{"id":"69de7a1b-8ee1-44d2-8aa9-339c44f0630e","arxiv_id":"2508.01590","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"MCG-IMM uses evolutionary multi-criteria optimization during sampling to increase the intra-batch diversity of in-betweening human motions generated by any pretrained model.","lead":"This paper proposes a plug-and-play method called MCG-IMM that guides a pretrained human motion generation model to produce more diverse and smoother in-betweening transitions, using a multi-objective optimization loop at sampling time. It claims to work across diffusion, VAE, and GAN backbones without retraining, and reports state-of-the-art diversity on four motion datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim hinges on Eq. (8)'s conditioning step, but the paper never specifies how a pretrained VAE or GAN accepts a full motion sequence Y_{i-1} as input; without this, the plug-and-play claim for these backbones is unsupported.","rationale":"The reader's weakest assumption is indeed the one on which the paper's central claim rests. The abstract's \"plug-and-play ... without introducing additional parameters\" and Section III.C's summary both depend on Eq. (8). The paper gives no architectural recipe; Section III.B's promise of Appendix details is unfulfilled. This is not a mere presentation gap: if the conditioning on Y_{i-1} requires training the backbone, then the central claim is contradicted; if it doesn't, the evolutionary loop may have no effect beyond resampling. Either way the reported improvements (Table I, especially APD) cannot be cleanly attributed to MCG-IMM. I also note the classifier used in the diversity criterion may be the same pretrained action-recognition model used for ACC/FID, which introduces a circularity, but the conditioning gap is the more load-bearing issue because it invalidates the method's stated generality. Therefore I retain the reader's REJECT verdict.","tokens_in":15938,"tokens_out":6043,"duration_ms":67144,"concrete_test":"Release the missing Appendix/code, then on HumanAct12 run Algorithm 1 for MCG-IMM(VAE) and MCG-IMM(GAN) with two fixed parents Y^{(a)} and Y^{(b)} at iteration i=1. Generate offspring sets from G(·|Y^{(a)}, X1, X2) and G(·|Y^{(b)}, X1, X2); if the two sets are statistically indistinguishable (e.g., FID/MMD below the run-to-run noise), Eq. (8) is not a real conditioning operation. Independently, inspect the released checkpoints/architecture to confirm whether the backbone was trained to accept a full motion-sequence condition; if any fine-tuning was used, the \"without retraining\" claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Eq. (8) / Algorithm 1 line 4: each offspring batch is drawn from G(Y_i | Y_{i-1}^{(l')}, X1, X2). For this to be a well-defined plug-and-play operation, the pretrained backbone must accept a full motion sequence Y_{i-1} as a conditioning input. The paper only asserts this and defers the construction to \"the Appendix\" (Section III.B), but no appendix is present in v1. For DDPM this could be realized by masking/inpainting, but no masking schedule or noise formulation is given; for VAE and GAN, a standard pretrained generator accepts a latent code plus a label/keyframe condition, not a previously generated sequence. If the backbones need to be trained or fine-tuned to consume Y_{i-1}, the headline \"no retraining / no additional parameters\" is false; if they are not trained to consume it, the offspring distribution is independent of the elite parent, and the evolutionary selection loop degenerates to repeated i.i.d. sampling from G(Y|X1,X2). In neither case is the reported APD improvement attributable to the claimed multi-criteria guidance. The missing appendix and absence of code make this impossible to verify from the paper alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MCG-IMM, a plug-and-play multi-criteria guidance method for diverse in-betweening human motion generation. It reformulates the sampling process of pretrained generative models (DDPM, VAE, GAN) as a multi-criteria optimization problem with objectives for diversity and smoothness, and solves it using an evolutionary algorithm with nondominated sorting and crowding distance. Experiments on BABEL, HumanAct12, NTU RGB-D, and GRAB are reported, and the paper claims consistent state-of-the-art performance without retraining or additional parameters.","tokens_in":16248,"tokens_out":7460,"duration_ms":84356,"significance":"If the central claim were established, the method would be a valuable lightweight add-on to existing generative models for motion in-betweening, addressing a real need for intra-batch diversity. The paper's ambition to be model-agnostic is attractive, and the inclusion of ablations and the attempt at theoretical support are commendable. However, the manuscript as written does not provide the necessary specification of the conditioning mechanism, and the evaluation is confounded by using the same classifier in both the objective and the metrics.","major_comments":[{"comment":"The offspring generation step conditions the generative model on a full previously generated motion sequence Y_{i-1}^{[l']}. For DDPM this could be realized by inpainting or classifier guidance, but no such mechanism is specified; for VAE and GAN, a standard pretrained generator does not accept a full motion sequence as a conditioning input. The paper refers to \"the Appendix\" for details, but no appendix is included in v1. If the backbones are not conditioned on Y_{i-1}, the offspring distribution is independent of the elite parent and the evolutionary selection degenerates to repeated i.i.d. sampling; if they are, the backbone must have been trained or modified to accept this input, contradicting the \"no retraining / no additional parameters\" claim. Either way, the reported APD improvements are not attributable to the proposed multi-criteria guidance.","section":"III.B, Eq. (8), Algorithm 1 line 4"},{"comment":"The evaluation metrics ACC and FID are computed using the same pretrained action classifier that is used in the diversity objective via C(Y) and P_c(Y). This is circular: the optimization directly moves generated motions toward the classifier's preferred categories and feature distribution, so the reported ACC and FID gains are at least partly an artifact of optimizing the evaluation signal. An independent evaluation classifier or metrics not derived from the same model are needed.","section":"IV.C and Eq. (1)"},{"comment":"The diversity component is ill-typed: C(Y) is an integer in {0,...,D-1}, while P_c(Y) is described as \"the probability that a given motion sequence belongs to each categorized motion type,\" which would be a D-dimensional vector; adding them is undefined. If P_c(Y) is instead the scalar probability of the predicted class, the notation should say so explicitly. Moreover, both α1 and α2 are per-sample functions; minimizing F1 and F2 does not directly measure or encourage pairwise diversity within the batch, so the connection between the multi-criteria problem and the APD metric is not established.","section":"III.A, Eq. (1)"},{"comment":"These results are near-tautological consequences of the definitions (α1+α2=1 and the form of β) and do not provide independent support for the diversity claim. Theorem 1 only asserts that minimizers of β are Pareto optimal, and Theorem 2 only states that two solutions in B with sufficiently distant objective vectors have different class labels. Neither theorem guarantees that the evolutionary algorithm will find such solutions or that the resulting batch will be diverse in the sense measured by APD.","section":"III.A, Theorems 1 and 2"},{"comment":"The empirical evidence is incomplete and overstates the case for consistency. No error bars or multiple seeds are reported. On NTU RGB-D, MCG-IMM (VAE) has FIDtr 81.64 versus 75.46 for MoFusion; on GRAB, MCG-IMM (VAE) has ACC 78.5 versus 79.21 for MoFusion. Also, the underlying pretrained backbones for the three MCG-IMM variants are not identified, so the comparison across methods is not controlled; differences in base models may account for the reported improvements.","section":"Table I, Section V.A"}],"minor_comments":[{"comment":"The symbol β is used both for the diffusion noise schedule in Section IV.B and for the smoothness function in Eq. (2), which is confusing and should be disambiguated.","section":"Section IV.B vs. Eq. (2)"},{"comment":"The phrase \"within a bath process\" appears to be a typo for \"within a batch process.\"","section":"Section V.A"},{"comment":"The final sentence of the abstract is grammatically incomplete; it should read something like \"consistently outperforms state-of-the-art methods in the in-betweening motion generation task.\"","section":"Abstract"},{"comment":"Equation (11) uses Y for the ground truth and Ŷ for the prediction, while elsewhere Y denotes generated sequences; the notation should be unified.","section":"Eq. (11)"},{"comment":"The text repeatedly refers to an appendix for details, but the v1 manuscript contains no appendix; this should be fixed before resubmission.","section":"Remark 1, Section III.B"},{"comment":"The claim \"without introducing additional parameters\" should be qualified to \"no additional trainable parameters,\" since the classifier C(Y) has its own parameters, even if they are frozen.","section":"Abstract and Section I"}],"recommendation":"reject","confidential_remarks":"The manuscript appears incomplete (the referenced appendix is missing), and the central claim is not supported as written. The unspecified conditioning mechanism and the circular evaluation are fundamental technical gaps that cannot be patched locally. The paper would need substantial reformulation and new experiments to be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is fresh: use NSGA-II-style nondominated sorting and crowding distance to guide the sampling of pretrained generative models for in-betweening motion, with two hand-designed criteria for diversity and smoothness. That specific combination is not in the cited literature, and the ablations show consistent APD gains across three backbones and four datasets. The variable-length transition mechanism and the Pareto-front visualization are also nice touches.\n\nThe problem is that the central mechanism is underspecified. Equation (8) and Algorithm 1 line 4 draw offspring conditioned on a previously generated full sequence Y_{i-1}, but the paper never says how a pretrained VAE or GAN accepts a full sequence as conditioning. The appendix is referenced but absent in v1. If the backbone ignores Y_{i-1}, the evolution loop degenerates to repeated i.i.d. sampling followed by selection, which could still improve APD but not for the reasons claimed. This is a load-bearing gap, not a minor omission.\n\nThe diversity criterion in Eq. (1) is also formally sloppy: it adds an integer class label C(Y) to a probability P_c(Y), and the intra-class term is never defined as a scalar. Theorems 1 and 2 are essentially restatements of the definitions; they add no real support. On evaluation, the same pretrained action classifier appears in the guidance objective and in the ACC/FID metrics, so the method directly optimizes the signal it is later judged on. No error bars, no code, no appendix—replication is impossible from the paper alone.\n\nThat said, the empirical improvements are consistent and the idea is salvageable. The flaws are fixable in principle: specify the conditioning mechanism (or retrain the backbones and drop the \"no retraining\" claim), fix Eq. (1), and use a held-out evaluator. As it stands, I would not accept it, but I would send it to a serious referee with those concerns front and center. The novelty justifies referee time; the presentation does not yet justify publication.","headline":"A genuinely new sampling-time evolutionary recipe for in-betweening motion, but the key conditioning step is unspecified and the evaluation is partly circular, so the claims outrun the evidence.","tokens_in":16686,"tokens_out":2282,"would_cite":false,"duration_ms":27426,"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":"Treating a frozen motion generator's sampling as a two-criteria Pareto search can raise the diversity of the motions it produces, with no retraining or added parameters, the paper argues.","keywords":["human motion in-betweening","diversity","multi-criteria optimization","Pareto optimality","plug-and-play guidance","diffusion models","evolutionary sampling","motion generation"],"falsifier":"Apply MCG-IMM to a pretrained VAE or GAN backbone whose forward pass accepts only keyframe poses and has no conditioning pathway for a full previously generated sequence; if that backbone cannot produce the offspring described in the generation equation, the method's model-agnostic plug-and-play claim fails in practice.","tokens_in":15804,"feed_emoji":"🏃","tokens_out":5168,"duration_ms":58774,"temperature":0.7,"pith_summary":"This paper tries to establish that the diversity of in-betweening human motions produced by pretrained generative models can be improved at sampling time, without retraining or adding parameters. The authors reformulate the sampling process as a multi-criteria optimization problem with two objectives—one that encourages variety across and within action categories and one that enforces smooth joins to the given keyframes—and then search for Pareto-optimal motion sequences using an evolutionary loop. If this works, it matters because intra-batch diversity is a known weakness of conditional motion generators, and the guidance is designed to be model-agnostic across DDPM, VAE, and GAN backbones. The paper reports consistent improvements in diversity and accuracy on four human motion datasets.","feed_headline":"Pareto-guided sampling adds diversity to frozen motion generators","feed_subtitle":"A no-retrain optimizer steers DDPM, VAE, and GAN samplers toward smoother, more varied in-betweening transitions.","key_machinery":"The central object is the multi-criteria optimization formulation with objectives $F_1(Y)=\\alpha_1(Y)+\\beta(Y)$ and $F_2(Y)=\\alpha_2(Y)+\\beta(Y)$, where $\\alpha_1(Y)=\\frac{1}{D}(C(Y)+P_c(Y))$, $\\alpha_2(Y)=1-\\alpha_1(Y)$, and $\\beta(Y)=\\|X_1[-1]-Y[0]\\|+\\|Y[-1]-X_2[0]\\|$ measures endpoint smoothness. The diversity component $\\alpha_1$ couples a classifier's category label $C(Y)$ with its class probabilities $P_c(Y)$, so that both inter-class and intra-class differences are rewarded. The theorems guarantee that minimizers of $\\beta$ are Pareto-optimal and that separating solutions along the Pareto front in objective space separates them by action category, which is what turns an optimization loop into a diversity engine.","core_discovery":"On the paper's own terms, the central claim is that any pretrained in-betweening motion generator can be steered toward more diverse and still smooth outputs by turning its sampling into a two-criteria search. The first criterion uses a frozen action classifier's labels and class probabilities to separate both different action categories and subtle variations within a category; the second measures the discontinuity at the two endpoints where the generated transition meets the user-provided sequences. Solving this two-objective problem with an evolutionary process—generating offspring motions from elite sequences via the frozen generator, then selecting by nondominated sorting and crowding distance—yields a batch of transitions that are mutually diverse and temporally coherent, with no additional training parameters. Two supporting theorems state that any motion minimizing the smoothness term is Pareto-optimal, and that two Pareto-optimal motions whose objective-space distance exceeds $4/D$ must belong to different action classes. The empirical claim is that this guidance consistently improves both diversity and accuracy over prior methods on BABEL, HumanAct12, NTU RGB-D, and GRAB, using DDPM, VAE, or GAN as the frozen backbone.","pith_inferences":["The plug-and-play claim quietly depends on the backbone accepting a full previously generated sequence as a conditioning input; generative models that only condition on keyframes or latent codes would need a wrapper to satisfy the offspring-generation step, which the paper does not specify.","Because the diversity term is driven by a classifier's category probabilities, the method's diversity ceiling is set by the classifier's ability to perceive differences; on noisy datasets such as NTU RGB-D, the guidance may struggle, which matches the weaker reported gains there.","The evolutionary loop is essentially using the frozen generator as a mutation operator, an idea that could transfer to other autoregressive or diffusion-based sequence tasks where a pretrained conditional model plus a classifier are available.","The two theorems rely on $\\alpha_1+\\alpha_2=1$ exactly, so any normalization change in the classifier probabilities would break the Pareto-optimality proof and the construction would need reworking."],"forward_implications":["Wrapping an existing pretrained in-betweening model with MCG-IMM should improve intra-batch diversity immediately, without the cost of fine-tuning or architectural changes.","Diversity and smoothness become inference-time controls: adjusting the criteria or the number of optimization iterations tunes the output batch without touching the trained weights.","The same Pareto-search wrapper could be dropped onto any conditional generative sequence model that accepts an additional conditioning sequence alongside the fixed keyframes, making the approach a general sampling-time plugin.","The diversity ceiling is set by the classifier's granularity: a more fine-grained action classifier would let the optimizer separate motions that share a coarse action label.","Variable-length transitions are handled by predicting the transition length from the cosine similarity of the endpoint poses, avoiding fixed-length assumptions in the backbone."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model backbone whose stochastic sampling MCG-IMM wraps and diversifies.","marker":"[1]"},{"why":"Provides the conditional human motion diffusion prior used to build the DDPM in-betweening backbone and documents the diversity shrinkage in conditional settings.","marker":"[12]"},{"why":"MoFusion is the diffusion-based in-betweening baseline that MCG-IMM improves upon in the comparisons.","marker":"[36]"},{"why":"The weakly-supervised action transition learning method (WAT) serves as a VAE/Transformer-based baseline and a comparison point.","marker":"[18]"},{"why":"RMI is a classical robust motion in-betweening baseline evaluated against the proposed method.","marker":"[17]"},{"why":"NSGA-II provides the fast nondominated sorting and crowding-distance selection used in the evolutionary optimization loop.","marker":"[52]"},{"why":"DDIM is cited as a deterministic sampler whose diversity suppression motivates the need for sampling-time guidance.","marker":"[21]"},{"why":"Action2Motion is a dataset and action-conditioned generation baseline used in the evaluation.","marker":"[24]"}],"fun_headline_variants":["Plug-and-play diversity boost for motion in-betweening","No-retrain optimizer diversifies in-betweening across models","Pareto-guided sampling adds variety to motion interpolation","Steer frozen motion generators to diverse, smooth transitions","Two-criteria search yields diverse human motion sequences"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pretrained generative model can be conditioned on a full previously generated motion sequence in addition to the two keyframe sequences, and that the frozen classifier's category probabilities provide a meaningful signal for within-category diversity; the paper does not spell out how VAE or GAN backbones would accept that sequence conditioning.","fun_headline_variants_meta":{"raw":{"variants":["Plug-and-play diversity boost for motion in-betweening","No-retrain optimizer diversifies in-betweening across models","Pareto-guided sampling adds variety to motion interpolation","Steer frozen motion generators to diverse, smooth transitions","Two-criteria search yields diverse human motion sequences"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1566,"prompt_tokens":1024,"completion_tokens":542,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":466}},"tokens_in":640,"tokens_out":542,"duration_ms":7052,"temperature":1.0,"reasoning_tokens":466,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:30:05.074995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply MCG-IMM to a pretrained VAE or GAN backbone whose forward pass accepts only keyframe poses and has no conditioning pathway for a full previously generated sequence; if that backbone cannot produce the offspring described in the generation equation, the method's model-agnostic plug-and-play claim fails in practice.","supporting_citations":[{"cited_title":"Human motion diffusion model","cited_arxiv_id":null,"evidence_quote":"Provides the conditional human motion diffusion prior used to build the DDPM in-betweening backbone and documents the diversity shrinkage in conditional settings."},{"cited_title":"Mofusion: A framework for denoising-diffusion- based motion synthesis","cited_arxiv_id":null,"evidence_quote":"MoFusion is the diffusion-based in-betweening baseline that MCG-IMM improves upon in the comparisons."},{"cited_title":"Weakly-supervised action transition learning for stochastic human motion prediction","cited_arxiv_id":null,"evidence_quote":"The weakly-supervised action transition learning method (WAT) serves as a VAE/Transformer-based baseline and a comparison point."},{"cited_title":"Robust motion in-betweening.ACM Transactions on Graphics (TOG), 39(4):60–1, 2020","cited_arxiv_id":null,"evidence_quote":"RMI is a classical robust motion in-betweening baseline evaluated against the proposed method."},{"cited_title":"Denoising diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"DDIM is cited as a deterministic sampler whose diversity suppression motivates the need for sampling-time guidance."},{"cited_title":"Action2motion: Conditioned generation of 3d human motions","cited_arxiv_id":null,"evidence_quote":"Action2Motion is a dataset and action-conditioned generation baseline used in the evaluation."}],"review_version":1}