{"id":"5445d13a-234e-4f23-bdc0-63849d85c54f","arxiv_id":"2412.05460","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CigTime generates corrective text instructions from a source-target pair of 3D human motions by training an LLM on triplets synthesized with a text-conditioned motion editor.","lead":"CigTime takes a person's current motion and a desired motion, then produces a text instruction that tells the person what to change. It is one of the first systems to treat corrective instruction generation as the inverse of text-driven motion editing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim's evidence is weakened by synthetic corrective labels and a partly circular editor loop; the paper's Table 3 softens but does not remove the concern, so a human-validated independent benchmark is required before the large-margin claim can be accepted.","rationale":"The reader's weakest assumption correctly identifies the synthetic triplets and the MDM scoring loop as the load-bearing premise. I agree that the paper's headline numbers in Table 1 are largely consistency checks with a model-generated world: target motions are produced by MDM under a random caption, and the reconstruction metric asks whether the predicted instruction can be executed by that same MDM to recover the target. This is circular in the sense that both the data-generation oracle and the evaluation oracle share the same text-to-motion mapping. However, I flag Table 3 as partial mitigation: when PriorMDM is used as the editor for evaluation, CigTime still reconstructs the MDM-defined target almost as well as the ground-truth caption (0.22 MPJPE vs 0.22 for PriorMDM-LW; 0.26 vs 0.25 for PriorMDM-RF). So the claim is not entirely a tautology of using the same editor; the generated instructions carry enough information to be executed by a different editor. The more serious residual flaw is label validity: the 'ground-truth' instruction is not an expert correction but a randomly selected caption, and BLEU/ROUGE/METEOR are computed against that label. The paper itself shows in Fig. 5 that multiple, semantically distinct instructions lead to near-identical reconstructions, which means n-gram overlap with a single random caption is not a faithful measure of instruction quality. The reconstruction metric, even with an independent editor, still measures agreement with the synthetic target definition, not real coaching effectiveness. No human-rated evaluation is reported; the §E real-world demo is a single anecdote with two participants and no quantitative outcome. For these reasons the paper's strong 'largely outperforms baselines' claim should remain conditional rather than being accepted as solid evidence. The proposed human rating experiment directly tests the coaching claim; a target-only ablation would further determine whether the model uses the source at all. I therefore keep the reader's conditional verdict unchanged and mark agreement as partial because I see Table 3 as mitigating part of the original circularity concern while sharpening the label-validity concern.","tokens_in":14566,"tokens_out":8027,"duration_ms":87812,"concrete_test":"Recruit at least 5 raters with movement coaching or physiotherapy background; collect 50 held-out source-target motion pairs from a real capture or pose-estimation setting (e.g., WHAM-extracted pairs like the paper's §E) that were not produced by MDM. For each pair, show raters the CigTime instruction, the Llama-3-8B in-context instruction, and the ground-truth HumanML3D caption in randomized blind order, and ask them to rate on a 1-5 Likert scale (a) whether the instruction would move a trainee's motion toward the target and (b) whether it is specific to the source-target discrepancy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that CigTime generates corrective instructions and outperforms baselines by a large margin rests on the validity of the synthetic triplets built in §3.2. In Eq. 5, each target xO is a mask-blend of the source xI with MDM's generation under a randomly drawn HumanML3D caption L; L is not an expert correction tied to a mistake in xI, and the conclusion explicitly concedes the dataset 'lacks targeted feedback on form and dynamics that are specific to actions and sports.' Because the same MDM is the editor that synthesized the targets and the scoring oracle in §4.2, the reconstruction metrics measure whether the generated text is executable by MDM to reach a self-consistently defined target, not whether it coaches a human. Table 3 does evaluate with PriorMDM, and CigTime remains close to ground truth there, which partially mitigates the 'same editor' circularity. However, the instruction-quality scores are still computed against the randomly selected caption, and the paper's own Fig. 5 shows semantically different instructions can produce nearly identical reconstructions, so BLEU/ROUGE/METEOR gains are not evidence of corrective quality. The learned model may therefore be captioning the target motion rather than comparing source and target; the source-input ablation needed to rule out that shortcut is absent. Without a human-rated benchmark on real source-target pairs, the central coaching claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CigTime, a method for generating corrective text instructions from a pair of source and target 3D human motions. The authors propose a data-collection pipeline that uses a pretrained motion diffusion model (MDM) as an editor: a randomly chosen HumanML3D action caption is used with MDM to edit the source motion, and the target is formed by mask-blending this edited motion with the source. The resulting (source, target, instruction) triplets are tokenized with a VQ-VAE and used to fine-tune Llama-3-8B. The model is evaluated with text-similarity metrics (BLEU, ROUGE, METEOR, CLIPScore) and with reconstruction accuracy, where the generated instruction is fed back into a motion editor and compared with the target via MPJPE and FID. Experiments on HumanML3D, KIT, and Fit3D, plus ablations over backbone and motion editor, are reported. The central claim is that CigTime substantially outperforms LLM and motion-language baselines on this new task.","tokens_in":14870,"tokens_out":3475,"duration_ms":34659,"significance":"If the claims are substantiated, the paper addresses a genuinely useful and understudied task: generating corrective coaching feedback from motion pairs. The idea of inverting motion editing to create training data is creative and could reduce annotation cost. The paper includes a real-world demonstration with WHAM-estimated motions, which is a good step toward practical coaching applications. It also attempts to probe generalization across datasets and motion editors. However, the current evidence is weakened by a partly circular evaluation loop, the use of non-expert synthetic corrective labels, and the absence of statistical uncertainty estimates. The ablation with PriorMDM (Table 3) is a constructive attempt to address editor circularity, and the extension to KIT shows some generalization, but the headline claim of large improvement is not yet convincingly established.","major_comments":[{"comment":"The evaluation loop is closed: the target motion xO is constructed by MDM from a randomly chosen instruction L via Eq. (5), and the reconstruction metrics in §4.2 feed the generated instruction back into MDM and compare the result with xO. This measures whether the generated text is executable by MDM to reproduce a self-consistently defined target, not whether it is a good corrective instruction for a user. Table 3 partially mitigates the same-editor concern by using PriorMDM, but the instruction-quality metrics still compare against the randomly selected caption, and Figure 5 shows that semantically different instructions can produce nearly identical reconstructions. To support the central claim, an evaluation with a motion editor not used during data generation, and ideally human ratings on real source-target pairs, is needed.","section":"§3.2, Eq. (5) and §4.2"},{"comment":"The corrective instruction label L is not an expert correction tied to a specific deficiency in xI; it is a randomly sampled HumanML3D action caption used to edit xI into xO. The paper's own limitation section admits that the dataset \"lacks targeted feedback on form and dynamics that are specific to actions and sports.\" This raises the risk that the model learns to caption the target motion given the source, rather than to compare the two motions and identify what needs to change. An ablation that removes the source motion tokens, or swaps the source and target, would directly test whether source information is actually used; no such ablation is reported.","section":"§3.2 and §5 (limitation 1)"},{"comment":"No error bars, standard deviations, or significance tests are reported, so the \"large margin\" claim in Table 1 is not shown to be statistically robust. On KIT (Table 5), the margins shrink substantially; e.g., BLEU is 0.14 for CigTime and 0.14 for Qwen-1.5-7B-LoRA, and ROUGE is 0.27 vs 0.25. The magnitude of the improvement is highly dataset-dependent, and the reader cannot tell whether the HumanML3D gains are within run-to-run variance. Reporting multiple seeds with standard deviations and appropriate significance tests is necessary to support the headline conclusion.","section":"Tables 1 and 5"},{"comment":"The paper's own qualitative examples show that different corrective instructions (e.g., \"forehand table tennis\" vs \"throwing a frisbee\") can lead to motions that are nearly identical to each other and to the target. This means the reconstruction metrics (MPJPE/FID) are insensitive to important semantic differences between instructions, so low MPJPE cannot be taken as evidence that the generated instruction is the correct one. The paper acknowledges this diversity but does not adjust its evaluation conclusions; a semantic or human-judgment evaluation is needed to establish instruction quality beyond n-gram overlap.","section":"Appendix B.3, Figure 5"}],"minor_comments":[{"comment":"There are several typographical inconsistencies: \"METERO\" appears in Tables 1 and 2 (should be \"METEOR\"), \"Geasture\" appears in Figure 5, and \"kerne_size\" appears in Table 7. The heading \"Ablation study with different network structurer\" contains a typo.","section":"Global"},{"comment":"The paper alternates between \"corrective\" and \"correctional\" instruction; please choose one consistent term throughout.","section":"Abstract/introduction"},{"comment":"The description \"training for 50000 epochs\" for MDM is likely a typo; please clarify the actual training schedule.","section":"Appendix C.1"},{"comment":"The text refers to \"the global response pdf\" in the figure caption; this is an artifact and should be replaced with a proper reference to the figure.","section":"Appendix E, Figure 6"},{"comment":"Reference [30] is listed as \"Meta. Llama3, 2024\" without a version or URL; please provide a complete citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper tackles an interesting and practical problem, and the data-generation pipeline is a reasonable starting point. However, the evaluation is not yet convincing because the main metrics are computed in a loop with the same motion editor used to create the data, and the labels themselves are not corrective in an expert sense. The missing source-input ablation and lack of error bars are fixable and would substantially strengthen the paper. I would not reject the paper, but the claims need to be tempered or the evaluation needs to be made independent before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper introduces a genuinely new task—corrective instruction generation from a general source-target motion pair—and builds a clean pipeline around it. The method is sensible: use a motion editor to synthesize triplets, tokenize motions with VQ-VAE, fine-tune an LLM to invert the editing. They compare against a reasonable set of LLM and motion-language baselines and win by wide margins. Credit where due: the inverse framing is not in the cited prior work, which is limited to static poses or specific exercises, and the paper includes a real-world demo and a generalization check on Fit3D.\n\nThe soft spots are real. The training triplets are generated by MDM from a randomly selected HumanML3D caption; that caption is not an expert correction tied to a mistake in the source. The main reconstruction metric then feeds the model's predicted instruction back into the same MDM to see if it reproduces the MDM-edited target. That is a closed loop, and it inflates the reported reconstruction gains. Table 3 helps: evaluating with PriorMDM, a different editor, still shows CigTime near ground truth, which says the generated instructions carry some transferable signal. But the instruction-quality metrics (BLEU, ROUGE, METEOR, CLIP) are computed against the random caption, not against any notion of corrective quality, and Fig. 5 demonstrates that semantically different instructions can reconstruct nearly the same motion. So the large-margin headline over baselines is not yet evidence that the model coaches a human. The paper itself concedes the dataset lacks action-specific form feedback.\n\nThe missing experiment that would sharpen the verdict is a source-input ablation: if you pair a target with a random source, does the model still produce the same caption? If yes, it is captioning the target and not comparing the pair. I'd want to see that, plus a human-rated benchmark on real source-target pairs, before trusting the coaching claim. The paper deserves a serious referee—it is not desk-reject material—but it needs revision, not acceptance as is.","headline":"New task formulation with a plausible method, but the synthetic data loop and self-scoring editor make the headline gains provisional until human-validated evaluation.","tokens_in":15383,"tokens_out":3269,"would_cite":true,"duration_ms":29685,"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":"CigTime turns a source-target motion pair into the corrective instruction that transforms one into the other.","keywords":["corrective instruction generation","motion editing","text-conditioned motion generation","motion diffusion","large language models","VQ-VAE tokenization","motion-language model","sports coaching"],"falsifier":"Take a set of source-target motion pairs from real coaching sessions and ask coaches whether the generated instructions describe the intended correction; if instructions that reconstruct well under the motion editor are frequently judged as describing a different action, the core claim fails. Concretely, a pair whose true correction is 'bend your knees on landing' that yields an instruction the editor reconstructs well but a coach reads as a different move would contradict the inverse-motion-editing assumption.","tokens_in":14358,"feed_emoji":"🎯","tokens_out":6541,"duration_ms":67835,"temperature":0.7,"pith_summary":"This paper introduces CigTime, a motion-language model that takes two motion sequences, the user's current movement and a target movement, and writes the corrective instruction that would turn the first into the second. The authors' central claim is that this task is the inverse of text-conditioned motion editing, so they can generate unlimited training triplets by asking a pretrained motion editor to edit source motions according to randomly selected captions. A VQ-VAE tokenizer converts both motions into discrete tokens, and a large language model is fine-tuned on the template pairing source tokens, target tokens, and the instruction. On a motion-caption benchmark, CigTime beats large-language-model and motion-language baselines on both text metrics and reconstruction accuracy. If the claim holds, coaching feedback can be synthesized without expert annotation, which matters for sports training and motor rehabilitation.","feed_headline":"A model writes coaching cues from a motion pair","feed_subtitle":"Training on synthetic edits from a diffusion motion editor lets a fine-tuned LLM produce instructions that reconstruct the target motion.","key_machinery":"The central machinery is the inverse of the paper's editing equation: the target motion is defined as $x_O = m \\odot x_L + (1-m) \\odot x_I$, where $x_L$ is the motion generated by a pretrained text-conditioned diffusion editor from an instruction $L$ and $m$ is a mask selecting the edited body part. CigTime is trained to invert this map: a VQ-VAE tokenizer turns $x_I$ and $x_O$ into discrete token lists, the token lists are arranged into a prompt, and a large language model is fine-tuned with cross-entropy loss to emit $L$. An anchor loss $\\lambda\\|W-W_0\\|_2^2$ keeps the newly learned motion-token embeddings from drifting away from the original vocabulary while still allowing them to adapt.","core_discovery":"On its own terms, the paper claims that corrective instruction generation is the inverse problem of text-conditioned motion editing, and that inverting an existing editor with synthetic triplets is enough to train a model to do it. Given any source motion and desired target motion, the trained model outputs a textual correction; when that text is fed back through the motion editor, it reproduces the target better than text from generic large language models or motion-captioning models. The paper reports a mean-per-joint position error of 0.13 and an FID of 1.44 on its benchmark, against best baseline values of 0.21 and 2.09, and reports that its advantage persists when the evaluation uses a different motion editor than the one that generated the training data.","pith_inferences":["Because the same pretrained motion editor is used to create the training triplets and to score reconstruction, the reported MPJPE and FID numbers partly measure how faithfully the model imitates that editor's caption-to-edit mapping, not how good the coaching is in absolute terms.","Generalization to a different motion-caption collection shows n-gram scores dropping sharply while reconstruction metrics stay similar; a plausible reading is that the model learns transferable motion-difference semantics but not dataset-specific vocabulary, and human evaluation would be needed to decide which property matters more.","A natural next test is to gather a small set of expert-written correction pairs: if the model can be fine-tuned on them or evaluated with human raters, one can separate the synthetic-editor bias from genuine coaching quality."],"forward_implications":["Corrective instruction data no longer has to be manually annotated: any source motion from a motion-caption dataset can be turned into training triplets through editing.","The generated text is actionable in the sense that feeding it back into the motion editor moves the source toward the target, so the same evaluation loop can be automated without human judges.","The method generalizes beyond the training editor: evaluated with a different motion-diffusion editor, the generated instructions still reconstruct targets better than baselines, suggesting the model captures motion-difference semantics rather than memorizing the data editor's behavior.","The trained generator can be plugged into pose-estimation-based coaching systems, as demonstrated in the paper's real-world example where motions extracted from monocular video are turned into corrective text."],"supporting_citations":[{"why":"Supplies the pretrained text-conditioned motion editor used to synthesize target motions from instructions and later to score reconstruction quality.","marker":"[40]"},{"why":"Supplies the source motion sequences and paired captions used to build the triplets, plus the feature extractor used for FID evaluation.","marker":"[13]"},{"why":"The decoder-only large language model that is fine-tuned to emit the corrective instruction.","marker":"[30]"},{"why":"The VQ-VAE scheme that quantizes motion frames into discrete tokens for the language model prompt.","marker":"[41]"},{"why":"Contributes the approach of treating motion as tokens in large language models and serves as a motion-language baseline in the comparison.","marker":"[20]"},{"why":"Provides an alternative motion-diffusion editor used to evaluate whether generated instructions reconstruct targets under editors other than the data editor.","marker":"[37]"}],"fun_headline_variants":["Inverse motion editing yields corrective coaching cues","Coaching text by inverting a motion editor's guidance","A model writes instructions to close the motion gap","From motion pair to targeted coaching feedback","Inverting motion editing to generate better coaching cues"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that synthetic triplets made by a pretrained motion editor, with a randomly chosen caption standing in for expert feedback, are valid corrective instruction data, and that the same editor is a faithful oracle for scoring them.","fun_headline_variants_meta":{"raw":{"variants":["Inverse motion editing yields corrective coaching cues","Coaching text by inverting a motion editor's guidance","A model writes instructions to close the motion gap","From motion pair to targeted coaching feedback","Inverting motion editing to generate better coaching cues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1381,"prompt_tokens":847,"completion_tokens":534,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":464}},"tokens_in":463,"tokens_out":534,"duration_ms":6423,"temperature":1.0,"reasoning_tokens":464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:42:33.260136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of source-target motion pairs from real coaching sessions and ask coaches whether the generated instructions describe the intended correction; if instructions that reconstruct well under the motion editor are frequently judged as describing a different action, the core claim fails. Concretely, a pair whose true correction is 'bend your knees on landing' that yields an instruction the editor reconstructs well but a coach reads as a different move would contradict the inverse-motion-editing assumption.","supporting_citations":[],"review_version":1}