{"id":"790ea71a-2df1-4477-b503-2c432d462453","arxiv_id":"2507.09704","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural system predicts key-joint intentions from motion history and uses adversarially regularized codebook matching to synthesize controllable, real-time reactive motions.","lead":"The authors built a system that generates one character's reactive movements in real time from another character's past motion, predicting the partner's likely future motion to make reactions stable. It matters because it turns offline reaction animation into an interactive tool a user can steer with keyboard input.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority over matching-based methods is not established: Table 2 shows Motion Matching beats the proposed method on every reported metric, and the user-control regime where the method is claimed to win has no quantitative evaluation.","rationale":"A good-faith reading of the paper shows a coherent architecture, and the ablations do demonstrate that the intention predictor and adversarial codebook matching improve over the no-intention and no-GAN variants. The method is plausible for its intended use. However, the headline claim requires two things: quantitative superiority over matching-based baselines, and stability/generalizability under user control. The paper's own Table 2 undermines the first in the only ground-truth setting: Motion Matching is better on every reported metric, which the text itself concedes. For the second, the only evidence is qualitative. The intention predictor is trained on recorded future root trajectories, which are part of the input representation in Section 3.1 (T_X^{-1→+1}); user control replaces that future root trajectory with novel paths, so the predictor operates out-of-distribution. The paper never reports intention error on perturbed inputs, and there is no quantitative comparison of reaction quality under control. These are not disagreements with outside consensus; they are missing measurements for the paper's specific claim. The proposed test, re-running Table 2 metrics on synthetically perturbed future root trajectories with Motion Matching as a baseline and comparing degradation against the intention-predictor ablation gap, would settle whether the claimed advantage survives in the control regime. The reader's weakest assumption overlaps with intention-predictor reliability but frames it as prediction from history; the paper actually feeds a user-specified future root trajectory into the predictor, so the more direct issue is the absence of any quantitative evidence that this out-of-distribution mapping works. The manuscript is plausible and correctable, so I do not move the verdict; it remains conditional pending these measurements.","tokens_in":12656,"tokens_out":7341,"duration_ms":85393,"concrete_test":"Take the held-out test set and synthesize user-control perturbations by replacing the input character's future root trajectory T_X^{0→+1} with rotated and shifted paths (e.g., ±15°, ±30°, and 0.5-1.5 m lateral offsets). Run the full trained pipeline and Motion Matching on these inputs and report the Table 2 metrics plus the intention-prediction MSE from Eq. 5. If Ours does not beat MM on MPJPE and FID under these controls, or if the controlled-vs-noncontrolled degradation in MPJPE for Ours exceeds the Complete-vs-w/o-intention ablation gap (0.88 MPJPE), then the claimed superiority and generalizability under user control is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the abstract's claim that the method 'outperforms other matching-based motion synthesis approaches' and delivers 'superior stability and generalizability,' the paper must show that the intention-guided codebook matching beats a strong matching baseline in the target setting. Table 2 (Section 4.3) is the only quantitative comparison with ground truth, and there Motion Matching dominates on every metric: MPJPE 1.88 vs 6.91, FID 0.47 vs 28.36, Diversity 81.65 vs 61.77 (GT 85.35), and codebook error 0.011 vs 0.048. The paper argues that MM lacks generalizability and is memory-intensive, but the 'superior stability and generalizability' claim is then supported only by qualitative figures and video for user-controlled inputs. In that regime, the key component is the intention predictor (Eq. 4-5), which must convert a user-modified future root trajectory into accurate five-key-joint intentions; its inference-time error is never measured, no quantitative metrics are reported for controlled trajectories, and no error bars or code are provided. If intention prediction degrades under user control, the encoder is conditioned on misleading intentions, and the claimed advantage over matching is unverified. Thus the strongest central claim is not supported by the evidence as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a real-time reactive motion synthesis method that, given the history of two interacting characters and an optional user-controlled future trajectory for the input character, predicts the next pose of the reactive character. The main ideas are an intention predictor that forecasts future trajectories of five key joints of the input character, and an adversarial codebook matching module that encodes input history plus predicted intention into a latent space, matches it to a codebook, and decodes the next pose. The method is intended to run online at about 7 ms per frame, with recursive feedback for long-term synthesis. Quantitative experiments are reported on a combination of ReMoCap, boxing, and DD100 datasets, comparing against Motion Matching and the codebook matching method of SSH*24, plus ablations that remove the intention predictor, the adversarial loss, or train the intention predictor end-to-end. User-controlled trajectory experiments are presented qualitatively.","tokens_in":12941,"tokens_out":5170,"duration_ms":55812,"significance":"If the claims were fully supported, the work would be a useful contribution to interactive character animation: it addresses a genuinely online setting where the reactive character cannot see the future of the input character, and it demonstrates a plausible mechanism for conditioning generation on a predicted intention while retaining real-time performance. The adaptation of categorical codebook matching to two-character interaction, the use of a relative coordinate frame, and the scheduled-sampling study are positive elements. However, the paper's central claim of superiority over matching-based approaches is not established by the reported evidence, because the strongest matching baseline dominates the proposed method on all offline metrics, and the user-controlled regime where the method is claimed to win is only evaluated qualitatively. The work is therefore promising but needs a substantial strengthening of the evaluation before the main conclusions can be accepted.","major_comments":[{"comment":"The central claim that the method 'outperforms other matching-based motion synthesis approaches' is directly contradicted by the quantitative comparison with Motion Matching (MM): MM achieves lower MPJPE (1.88 vs 6.91), lower FID (0.47 vs 28.36), higher Diversity closer to ground truth (81.65 vs 61.77 against GT 85.35), and lower codebook error (0.011 vs 0.048). Since MM is the matching-based baseline used in the paper, the evidence as presented supports the opposite of the claimed superiority in the offline accuracy setting. The argument that MM lacks generalizability is deferred to qualitative controlled experiments, so the claim needs to be restricted or replaced by quantitative evidence in the setting where the method is claimed to win.","section":"Section 4.3, Table 2"},{"comment":"The user-controlled regime, which is the main claimed advantage over Motion Matching, is evaluated only qualitatively. No quantitative metrics are reported for the proposed method or for MM and [SSH*24] under modified root trajectories, and the paper does not report error bars, perceptual studies, or any measure of physical plausibility such as foot sliding or codebook error in the controlled setting. Figure 13 shows a density plot, but it lacks axis labels and numeric comparisons. Without such evidence, the conclusion that the method 'outperforms' MM under user control is unsupported.","section":"Section 4.3.2, Figures 7 and 9"},{"comment":"The intention predictor is trained with MSE against ground-truth future five-key-joint trajectories, but at inference the user modifies the future root trajectory T^{0→+1}_{X_t}. The paper never measures the intention prediction error under user-controlled inputs or how this error degrades as the controlled trajectory diverges from training data. Since the codebook encoders are conditioned on the intention feature, an inaccurate intention prediction could reduce the method to history-only codebook matching, which the ablation shows performs worse. Reporting inference-time intention error for controlled trajectories is necessary to validate the key mechanism.","section":"Section 3.2, Eqs. (4)-(5), Section 4.3.2"},{"comment":"The improvements attributed to the adversarial loss and to the intention predictor are small (e.g., MPJPE 6.91 vs 7.13 vs 7.79; codebook error 0.048 vs 0.054 vs 0.063) and are reported without variance across seeds or statistical significance. The paper also does not report the sensitivity of these results to the codebook capacity and loss weights, even though lambda_G/lambda_C is shown to have a sharp optimum at 500 in Figure 14. The evidence for the specific contribution of adversarial training is therefore suggestive rather than conclusive.","section":"Section 4.4, Table 3"},{"comment":"The claim that 'Motion Matching exhibits significantly poorer codebook alignment' under user control conflicts with Table 2, where MM has the lowest codebook error (0.011 vs 0.048) in the standard evaluation. If Figure 13 is restricted to the controlled setting, this needs to be stated explicitly, and the axes and quantitative values must be provided; as written, the figure cannot support the comparison.","section":"Section 4.4.2, Figure 13"}],"minor_comments":[{"comment":"The reference 'Figure ??' should be a numbered figure reference, and Table 1 is discussed before being introduced.","section":"Section 4.1"},{"comment":"The text says the data representation is 'as stated in Section 5' and the five key joints are 'as described in Section 5', but both are described in Section 3.1.","section":"Section 3.2"},{"comment":"The scheduled sampling probability p=e/E increases with epoch, so early training uses ground truth and later training uses predictions; the direction of the schedule is unusual and should be stated explicitly, since standard scheduled sampling typically decreases teacher forcing over time.","section":"Section 4.4.1, Eq. (15)"},{"comment":"The table header contains the typo 'Diversiy' instead of 'Diversity'.","section":"Table 2"},{"comment":"The paper says 'throughout' where 'thorough' is intended, and in Section 4.3.1 'lays' should be 'lies'.","section":"Section 4.3"},{"comment":"The limitation about periodic motions is useful, but it also implies the method's performance depends heavily on dataset regularity; the paper should connect this limitation to the claimed generalizability under user control.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant and timely problem, and the proposed pipeline is reasonable. The main obstacle is that the paper's headline claim is contradicted by its own Table 2 against Motion Matching, while the regime in which the claim might hold is only qualitatively demonstrated. I would be willing to look at a revised version that either softens the claim or, preferably, adds quantitative controlled-trajectory evaluation, intention prediction error under control, and variance/statistical significance for the ablation results. The absence of released code or a supplemental video link also weakens reproducibility, though this is not by itself a blocker."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know before reading. First, the paper's central claim—that intention-guided codebook matching beats other matching-based approaches—is not actually supported by its own table. Second, the underlying method is real and the ablations are useful, so it's a revision problem, not a reject.\n\nThe genuinely new piece is the assembly: an intention predictor that forecasts the partner's five key joint trajectories, feeding those predictions into a codebook-matching controller adapted from SSH24, with adversarial regularization on the latent space. The adaptation to two-character interaction in the reactor's root space is sensible. The ablations are clean: adding the intention predictor drops MPJPE from 7.79 to 6.91 and FID from 43.77 to 28.36; the GAN loss reduces codebook error from 0.054 to 0.048. Seven milliseconds per frame is a genuine real-time number, and the limitation paragraph on non-periodic motions is honest.\n\nThe soft spot is the matching-baseline comparison. In Table 2, Motion Matching is evaluated with query features that include the future root trajectories of both characters—T^{0→+1}_Xt and T^{0→+1}_Yt. That means MM is given ground-truth future information that the proposed method is explicitly denied. It then beats the proposed method on every metric: MPJPE 1.88 vs 6.91, FID 0.47 vs 28.36. Calling that an 'inferior' matching-based approach while claiming to 'outperform' it is a misread of the table. The right comparison would be MM without future features, or a quantitative user-control experiment. The user-control evidence is purely qualitative, and intention-prediction error under steering is never measured. No error bars are reported, no code is released, and the GAN/matching loss weight is tuned on the evaluation metric.\n\nNone of that kills the method. The idea of predicting intentions to stabilize online codebook matching is legitimate, and the ablations suggest it works. But the strongest written claims outrun the evidence. For a reader working on interactive motion synthesis, the paper is worth reading for the architecture and the ablation results; the matching comparison should be ignored until fixed.\n\nMy recommendation: send it to peer review. The editor should not desk-reject, but a referee should require the authors to either rerun the matching baseline with only historical inputs or provide quantitative results for the controlled trajectory experiments, and to soften the abstract's claims accordingly.","headline":"Plausible online reactive-motion method with clean ablations, but the headline claim against Motion Matching is contradicted by the paper's own table.","tokens_in":13462,"tokens_out":4453,"would_cite":false,"duration_ms":43103,"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":"This paper claims that a character's reactive motion can be synthesized online at about 7 ms per frame by first predicting the input character's key-joint intentions and then matching those intentions to an adversarially trained codebook…","keywords":["reactive motion synthesis","intention prediction","codebook matching","adversarial training","real-time character control","human interaction animation","motion capture","user-controlled motion"],"falsifier":"Measure the intention predictor's error, for example MPJPE between predicted and actual future trajectories of the five key joints, at inference time on user-controlled test trajectories: if intention error grows large precisely when output quality degrades or collapses to the noisy history-only baseline, the paper's central explanation is falsified. A complementary check is to run the system on non-periodic inputs such as a standing character clapping, where the paper itself predicts trouble, and see whether the reactive character freezes or jitters beyond acceptable quality.","tokens_in":12449,"feed_emoji":"🕺","tokens_out":6504,"duration_ms":62100,"temperature":0.7,"pith_summary":"This paper tries to establish that a single character's reactive motion can be synthesized online, frame by frame, from only the observed history of both characters plus user-specified control, without knowing the input character's future motion. The key move is to insert an intention predictor that forecasts the future trajectories of five key joints of the input character, turning an ill-posed prediction problem into a better-conditioned one. That predicted intention is encoded into the latent space of the reactive motion and matched against a learned codebook, with adversarial training aligning the estimated and target latent distributions. If the claims hold, interactive animation systems can produce long, realistic, user-steerable interactions in real time instead of relying on precomputed full trajectories or large motion databases. The paper's ablations support the claim that the intention predictor, not just the codebook, is what stabilizes the output.","feed_headline":"Reactive partners animate in 7 ms per frame via predicted intent","feed_subtitle":"The system predicts a partner's key-joint future from history, then codebook-matches a stable, steerable reaction.","key_machinery":"The central object is the intention predictor combined with an adversarial codebook matching module. The intention predictor is a small fully-connected network that takes the current input features of both characters and outputs the predicted future trajectories of five key joints of the input character; these trajectories are represented in the local root space of the reactive character so that the model sees relative interaction rather than absolute positions. The codebook is a learned set of $c=16$ vectors of dimension $128$; the input encoder and output encoder each produce a categorical distribution over these vectors via Gumbel-Softmax, the quantized vector is decoded into the next pose, and a discriminator forces the estimated latent distribution to align with the target distribution before quantization. That adversarial alignment is what the paper claims prevents distinct interaction patterns from collapsing onto the same codebook entry, which they identify as the failure mode of plain codebook matching on two-character interaction.","core_discovery":"On the paper's own terms, the central discovery is that reactive motion synthesis, previously treated as an offline problem in which the input character's full future is known, can be made online and controllable by predicting a compact 'intention'—the future trajectories of the head, hands, and legs of the input character in the reactor's local space—and using that prediction as an auxiliary conditioning signal for adversarial codebook matching. With this intention signal, the same historical input no longer maps to ambiguous codebook vectors, so the output poses stay coherent over long recursive generation. The method predicts the next pose stochastically by sampling a categorical distribution through Gumbel-Softmax, then decodes the sampled codebook vector into the pose, and a discriminator aligns the continuous latent space so that similar interactions map to similar regions. The authors report that this beats motion matching and plain codebook matching on MPJPE, FID, diversity, and codebook matching error, and that it keeps producing interactive motion even when user control pushes the input trajectories far from anything in the training set.","pith_inferences":["Beyond the paper: the method's success likely hinges on how predictable the input motion is; the paper's own limitation section notes that non-periodic motions like clapping cause jitter or frozen output, so an immediate testable extension is to measure intention prediction error on periodic versus non-periodic inputs.","Beyond the paper: because the intention predictor is trained only on ground-truth future trajectories, online errors from feeding predicted intentions back into the encoders are never measured; an inference-time evaluation of intention error under user steering would show how much of the claimed stability survives prediction drift.","Beyond the paper: the same intention-plus-codebook recipe could transfer to other paired interaction tasks, such as two-agent navigation, handover motions, or co-speech gesture pairs, wherever a compact future signal can be estimated from history."],"forward_implications":["Real-time interaction systems such as games, VR, and social simulation can generate reactive partner motions at about 7 ms per frame with no access to the partner's future motion.","Users can steer the input character's path with keyboard controls, and the reactor continues to produce plausible, non-frozen motion even for trajectories outside the training distribution.","Long-term synthesis is feasible by recursively feeding generated poses back as history, since the intention predictor and matcher operate on one frame at a time.","Stochasticity is preserved through Gumbel-Softmax temperature, so the same input can yield multiple plausible reactions while remaining stable across samples.","The intention predictor is the load-bearing component: removing it degrades codebook matching error and FID substantially, so future methods in this line must keep or replace that conditioning signal."],"supporting_citations":[{"why":"Supplies the codebook matching framework that maps input latent features to output poses and the categorical sampling scheme that this paper adapts to two-character interaction.","marker":"[SSH∗24]"},{"why":"Provides the ReMoCap two-person dancing dataset used for training and evaluation and represents the offline reaction-synthesis baseline that motivates the online formulation.","marker":"[GDG∗24]"},{"why":"Motion matching baseline that retrieves next poses from the data; the paper compares against it as a standard real-time alternative and shows it lacks generalizability under user control.","marker":"[MZ20]"},{"why":"Learned motion matching variant used as the other real-time matching baseline, sharing the query-feature retrieval limitation.","marker":"[HKPP20]"},{"why":"Supplies the local-space pose and trajectory representation, including foot-contact features and root updates, that the paper uses for both input and output.","marker":"[SZKZ20]"},{"why":"Gumbel-Softmax reparameterization is used to sample the categorical distribution over codebook vectors so gradients flow through quantization.","marker":"[JGP17]"},{"why":"Provides the DD100 dance accompaniment dataset used as a third training source, widening the motion distribution beyond boxing and ReMoCap.","marker":"[SGY∗24]"}],"fun_headline_variants":["Intention predictor makes reactive motion real-time and user-controllable","Codebook matching with intention cues enables real-time interactive motion","Predict future joints to synthesize responsive partners in milliseconds","From offline to online: intention guidance for controllable reactive motion","Steerable avatars: real-time reaction via predicted intention and codebook"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method stands on the assumption that the future motion of the input character can be predicted well enough from history alone that the predicted intention is a reliable conditioning signal; if that prediction becomes inaccurate under user steering, the system degrades to history-only codebook matching, which the paper's own ablation shows is markedly worse.","fun_headline_variants_meta":{"raw":{"variants":["Intention predictor makes reactive motion real-time and user-controllable","Codebook matching with intention cues enables real-time interactive motion","Predict future joints to synthesize responsive partners in milliseconds","From offline to online: intention guidance for controllable reactive motion","Steerable avatars: real-time reaction via predicted intention and codebook"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1550,"prompt_tokens":919,"completion_tokens":631,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":560}},"tokens_in":535,"tokens_out":631,"duration_ms":7850,"temperature":1.0,"reasoning_tokens":560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:49:35.549942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the intention predictor's error, for example MPJPE between predicted and actual future trajectories of the five key joints, at inference time on user-controlled test trajectories: if intention error grows large precisely when output quality degrades or collapses to the noisy history-only baseline, the paper's central explanation is falsified. A complementary check is to run the system on non-periodic inputs such as a standing character clapping, where the paper itself predicts trouble, and see whether the reactive character freezes or jitters beyond acceptable quality.","supporting_citations":[],"review_version":1}