neighbor entities, so the student learns both types of label correlation.","core_discovery":"The paper introduces DSoM, a framework that replaces one-hot hard-label training of a multimodal knowledge graph reasoning model with a distillation routine in which a unimodal ComplEx student learns to imitate soft label distributions emitted by separately pre-trained structural, visual, and textual teacher models. A small policy network receives the teachers' score distributions as state and samples a subset of teachers for each triple, receiving a positive or negative reward depending on whether the averaged chosen teacher logits beat the student's own cross-entropy on that triple. The student is then trained with hard-label cross-entropy plus a neighbor-decoupled knowledge distillation loss that splits teacher and student logits into neighbor and non-neighbor entity groups, so the student learns both correlations among true answers and correlations among distractors. The reported result is consistent gains over adaptive-fusion and ensemble baselines, with the unimodal student reaching 42.68 MRR on DB15K versus the previous best 37.72.","pith_inferences":["Extension beyond the paper: the non-stationary reward means the learned policy could be a bystander; a simpler per-triple selector that picks the teacher combination with lowest validation-set cross-entropy would settle whether reinforcement is essential to the reported gains.","Neighbouring problem: the neighbor-decoupled objective is a general treatment of link prediction with multiple true answers, so other knowledge graph completion settings with plural true tails could reuse it without the multimodal framing.","Engineering consequence: because the student is unimodal and can run at 1.72M parameters while outperforming a 22.0M-parameter baseline, the framework licenses replacing heavy multimodal encoders at serving time with a compact embedding model.","Testable extension: training the same student with the policy frozen after the first epoch, versus letting it co-train throughout, would reveal how much of the gain comes from the agent adapting to the student rather than from modality quality alone."],"forward_implications":["A unimodal student can carry multimodal knowledge: at inference, only the student's entity and relation embeddings score triples, so multimodal encoders are needed only during training.","Discrete, per-triple exclusion of a modality removes misleading signals entirely, which the paper argues is stronger than the low-weight treatment used by gating and attention fusion.","Soft-label distillation supplies supervision about all entity candidates, so correlations among true answers and among distractors are learned even when they never appear as targets in a one-hot objective.","The framework degrades gracefully: with up to 80% of visual and textual modalities missing, and at embedding dimensions as low as 64, the student still beats the strongest full-size baseline on DB15K.","The recipe is modality-agnostic: adding another teacher modality only enlarges the policy pool, so the same reinforced distillation could absorb audio, video, or other future modalities."],"supporting_citations":[{"why":"Introduces knowledge distillation, the soft-label supervision mechanism the student uses.","marker":"[14]"},{"why":"Provides ComplEx, the score function used for the teachers and the unimodal student.","marker":"[40]"},{"why":"Contributes the multi-teacher KD baseline (MulDE) that DSoM is compared against.","marker":"[42]"},{"why":"Supplies the REINFORCE-style policy gradient used to train the teacher-combination agent in Equation (9).","marker":"[48]"},{"why":"Provides the MyGO model, the strongest adaptive-fusion baseline and the efficiency comparison partner.","marker":"[58]"},{"why":"Supplies decoupled knowledge distillation, which DSoM extends to neighbor entities.","marker":"[64]"},{"why":"Provides MoSE, the modality-split pretraining and teacher-average baseline used for the ablation study.","marker":"[65]"}],"fun_headline_variants":["RL teacher choice outperforms adaptive multimodal fusion","Distill dark knowledge from selected teachers only","Reinforced distillation beats static multimodal baselines","Per-triple teacher selection via RL improves KG reasoning","Drop unhelpful modalities with a policy network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The agent is taught only by whether the averaged chosen teachers have lower loss on a training triple than the student currently does; if that signal tracks the student's memorization instead of genuine modality quality, the selection mechanism and the reported gains could come from the co-training loop itself.","fun_headline_variants_meta":{"raw":{"variants":["RL teacher choice outperforms adaptive multimodal fusion","Distill dark knowledge from selected teachers only","Reinforced distillation beats static multimodal baselines","Per-triple teacher selection via RL improves KG reasoning","Drop unhelpful modalities with a policy network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1715,"prompt_tokens":1006,"completion_tokens":709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":639}},"tokens_in":622,"tokens_out":709,"duration_ms":9551,"temperature":1.0,"reasoning_tokens":639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:18:56.765673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the learned policy's teacher choices on held-out triples with the teacher subset that actually minimizes cross-entropy on those triples; if agreement is near chance, or if replacing the learned agent with a static per-triple best-teacher-by-validation-CE heuristic reproduces the same MRR, the reinforcement component is not doing the claimed work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides ComplEx, the score function used for the teachers and the unimodal student."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the multi-teacher KD baseline (MulDE) that DSoM is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the REINFORCE-style policy gradient used to train the teacher-combination agent in Equation (9)."},{"cited_title":"Tokenization, Fusion, and Augmentation: Towards Fine-grained Multi-modal Entity Representation","cited_arxiv_id":"2404.09468","evidence_quote":"Provides the MyGO model, the strongest adaptive-fusion baseline and the efficiency comparison partner."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies decoupled knowledge distillation, which DSoM extends to neighbor entities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides MoSE, the modality-split pretraining and teacher-average baseline used for the ablation study."}],"review_version":1}