{"id":"e35b18ed-958d-4f29-8f5b-0ba648f12df2","arxiv_id":"2508.10897","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Human-in-Context is a single in-context-learning model that handles 10 pose- and mesh-based 3D human motion tasks across 4 datasets and outperforms prior unified and task-specific baselines.","lead":"This paper trains one model to handle multiple 3D human motion tasks (pose estimation, mesh recovery, motion prediction, motion in-between, and joint completion) across several datasets and two data formats, using examples, called prompts, to tell the model which task to perform. The authors report that their unified model beats prior multi-task and single-task baselines on both known datasets and an unseen dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline re-implementation is the load-bearing concern. PoseRetNet's six mesh-task entries in Tables 3–4 (312–482 mm vs 43–145 mm for all other methods) are non-functional, yet they are included in the 21.8% average; Section 5's one-sentence protocol does not establish fair baselines.","rationale":"This paper has genuine strengths: a single-stage, head-free unified model spanning pose and mesh through one token format (Section 3.1, virtual joints and beta=0) is a real design contribution; the ablations in Tables 5–9 cover anchors, soft anchors, sampling, aggregation levels, layers, and feature dimension and are internally consistent; HiC wins on 36 of 36 task-dataset entries; and the announced code release supports reproducibility. The load-bearing risk is the comparison protocol, which is exactly the reader's weakest assumption, sharpened with table-level evidence. The Experimental Setting paragraph of Section 5 is the only description of how all five baselines were adapted, and it omits architecture changes, losses, schedules, and compute. PoseRetNet†'s mesh entries at 312–482 mm MPVE — 4–11x worse than every other method on identical entries — cannot represent a competently adapted baseline, and since the text never defines the 21.8% average, the headline margin is not verifiable. The pose baselines are also far below their published native results (MotionBERT† 98.36 mm on H3.6M pose estimation vs ~64 mm published), with no sanity check that the re-implementations were faithful. Two secondary issues reinforce conditionality. First, the SPS algorithm as written (Algorithm 1, Eqs. 4–7) is mathematically identical to farthest-point sampling: argmin over sequences of max-similarity-to-anchor-set equals argmax of distance to the nearest anchor, so the sizable SPS-vs-FPS gap in Table 7 and the qualitative distinction in Figure 8 are unexplained unless the implementation differs from the equations; the prompting-novelty claim needs reconciliation. Second, '10 tasks across 4 datasets' overstates native coverage: FreeMan is evaluated on 6 tasks only (Table 3c), AMASS's 2D pose is orthographically projected from 3D (Section 3.1), and H3.6M mesh is MoSh-generated; these are disclosed but temper the scope claim. Related unified models (UPS, LMM, UniHCP) are discussed but not compared, so the 'first unified model' claim is defensible only within the paper's specific constraints. I keep the reader's CONDITIONAL verdict rather than rejecting: most per-task margins are large, the wins are consistent, and the ablations are coherent, so the model may well be as strong as reported. The condition is precise and testable: release the code, validate the baseline adaptations, and recompute the aggregate.","tokens_in":28468,"tokens_out":25880,"duration_ms":264096,"concrete_test":"Using the released code/checkpoints, recompute the aggregate improvement behind the 21.8% claim after correcting the six PoseRetNet† mesh-task entries (AMASS MR/FMR 478.51/481.95, H3.6M MR/FMR 328.24/343.60, 3DPW MR/FMR 314.26/311.91 mm; all other methods score 43–145 mm on the same entries): re-run PoseRetNet with a valid SMPL-output adaptation, or if no such adaptation exists, exclude those entries and recompute the average over the remaining baseline-task pairs. If the average drops below roughly 15%, or if MotionBERT†'s H3.6M pose-estimation result (98.36 mm) is not reproducible from the stated recipe given MotionBERT's published fine-tuned value near 64 mm, then the reported '21.8% average improvement' reflects handicapped baselines rather than model quality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim — 'outperforms prior domain-specific and cross-domain models by 21.8% on average' — is computed against five baselines re-implemented by the authors under a protocol described in one sentence in the Experimental Setting paragraph of Section 5: 'All models are re-implemented to align with the setting of unified cross-domain 3D human motion modeling, i.e., one-time training for all tasks and datasets without any domain-specific model heads.' No per-baseline architecture adaptation, loss, training schedule, or compute budget is given, and no error bars are reported. Tables 3–4 contain direct evidence that this protocol is not producing valid baseline behavior. PoseRetNet†, a pose-estimation method with no native SMPL output mechanism, reports MPVE of 478.51 and 481.95 on AMASS mesh recovery and future mesh recovery, 328.24 and 343.60 on H3.6M, and 314.26 and 311.91 on 3DPW — 4–11x worse than every other method on those same entries, which all fall in the 43–145 mm range. These are not merely weak numbers; they indicate a non-functional pose-to-mesh adaptation, and under any average over baseline-task pairs they inflate the headline margin. The text never defines the formula for the 21.8% average. The pose entries are also far below published native-task numbers: MotionBERT† scores 98.36 mm on H3.6M pose estimation versus MotionBERT's published fine-tuned value near 64 mm, with no evidence that the re-implementations received comparable capacity, tuning, or training effort to HiC's 120 epochs on four A6000 GPUs. The model may genuinely be strong — wins are consistent across 36 task-dataset entries and the ablations are coherent — but the magnitude of the central claim is not established until the baseline protocol is validated and the aggregate margins are recomputed. This confirms, rather than changes, the reader's CONDITIONAL verdict.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Human-in-Context (HiC), a single model trained once to handle ten 3D human motion tasks across pose and mesh modalities and four datasets (AMASS, Human3.6M, FreeMan, and the out-of-domain 3DPW). The method extends the authors' previous Pose-in-Context (PiC) with a max-min similarity prompt sampling strategy that retrieves hard anchors and paired soft anchors, and with X-Fusion Net, a dual-branch network using self-attention, graph convolution, and state-space aggregation followed by cross-level context update. The paper claims consistent improvements over re-implemented baselines, reporting a 9.9% gain over PiC and a 21.8% average gain over prior domain-specific and cross-domain models on MPJPE/MPVE metrics. The main evidence is in Tables 3 and 4, supported by ablations on anchors, soft anchors, sampling, fusion, dimensions, and layers. Code and models are promised at a public repository.","tokens_in":28794,"tokens_out":3654,"duration_ms":42888,"significance":"If the empirical claims hold, HiC would be a useful step toward a single fully unified 3D human motion model, and its design choices (prompt sampling with hard/soft anchors, multi-level context fusion) are reasonable and well motivated. The paper also contributes a larger in-context formulation than PiC, covering two modalities and ten tasks, and it releases code and models. The significance, however, is contingent on the credibility of the baseline comparison. The current evaluation is weakened by the fact that all baselines are re-implemented by the authors under a protocol described in one sentence, that the main comparison is against the authors' own prior work (PiC), and that some re-implemented baseline entries (PoseRetNet on mesh tasks) are two orders of magnitude worse than other methods, which suggests a non-functional adaptation rather than a fair comparison. The central quantitative claims therefore need substantial verification before the paper can be accepted.","major_comments":[{"comment":"The claimed 21.8% average improvement is not credible as reported because several baseline entries appear non-functional. PoseRetNet is listed with MPVE values of 478.51 and 481.95 on AMASS mesh recovery and future mesh recovery, 328.24 and 343.60 on Human3.6M, and 314.26 and 311.91 on 3DPW, while every other method on those same entries is in the 43-145 mm range. Including these entries in an average comparison inflates the headline margin. The paper never defines the formula for the 21.8% average (which tasks/datasets are averaged, whether the average is over all 50 baseline-task entries, and whether non-functional entries are included). The authors should report per-baseline averages, exclude or separately analyze non-functional task adaptations, and give the exact formula.","section":"Section 5, Tables 3 and 4"},{"comment":"The baseline re-implementation protocol is described only as 'all models are re-implemented to align with the setting of unified cross-domain 3D human motion modeling, i.e., one-time training for all tasks and datasets without any domain-specific model heads.' This is insufficient to establish that the comparison reflects model quality rather than the authors' implementation choices. No per-baseline architecture adaptation, loss weighting, training schedule, data augmentation, or compute budget is given. The suspicion is reinforced by MotionBERT obtaining 98.36 mm on H3.6M pose estimation, far above the published native-task performance of that method. The authors should provide per-baseline implementation details and, where possible, validate that each re-implementation reproduces its published native-task result before it is used in the unified setting.","section":"Section 5, Experimental Setting"},{"comment":"The similarity measure used for anchor sampling and prompt retrieval computes a Euclidean distance between raw sequence values, but those values are semantically heterogeneous: pose entries are 3D position coordinates in millimeters, while mesh entries are axis-angle rotation vectors in radians. Directly comparing these quantities in a single distance is not meaningfully justified, especially for the claimed cross-modal prompt retrieval where a pose query may be compared with mesh anchors. The paper should either define a common normalized or semantically aligned representation before applying Eq. (4), or demonstrate empirically that the retrieval behavior is robust to this unit mismatch.","section":"Section 4.2, Eq. (4)"},{"comment":"No error bars or multiple-seed results are reported for any experiment, despite the fact that prompt retrieval, mask generation, and training involve stochastic choices. Since the paper's central claim is a set of consistent margins (9.9% over PiC, 21.8% over the field), the absence of variance estimates makes it impossible to tell whether the reported gaps are statistically significant. Reporting at least mean and standard deviation over three seeds for the main tables and the key ablations (Tables 5-8) is necessary to support the strength of these claims.","section":"Tables 3-9"}],"minor_comments":[{"comment":"There are typos in the experimental section: 'perfromance evalaution' should be 'performance evaluation', and the Figure 6 caption contains 'Recvoery' instead of 'Recovery'.","section":"Section 5"},{"comment":"The captions describe ablations of 'multi-level context update', but the component being ablated is the multi-level context aggregation in X-Fusion blocks; the wording should be aligned with Section 4.3.2.","section":"Figure 9 and Figure 10 captions"},{"comment":"The notation [E(QD_j), U] for the query feature says concatenation is along the temporal axis, but the dimensions of E(QD_j) and the TUP U are not specified; please clarify how the temporal lengths match.","section":"Eq. (2)"},{"comment":"The description of the relative motion similarity space is intuitive but not fully specified; to make the anchor sampling reproducible, the paper should state how the canonical T-body is represented in both pose and mesh forms and how sequences of different lengths are handled.","section":"Section 4.2"},{"comment":"The row for Future Mesh Recovery lists the output as {X_mesh_{F+1:2F}, beta} while the input is 2D pose from frames 1:F, but the text describing future mesh recovery in Section 3.1 does not explicitly say that the output is future frames; please make the formulation consistent.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is built around the authors' own previous model PiC as the primary baseline, and all external baselines are re-implemented by the authors. I would advise the editor to ask for either official or independently reproduced baseline numbers, or at least a demonstration that each re-implementation matches its published native-task performance, before the claimed margins are considered trustworthy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Name],\n\nQuick take: this is a genuine extension of the authors' own PiC line, and the model itself looks coherent. But the headline number—21.8% average improvement over prior methods—is not something I would quote until the baseline protocol is fixed.\n\nWhat's new: HiC takes the in-context learning idea from PiC and pushes it across modalities. Pose and mesh are unified into one token format, the task set doubles to ten, data scale grows from 0.18M to 3.83M sequences, and the prompting strategy is upgraded from random TGP selection to a max-min similarity sampling that picks diverse anchors and retrieves the closest one per query. The X-Fusion Net with self-attention, graph conv, and SSM branches plus a learned cross-level weighting is a sensible way to mix context. The ablations are systematic and tell a consistent story: SPS beats random, FPS, and k-means; soft anchors help; the dynamic update beats static averaging. The model wins on all 36 task-dataset entries, including out-of-domain 3DPW. Internally, it is a clean, credible study.\n\nThe soft spot is the evaluation protocol, and it's not minor. All baselines are 're-implemented to align with the setting of unified cross-domain 3D human motion modeling'—that's the whole description. There's no per-baseline architecture adaptation, no compute budget, no training schedule, no error bars. And the tables show the protocol is producing garbage for at least one baseline: PoseRetNet† reports MPVE of 478 mm on AMASS mesh recovery while every other method falls in the 38–72 mm range. That's a non-functional pose-to-mesh adaptation, and those entries are still averaged into the 21.8% headline. MotionBERT† at 98 mm on H3.6M pose estimation also looks far below its published ~64 mm. Wins this consistent are unlikely to vanish entirely, but the margin is unverified. I'd also want the 21.8% average defined, and a comparison to genuinely unified models like UniHCP, LMM, or UPS rather than only task-specific ones re-implemented by the authors.\n\nBottom line: this deserves a serious referee. The architecture and prompting ideas are worth engaging with, and the claims are significant if they hold. But the revision needs to document the re-implementations, recompute aggregates without the broken entries, and ideally add variance. For anyone working on unified human motion models, this is the paper to read—with a bit of skepticism about the numbers.","headline":"A coherent extension of PiC with real architectural ideas, but the headline improvement rests on under-specified re-implementations of baselines, including at least one broken one.","tokens_in":29475,"tokens_out":3339,"would_cite":false,"duration_ms":31930,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Single model, trained once, handles ten 3D human-motion tasks across pose and mesh, beating prior domain-specific and cross-domain models by 21.8% on average.","keywords":["3D human motion modeling","in-context learning","cross-domain model","pose and mesh representation","prompt sampling","human motion prediction","unified model","out-of-domain generalization"],"falsifier":"Train each baseline twice: once under the paper's unified single-process protocol and once with its native domain-specific heads and multi-stage training, then compare on the same ten tasks and on 3DPW. If the unified-protocol baselines lose substantially to their native versions, or if HiC's margin shrinks when baselines are given their own optimizers and loss schedules, the claim that the architecture and prompting strategy cause the improvement is falsified.","tokens_in":28218,"feed_emoji":"🤸","tokens_out":8567,"duration_ms":81794,"temperature":0.7,"pith_summary":"The paper sets out to prove that a single 3D human-motion model can be trained once, with no domain-specific heads and no multi-stage pipeline, and still handle ten tasks spanning two modalities (pose and mesh) and four datasets. It positions HiC as the first fully unified cross-domain 3D human motion model, with in-context learning—showing the model one input–output example, or prompt, that tells it which task to perform—as the unification mechanism. Two design choices carry the argument: a max-min similarity prompt sampling that selects representative anchors matched to each query, and a dual-branch network, X-Fusion Net, that injects prompt context into the query path while aggregating dependencies at multiple levels. The paper reports that this single model outperforms prior domain-specific and cross-domain models by 21.8% on average (and its own pose-only predecessor PiC by 9.9%) on both the three in-domain datasets and the unseen 3DPW benchmark. If right, the practical payoff is that one deployable model could replace a zoo of task-specific and dataset-specific motion models.","feed_headline":"One model, one training run: 10 human-motion tasks in pose and mesh","feed_subtitle":"Trained once, it beats task-specific models by 21.8% on average, including on an unseen dataset.","key_machinery":"Three mechanisms carry the argument. (1) A unified cross-modal representation: pose and mesh sequences are padded with virtual joints and zero shape parameters so every task input and output lives in $\\mathbb{R}^{F\\times J\\times C}$, making all ten tasks share one interface. (2) Max-min similarity prompt sampling (SPS): each training sequence is embedded in a relative motion similarity space anchored at the canonical T-pose; anchors are chosen iteratively by grouping unsampled sequences by maximum similarity to existing anchors and then sampling the sequence with minimum similarity to its group representative, giving $K=800$ anchors that cover dense and sparse regions; at inference the query retrieves its closest hard anchor plus a learnable soft anchor. (3) X-Fusion Net: a dual query/prompt branch network whose blocks perform multi-level context aggregation—self-attention for global embedding-space dependencies, graph convolution for local graph-space dependencies, and a state-space model for local state-space dependencies—followed by a cross-level context update that compresses the levels into influence scores and combines them with a learned frame-wise and joint-wise weighted sum, with prompt context added into the query branch each layer.","core_discovery":"The central claim is that cross-domain 3D human motion modeling can be formulated as in-context learning: the task is communicated entirely by a prompt pair $[\\text{input}, \\text{target}]$ drawn from the same domain as the query, and the model performs the corresponding task on the query without any task-specific head or fine-tuning. To make this work across modalities, the paper unifies pose and mesh sequences into one tensor format: pose joints are position coordinates, SMPL mesh joints are axis-angle rotation vectors, and gaps are filled with virtual joints and zero shape parameters. The paper then claims that two additions make the unified model succeed: max-min similarity prompt sampling, which spreads a small set of hard anchors across the data distribution and retrieves the closest one for each query, and X-Fusion Net, whose dual-branch architecture aggregates context with self-attention, graph convolution, and state-space models and updates features with learned cross-level weights. The reported evidence is a consistent win on all ten tasks on AMASS, Human3.6M, and FreeMan, and on the out-of-domain 3DPW dataset, with an average improvement of 21.8% over domain-specific and cross-domain baselines and 9.9% over PiC.","pith_inferences":["An untested extension follows directly: the same unified tensor format could absorb additional modalities, such as RGB video or point clouds, by adding a modality-specific encoder in front of the shared in-context model, potentially yielding a single human-motion foundation model.","The anchor set is a non-parametric memory; replacing exhaustive similarity search with approximate nearest-neighbor retrieval could let the anchor count and training scale grow far beyond the 800 used here, which the ablations suggest would help up to a point.","The max-min sampling objective is task-agnostic and could be transferred to other sequential domains, such as robotic manipulation trajectories, animal motion, or hand motion, wherever a prompt pair can define the task.","A decisive experiment the paper does not run: test on a fifth dataset with a different body model or joint topology (for example, SMPL-X) to see whether virtual-joint padding generalizes across body definitions or degrades silently."],"forward_implications":["A single trained model can serve ten tasks across pose and mesh modalities, replacing per-task heads and multi-stage pipelines.","Out-of-domain generalization improves: training on AMASS, Human3.6M, and FreeMan transfers to unseen 3DPW better than every compared baseline on all ten tasks.","Distribution-aware prompt retrieval helps more than random prompts: SPS beats random, farthest-point, and cluster-based sampling on both in-domain and out-of-domain benchmarks.","Dynamic cross-level weighting in X-Fusion Net outperforms static averaging of aggregation levels, and all three aggregators—self-attention, graph convolution, and the state-space model—contribute to the gain.","Scaling data roughly 21x while adding mesh tasks and a fourth dataset improves rather than degrades the model's generality."],"supporting_citations":[{"why":"TCPFormer is one of the domain-specific baselines re-implemented under the unified setting that HiC must beat.","marker":"[8]"},{"why":"HoT is a baseline used in the quantitative and qualitative comparisons, including mesh recovery and motion in-between.","marker":"[9]"},{"why":"Supplies the in-context learning paradigm that the paper adapts to 3D human motion modeling.","marker":"[16]"},{"why":"MotionBERT is the primary baseline and the source of the shared encoder and loss design the paper builds on and must outperform.","marker":"[18]"},{"why":"PiC (Skeleton-in-Context) is the authors' prior pose-only in-context model that HiC extends, and the baseline for the 9.9% improvement claim.","marker":"[30]"},{"why":"SMPL parametrization defines the mesh modality, with joint rotations and shape parameters used in the unified representation.","marker":"[49]"},{"why":"Human3.6M is one of the three in-domain training and test datasets for the ten tasks.","marker":"[53]"},{"why":"3DPW is the unseen out-of-domain test set used to claim generalization.","marker":"[54]"},{"why":"AMASS provides large-scale mesh motion data and is a second in-domain dataset.","marker":"[55]"},{"why":"FreeMan supplies real-world pose and mesh data as the third in-domain dataset.","marker":"[56]"}],"fun_headline_variants":["In-context learning unifies 3D motion across tasks and datasets","One model, any task: in-context 3D human motion","Cross-domain motion via in-context learning: pose and mesh","Human-in-Context: unified 3D motion without fine-tuning","Pose and mesh in one context: single-model motion modeling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison rests on the assumption that baseline models can be fairly re-implemented under the authors' unified setting—one training run, no domain-specific heads, one shared task formulation—so the reported 21.8% gap reflects model quality rather than the authors' choices of training recipes, hyperparameters, or losses for the baselines.","fun_headline_variants_meta":{"raw":{"variants":["In-context learning unifies 3D motion across tasks and datasets","One model, any task: in-context 3D human motion","Cross-domain motion via in-context learning: pose and mesh","Human-in-Context: unified 3D motion without fine-tuning","Pose and mesh in one context: single-model motion modeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1530,"prompt_tokens":1087,"completion_tokens":443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":703,"completion_tokens_details":{"reasoning_tokens":352}},"tokens_in":703,"tokens_out":443,"duration_ms":4524,"temperature":1.0,"reasoning_tokens":352,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:28:38.530736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train each baseline twice: once under the paper's unified single-process protocol and once with its native domain-specific heads and multi-stage training, then compare on the same ten tasks and on 3DPW. If the unified-protocol baselines lose substantially to their native versions, or if HiC's margin shrinks when baselines are given their own optimizers and loss schedules, the claim that the architecture and prompting strategy cause the improvement is falsified.","supporting_citations":[{"cited_title":"Language models are few-shot learners,","cited_arxiv_id":null,"evidence_quote":"Supplies the in-context learning paradigm that the paper adapts to 3D human motion modeling."},{"cited_title":"Motionbert: a unified perspective on learning human motion representations,","cited_arxiv_id":null,"evidence_quote":"MotionBERT is the primary baseline and the source of the shared encoder and loss design the paper builds on and must outperform."},{"cited_title":"Skeleton-in-context: unified skeleton sequence modeling with in-context learning,","cited_arxiv_id":null,"evidence_quote":"PiC (Skeleton-in-Context) is the authors' prior pose-only in-context model that HiC extends, and the baseline for the 9.9% improvement claim."},{"cited_title":"SMPL: a skinned multi-person linear model,","cited_arxiv_id":null,"evidence_quote":"SMPL parametrization defines the mesh modality, with joint rotations and shape parameters used in the unified representation."},{"cited_title":"Human3. 6m: large scale datasets and predictive methods for 3d human sensing in natural environments,","cited_arxiv_id":null,"evidence_quote":"Human3.6M is one of the three in-domain training and test datasets for the ten tasks."},{"cited_title":"Recovering accurate 3d human pose in the wild using imus and a moving camera,","cited_arxiv_id":null,"evidence_quote":"3DPW is the unseen out-of-domain test set used to claim generalization."},{"cited_title":"Amass: archive of motion capture as surface shapes,","cited_arxiv_id":null,"evidence_quote":"AMASS provides large-scale mesh motion data and is a second in-domain dataset."},{"cited_title":"Freeman: towards benchmarking 3d human pose estimation under real-world conditions,","cited_arxiv_id":null,"evidence_quote":"FreeMan supplies real-world pose and mesh data as the third in-domain dataset."}],"review_version":1}