Paper deep dive
Aggregate, Don't Adapt: Subject-Level Posterior Aggregation and Transductive Calibration for Cross-Site Parkinsonian Gait Severity
Junlong Shen
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/24/2026, 4:48:35 AM
Summary
This paper describes the winning entry to the MoCha 2026 Benchmark and Challenge on Parkinsonian Gait, achieving a macro-F1 score of 0.69447. The system utilizes a frozen public motion encoder (MotionAGFormer-S) and a linear probe, with performance gains primarily driven by subject-level posterior aggregation, exact reproduction of the reference benchmark's training recipe, and label-free transductive calibration. The study highlights that fine-tuning the encoder and complex architectural changes were less effective than leveraging the subject grouping structure inherent in the input data.
Entities (8)
Relation Signals (6)
Junlong Shen → achievedscore → 0.69447
confidence 95% · The system reaches 0.69447 macro-F1 on the hidden test
MoCha 2026 → haswinner → Junlong Shen
confidence 95% · We describe the winning entry to the MoCha 2026 Benchmark... Junlong Shen University of Alberta
Subject-Level Posterior Aggregation → improves → Macro-F1
confidence 95% · averaging per-walk posteriors within the subject grouping... is worth +0.143 macro-F1
CARE-PD → usedfortraining → MoCha 2026 System
confidence 90% · Training data is the CARE-PD benchmark [2]
MotionAGFormer-S → usedin → MoCha 2026 System
confidence 90% · A MotionAGFormer-S encoder [15] pretrained on Human3.6M [9], used unmodified
Transductive Calibration → improves → Macro-F1
confidence 85% · label-free transductive calibration... account for the remainder
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We describe the winning entry to the MoCha 2026 Benchmark and Challenge on Parkinsonian Gait, which predicts MDS-UPDRS gait severity from canonicalized SMPL motion recorded at clinical sites unseen during training. The system reaches 0.6945 macro-F1 on the hidden test and ranked first of 58 entries, ahead of the runner-up at 0.5807 and the organizers' baseline at 0.4289, on a frozen public motion encoder with a single $4\times512$ linear layer. Nearly all of the margin comes from three stages usually treated as bookkeeping: reproducing the reference benchmark's exact head recipe, averaging per-walk posteriors within the subject grouping the organizers ship, and a label-free transductive calibration of the feature mean and the decision operating point. Fine-tuning the encoder lost in four distinct forms, and ten alternative encoders were worse. Every ablation number is a paid read on the hidden test, because our own leave-two-cohort-out cross-validation proved anti-correlated with the deciding score over eleven configurations. We give the negative record in full, and identify our largest gain, subject-level aggregation, as the binding ceiling on this benchmark.
Tags
Links
- Source: https://arxiv.org/abs/2608.20587v1
- Canonical: https://arxiv.org/abs/2608.20587v1
Trouble viewing inline? Open PDF directly →
Full Text
38,053 characters extracted from source content.
Expand or collapse full text
Aggregate, Don’t Adapt: Subject-Level Posterior Aggregation and Transductive Calibration for Cross-Site Parkinsonian Gait Severity Junlong Shen University of Alberta, Edmonton, Canada junlong6@ualberta.ca Abstract. We describe the winning entry to the MoCha 2026 Bench- mark and Challenge on Parkinsonian Gait, which predicts MDS-UPDRS gait severity from canonicalized SMPL motion recorded at clinical sites unseen during training. The system reaches 0.69447 macro-F1 on the hidden test and ranked first of 58 entries, ahead of the runner-up at 0.5807 and the organizers’ baseline at 0.4289, on a frozen public mo- tion encoder with a single 4× 512 linear layer. Nearly all of the margin comes from three stages usually treated as bookkeeping: reproducing the reference benchmark’s exact head recipe, averaging per-walk posteriors within the subject grouping the organizers ship, and a label-free trans- ductive calibration of the feature mean and the decision operating point. Fine-tuning the encoder lost in four distinct forms, and ten alternative encoders were worse. Every ablation number is a paid read on the hidden test, because our own leave-two-cohort-out cross-validation proved anti- correlated with the deciding score over eleven configurations. We give the negative record in full, and identify our largest gain, subject-level aggregation, as the binding ceiling on this benchmark. Keywords: Parkinson’s disease· gait analysis· cross-site generalization · transductive inference· clinical motion analysis 1 Introduction Predicting parkinsonian gait severity at a clinical site never seen in training is a cross-site generalization problem, and on this benchmark the decisive levers turn out not to live in the motion representation. The MoCha 2026 challenge [1] asks for the MDS-UPDRS gait severity class [6] of a walking sequence given as canonicalized SMPL motion [13]. Training data is the CARE-PD benchmark [2]; the hidden test comes from clinical sites absent from training. Entries are ranked by macro-F1. Submissions are code: participants upload a predict function with model weights, and the organizers run it on hidden motion files. 1 Our starting hypothesis was the usual one. We assumed the gap to a useful system was a representation gap, and that a better or better-adapted motion 1 This is the arXiv version of the challenge report submitted to the MoCha 2026 organizers for prize verification. It is not part of the workshop proceedings. arXiv:2608.20587v1 [cs.CV] 20 Aug 2026 2Junlong Shen canonicalized SMPL walk joints, side view 81-frame clips frozen MotionAGFormer-S transductive centering 4×512 linear head, 14 KB per-walk posteriors subject mean over data[sid][wid] subject-kNN pooling q-divisor operating point one label per subject subject-mean aggregation +0.143 benchmark-exact head recipe +0.057 transductive centering + operating point +0.017 subject-kNN pooling +0.0075 4 fine-tuning forms, 10 alternative encoders all lost Fig. 1: The winning margin came from reading the input format, not from the motion representation. The encoder is public and frozen; the only trained parameters are a 14 KB linear layer (top). Averaging per-walk posteriors within the subject grouping that the organizers ship inside the input is worth +0.143 macro-F1, more than every representation change we made combined (bottom); every bar is a paid read on the hidden test. That same aggregation is also this benchmark’s ceiling (Sec. 7). encoder would close it. That hypothesis failed in every form we tested. What moved the score instead was aggregating predictions over the subject grouping that ships inside the input format, and reproducing the reference benchmark’s training recipe exactly rather than approximately. Two label-free transductive corrections that read only the unlabeled test distribution account for the remain- der. Fig. 1 puts the four measured contributions on one axis. Contributions. 1. A system that reaches 0.69447 macro-F1 on the MoCha 2026 hidden test, rank 1 of 58, +0.114 over the runner-up and +0.266 over the released base- line, from 14 KB of trained parameters on a frozen public encoder. 2. A component ablation in which every number is a paid read on the deciding test set (Sec. 5), including the finding that subject-level posterior aggregation is worth more (+0.143) than every model change we made combined. 3. A measurement result that generalizes beyond this challenge: our leave-two- cohort-out cross-validation was anti-correlated with the hidden test (Sec. 4). 4. The full negative record (Sec. 6), and the ceiling our own best stage imposes (Sec. 7). 2 Task, data and metric 2.1 What goes in and what comes out The unit of prediction is a walk, and walks arrive already grouped by subject. The evaluation server calls predict(data) once with the entire hidden test. Aggregate, Don’t Adapt3 walk 1 → 1 walk 2 → 1 walk 3 → 2 walk 4 → 2 22112212221 one label per walk ≠ one label per subject all 11 walks of T-SDU-PD/P003 Fig. 2: Severity is labeled per walk, and a subject’s walks do not have to agree. Four walks of one T-SDU-PD subject, drawn from the released canonicalized SMPL motion through forward kinematics and projected to the side view the encoder reads; three poses are shown per walk. Two carry ground-truth class 1 and two carry class 2. The strip gives all eleven walks of that subject. Averaging over this grouping is what buys +0.143 macro-F1 (Sec. 5); the residual disagreement inside it is what caps the system (Sec. 7). data[subject_id][walk_id] holds one canonicalized SMPL sequence: an axis- angle pose array of shape (T, 72), a global translation of shape (T, 3), a shape vector zeroed for privacy, and the original capture frame rate. The return value is predictions[subject_id][walk_id], an integer severity class in 0, 1, 2, 3, one per walk. Missing or invalid predictions count as wrong. That two-level dictionary is the single most important fact about the task. The subject identity is anonymized, but the grouping is given, at test time, for free. Fig. 2 shows what one subject’s walks look like and how their labels behave. 2.2 Label space Three properties of the label space drive every design decision later, and Tab. 1 holds all three. Training uses CARE-PD [2], a multi-site anonymized clinical dataset released as nine cohorts of canonicalized SMPL walks, four of which carry the UPDRS_GAIT annotation this task predicts. Severity 3 is rare in those four, at 44 walks in 2952, and absent from two cohorts entirely. Cohort class frequencies differ sharply, with class 2 ranging from 0.15 in PD-GaM to 0.44 in T-SDU-PD. And 62 of 110 subjects carry more than one label across their own walks. The remaining five cohorts (DNE [7, 8], E-LC [11, 14], KUL-DT-T [5, 21], T-LTC and T-SDU [2]) ship without severity labels. We used them only as unlabeled data in a semi-supervised arm that we report as a negative result in Sec. 6. No cohort outside CARE-PD entered the system. 4Junlong Shen Table 1: The four UPDRS-labeled CARE-PD cohorts, counted from the released pickles. Columns 0–3 count walks per severity class, and mixed counts subjects whose own walks do not all carry the same label. Class 3 is rare and missing from two cohorts, and more than half of all subjects are mixed. cohortsubjects walks 0 1 2 3 mixed 3DGait43 90 24 42 14 10 16 PD-GaM30 1700 783 635 248 34 29 BMCLab23 781 341 276 164 08 T-SDU-PD14 381 96 118 167 09 all four110 2952 1244 1071 593 44 62 2.3 Metric The ranking metric has one degree of freedom that no provided artifact fixes, and it is worth a quarter of the score. Entries are ranked by macro-F1, with macro-precision, macro-recall, accuracy and quadratic weighted kappa (QWK) reported alongside. The challenge pages state the primary metric as “Macro F1” and no further. They do not say which class set the average runs over, and the starting kit ships a baseline and a submission API but not the scoring program, so the question cannot be settled by reading provided code. It matters because severity 3 is rare enough that a system can miss it entirely, and dividing by three rather than four changes the score by roughly a quarter of one class. We resolved it from the leaderboard, by submitting a constant predictor and reading back the score it earned. The server divides by four, so an unpredicted class costs a full quarter of the metric. Every number in this report is that four-class macro-F1. 3 Method The system is a frozen encoder with a linear probe and three post-hoc stages, and no encoder is fine-tuned anywhere in it. 1. Geometry. SMPL pose and translation go through forward kinematics to 17 3D joints. Those joints are decimated to approximately 30 fps by striding on the fps field shipped with each sequence, projected to a single side view, cut into 81-frame clips, and normalized by a per-clip crop-scale. 2. Frozen features. A MotionAGFormer-S encoder [15] pretrained on Hu- man3.6M [9], used unmodified, produces a per-clip (81, 17, 512) representa- tion. We average over valid frames, then over clips, then over the 17 joint tokens, giving one 512-d vector per walk. Joint-mean pooling beat keeping the per-joint 8704-d representation, which is why the head is small. 3. Standardization. Features are z-scored with the feat_mean and feat_std stored inside the trained head. Aggregate, Don’t Adapt5 4. Transductive centering. We subtract a shrunken estimate of the test fea- ture mean, X ← X − c ̄ X test with c = 0.90. This is the only distribution- alignment operation that transferred; Sec. 6 lists the eight that did not. 5. Linear head. A single 4×512 linear layer trained with focal loss [12] (α = 1, γ = 1) and AdamW, producing per-walk posteriors. 6. Subject-mean aggregation. Per-walk posteriors are averaged within each subject id of the released data[subject_id][walk_id] grouping, at λ = 1.0, so every walk of a subject receives the same label. The aggregator is the uniform arithmetic mean. 7. Subject-kNN posterior pooling. Each subject’s posterior is blended to- ward its k = 5 nearest subjects by cosine similarity on subject-mean embed- dings, with weight λ nbr = 0.40. 8. Operating point. A q-divisor logit adjustment [16], P ′ ∝ P/q τ with τ = 0.50, where q is the model’s own predicted class marginal on the test set, followed by argmax. Two properties we want stated explicitly. The system is transductive but label- free. predict() receives the entire hidden test at once, and stages 4 and 8 read only unlabeled test statistics: its feature mean and its own predicted class marginal. No hidden-test label is used, inferred or reconstructed at any point, no hidden-test data was downloaded, scraped or exported from the evaluation server, and no attempt was made to re-identify any subject. The only grouping used is the anonymized subject id supplied in the input. The system also ex- ploits released structure, not leakage: that grouping is part of the input format the organizers ship, and aggregating over it is the single largest contribution here. Training used the CARE-PD cohorts and a public pretrained encoder and nothing else. Training. Feature extraction runs the frozen encoder over the four UPDRS- labeled CARE-PD cohorts (110 subjects, 2952 walks) under three augmentations that transferred, namely spatial jitter, temporal speed-warp and frame dropout, giving 44280 augmented rows. The head is then trained in two phases following the reference benchmark protocol: the epoch count is tuned on a 15% within- train split by macro-F1, then the head is refit on 100% of the labeled data for that epoch count. Total training cost is under one GPU-hour, dominated by feature extraction. Inference needs only torch and numpy, since the SMPL forward kinematics and the joint regressor are vendored into the submission. It is far inside the 3600 s evaluation limit: re-runs of the delivered artifact took 22–41 s for a 284-walk probe including model load on one partition of an H100 GPU, against a 372-walk real test. 4 How we measured Our cross-validation was anti-correlated with the score that decides the compe- tition, and finding that out was the most useful measurement we made. Partway 6Junlong Shen 0.430.440.450.460.470.48 leave-two-cohort-out CV macro-F1 0.550 0.575 0.600 0.625 0.650 0.675 0.700 hidden-test macro-F1 server best server worst ρ= −0.373, r= −0.760 (n= 11) full aggregation (λ= 1) partial (λ< 1) Fig. 3: Our local cross-validation ranked configurations backwards. Each point is one of the eleven configurations for which we hold both a leave-two-cohort-out score and a paid hidden-test read. Filled points use full subject aggregation, open points use less. The inversion is concentrated on the aggregation axis, which is also the axis worth the most. through the challenge we ran a calibration we should have run at the start. We replayed the eleven configurations for which we held paid hidden-test reads through our own local cross-validation, a nested leave-two-cohort-out protocol over the four labeled cohorts, and rank-correlated the two. Fig. 3 shows the result: Spearman ρ = −0.373 and Pearson r = −0.760. The server’s best con- figuration scored close to the lowest locally, and the server’s worst scored the highest. The inversion is worst exactly where it costs most. Cross-validation preferred no subject aggregation by−0.042, while the hidden test preferred full aggregation by +0.143. This is a property of the estimator rather than a bug in it. Shuffling the training labels collapses the same estimator from 0.4416 to 0.2234, so it does measure real signal. It simply measures a quantity whose ordering is inverted with respect to cross-site transfer for these levers. Our reading is that leave-two- cohort-out folds reward configurations that fit the idiosyncrasies of whichever cohorts remain, and subject aggregation deliberately discards exactly that kind of within-cohort detail. We therefore retired cross-validation as a selector for every decision, aggrega- tion and operating-point lever, and selected on hidden-test reads instead. That is defensible here because the challenge is a single-phase evaluation on a fixed test set with a deterministic scorer. There is no private re-split behind the visible board, so maximizing over reads is exact optimization of the announced objec- tive rather than selection on noise. Every number in Sec. 5 is consequently a paid read on the hidden test. Two practical consequences shaped the rest of the run. First, each submission spends one of three daily evaluation slots, which is why we call every hidden-test measurement a paid read. Reads are scarce, so we screened candidates offline for Aggregate, Don’t Adapt7 Table 2: Final standing on the MoCha 2026 hidden test. Every column is the orga- nizers’ scorer on the hidden test; our ranked entry is Codabench submission 882979. Leaderboard 18564, phase 27428, read at the close of the phase on 2026-08-16, when the board carried 58 entries. Boards stay editable after a phase closes: a re-read on 2026-08-20 listed 50 entries with this ordering unchanged. # teammacro-F1 macro-P macro-R accuracy QWK 1 JLShen (ours)0.69447 0.72347 0.67659 0.65591 0.60007 2 brady_kinesia0.58070 0.64650 0.55449 0.53495 0.40674 3 unist_visionlab0.57358 0.56464 0.58659 0.54301 0.42932 4 Nottingham_RVCE0.55577 0.59830 0.53448 0.53763 0.42050 5 tuananh10070.54907 0.59898 0.52355 0.48925 0.43321 6 anhnamxtanh0.54097 0.58588 0.51573 0.50806 0.40438 organizers’ released baseline 0.42890— distinctness, running each deployable artifact on a local probe and diffing the returned label dictionaries. Configurations that emit identical predictions are zero-information submissions. That screen killed a k-sweep, an entire evidence- weighting family, and four of six operating-point re-tune cells for zero submis- sions. Second, we re-derived the evaluation set’s size from the leaderboard’s own accuracy column, since every reported accuracy is an exact rational k/N: 40 readings of ours and all ten visible competitors’ land on N = 372. This is arith- metic on publicly displayed leaderboard numbers, it involves no test datum or label, and it conferred no scoring advantage. It did correct an internal premise that had wrongly closed a lever family. A premise-based kill dies with its premise. 5 Results 5.1 Final standing The field converged on the reference recipe and we did not, which is where the 0.114 gap comes from. In Tab. 2 the entire visible board except our entry sits inside the published CARE-PD cross-site band of roughly 0.52–0.55 macro-F1 for a single frozen encoder with a linear probe [2]. The rivals sit about 0.11 below us on accuracy while sharing our precision-leaning profile, so the difference is base separability rather than operating-point tuning. 5.2 Which stage bought the margin One row of Tab. 3 is worth more than all the others together, and the four questions below take the ladder in order of what each stage settles. Does the architecture or the training recipe carry the reference system? The recipe, worth +0.057. Our first probe on frozen MotionAGFormer-S features 8Junlong Shen Table 3: Component ladder. Every row is a paid read of the organizers’ scorer on the hidden test, never a local estimate. Rows were measured at different points in the run, so each ∆ is against the row above it on the pipeline as it then stood, not a strictly nested leave-one-out ablation. stagemacro-F1 ∆ organizers’ released baseline bundle0.4289— frozen MotionAGFormer-S + plain linear probe0.467 +0.038 + benchmark-exact head recipe (focal loss, AdamW, z-score) 0.524 +0.057 + transductive centering & q-divisor operating point0.5407 +0.017 + subject-mean posterior aggregation (λ = 1)0.68371 +0.143 + centering shrinkage retuned c = 0.8→ 0.90.68699 +0.0033 + subject-kNN posterior pooling (k=5, λ nbr =0.40)0.69447 +0.0075 was a plain logistic-regression head with cross-entropy loss, and it plateaued at 0.467 for weeks in a way that read like a task ceiling. It was not a ceiling. Repro- ducing the CARE-PD benchmark’s exact head recipe moved the same features to 0.524: focal loss with α = 1,γ = 1, AdamW, and z-scored features, taken from the benchmark’s vendored winning-configuration JSON files rather than from prose. The architecture was never the problem. When reproducing a reference system, the training recipe is part of the system. How much is the released subject grouping worth? More than every model change combined, at +0.143. The input format groups walks by subject, and the hidden test is no exception. Averaging per-walk posteriors within a subject and emitting one label per subject beats every representation change we made put together. The sweep is monotone to the boundary: λ = 0.00 → 0.5407, 0.75 → 0.6489, 0.90 → 0.65828, 1.00 → 0.68371. The choice of aggregator matters too. The uniform arithmetic mean beat confidence-weighted pooling (0.66416), trimmed- 25% pooling (0.65837) and logit-space geometric pooling (0.60777); geometric pooling sharpens posteriors and cripples the downstream operating point. On this benchmark, aggregating over a grouping the organizers shipped is a first- class modeling decision rather than data plumbing. Can evidence be borrowed across subjects as well as within one? Yes, but only locally, and only for +0.0075. Blending a subject’s posterior toward its k nearest subjects has a sharp interior optimum: λ nbr of 0 → 0.68699, 0.40 → 0.69447, 0.45 → 0.67256, 0.70 → 0.62769, 1.00 → 0.4012. Its own limit confirms the mechanism. Blending toward the global test-mean posterior, the k = ∞ case, scores 0.67517, below base. Local pooling denoises and global pooling washes out genuine between-subject severity differences. The value of k itself is flat, with k ∈3, 8, 12 byte-identical to k = 5 at λ nbr = 0.40. Are the two continuous knobs tuned or merely lucky? Both are saturated and piecewise-constant. Centering shrinkage gives c 0.60→ 0.68042, c 0.70 = c 0.80→ Aggregate, Don’t Adapt9 0.68371, and c 0.90 = c 0.95 = c 1.00 → 0.68699. The operating point gives τ 0.40 = τ 0.45 → 0.66843, τ 0.50 = τ 0.52 → 0.68371, τ 0.55 → 0.68042, and τ 0.60 → 0.64783. Both were re-swept after the pooling stage was added, since pooling changes exactly the posterior peakedness that the operating point reads, and both held. 6 What did not work For a challenge report this is the more informative half, and we give it in full. Tab. 4 lists each family we closed by direct measurement rather than by intuition. Does an ensemble still help once the pipeline already averages? Not measurably. Eight-head same-recipe averaging cost −0.04 on the pre-aggregation base, at 0.5007 against 0.5407. We re-measured it on the final system, since a negative result is only valid for the base it was measured on. Two-head averaging of the two strongest heads gives 0.69106, five-head averaging also gives 0.69106, and two- head averaging with the most decorrelated head gives 0.68764, against 0.69447 for the single head. On the aggregated base, ensembling is roughly neutral at −0.003 rather than a collapse: subject-mean and kNN pooling already perform the variance reduction an ensemble would supply. A single-head no-op control through the same code path reproduced the deployed predictions exactly, so these numbers isolate the averaging. Selecting a voter by solo score beat selecting it by decorrelation, which is the opposite of the usual guidance. 7 Where the ceiling is 7.1 Our best stage is also our ceiling The stage that bought the margin is the same stage that now bounds it. On the labeled cohorts, 62 of 110 subjects carry mixed walk labels: 44 span two severity classes, 16 span three, and 2 span all four (Fig. 4). Forcing one label per subject therefore makes 901 of 2952 walks unreachable, or 30.5%, a hard walk-accuracy ceiling of 69.5%. We realize 65.59%. The λ = 1.0 family is thus about 95% saturated, and every lever inside it fights over a handful of walks while the aggregation stage discards far more. Every attempt we made to re- admit per-walk evidence on top of aggregation lost. Recovering within-subject variation without giving up the variance reduction that aggregation buys is the open problem this benchmark now poses. 7.2 The head is unstable to its random initialization Part of our margin is a favorable draw, and the paper would be misread without that stated. Retraining the identical recipe with different seeds gives hidden-test macro-F1 of 0.694, 0.681, 0.652, 0.606 and 0.558 across five draws, a spread of 0.14 on a 4× 512 linear probe fit to 110 subjects. The shipped head is the best 10Junlong Shen 1234 distinct severity labels within a subject 0 10 20 30 40 50 subjects 48 44 16 2 62 of 110 subjects are mixed ceilingrealised 0.0 0.2 0.4 0.6 0.8 1.0 walk accuracy 0.695 0.656 one label per subject Fig. 4: One label per subject is worth +0.143, and it is also a hard bound. Left: most subjects in the labeled cohorts carry more than one severity label across their own walks. Right: predicting one label per subject can reach at most 69.5% walk accuracy on those cohorts, and the deployed system realizes 65.6% on the hidden test. Both panels count released labels only. of those draws, which is also why we release the exact trained head rather than a training script alone. Seed ensembling does not remove the variance, as Sec. 6 shows. A difference of±0.02 between two systems on this benchmark should not be read as a method effect. 7.3 A single operating point is wrong for every site at once Per-cohort severity distributions differ sharply (Tab. 1): class 3 is absent from two of four cohorts, and class 2 ranges from 0.15 to 0.44. A per-site operating point therefore has real headroom, and an oracle-site-label arm gains +0.0221 in cross-validation. Every deployable version we built lost. Site clustering in the embedding space reaches only 0.81–0.84 purity, while the features that cluster sites at 1.00 purity are precisely the hand-crafted features that are unusable as severity inputs. The mechanism is real and the estimator for it is not. 7.4 A possible irreducible component The same gait may receive different UPDRS ratings across raters and protocols. If cohort-specific label calibration is present, part of unseen-site generalization is not recoverable from the source labels at all, and no amount of representation work would close it. We could not test this without rater-level metadata. We flag it as the most likely explanation for the residual we could not move, and note that rater-level annotations would falsify or confirm it directly. Aggregate, Don’t Adapt11 7.5 Scope Everything above is measured on one benchmark, one metric and one hidden test. The claim we defend is that on this task the released grouping and a first- moment transductive correction dominate the representation, not that this holds for clinical gait assessment in general. The negative results bound the families we ran, not the space of methods: they say that no encoder, alignment, ensemble or decision rule we tested converts, and a different family could. 8 Reproducibility and data availability Code. The code accompanying this report is public and MIT-licensed at https: //github.com/jlshen025/codabench/tree/main/mocha. It carries the infer- ence code of the ranked entry, the 14 KB trained head that is the system’s only trained parameters, and the three-step chain that produced it. It also carries a verification script. That script assembles the exact runtime layout, checks every binary’s md5, and runs predict() on a probe built from the released cohorts, so the check needs no hidden data and reads no label. Running it returns the de- ployed head’s class marginal [112, 92, 74, 6], the marginal recorded for the ranked entry. The three third-party binaries the entry loads, namely the pretrained en- coder, the SMPL body model and the joint regressor, are referenced with check- sums rather than republished; all three ship with the CARE-PD release. Data. All training and evaluation data is CARE-PD [2], used under C BY- NC 4.0 and obtained from the official release. The model was trained on the four UPDRS-labeled cohorts: 3DGait [24], BMCLab [19], T-SDU-PD [2] and PD-GaM [3, 4]. The five unlabeled cohorts named in Sec. 2 were read only by the semi-supervised arm reported as a negative result. The frozen encoder is MotionAGFormer-S [15] as released by its authors, pretrained on Human3.6M [9] and not further trained here. The hidden test labels were never accessible to us. 9 Conclusion The winning margin on this benchmark did not come from a better motion rep- resentation. It came from taking the reference recipe literally, from aggregating over a grouping that sat in the input format all along, and from two label- free transductive corrections, on a frozen public encoder with 14 KB of trained parameters. For the next entrant, that ordering is the practical message: read the input format before the model zoo, and price a benchmark’s exact training recipe as part of the benchmark. For the next iteration of the challenge, the ceiling analysis in Sec. 7 suggests the benchmark is now measuring subject-level severity rather than walk-level severity, and that scoring subjects explicitly, or releasing rater metadata, would sharpen what it asks. The methodological finding we would most want carried forward is the one that cost us the most. We spent weeks trusting a local cross-validation that was 12Junlong Shen anti-correlated with the score deciding the competition. A handful of paid reads spent early on calibrating that estimator against the objective would have paid for themselves many times over. Where a competition exposes a deterministic scorer on a fixed test set, calibrating the local estimator against it is not overhead. It is the first experiment. Acknowledgements We thank the MoCha 2026 organizers for the benchmark, and the CARE-PD team for the dataset, which is used here under C BY-NC 4.0. The motion encoder is used frozen and unmodified. Compute was provided by the Digital Research Alliance of Canada. The author declares no competing interests. References 1. MoCha 2026: Workshop on human motion challenges in real-world and clinical settings, benchmark and challenge on Parkinsonian gait. https://mocha.care- pd.ca/, Codabench competition 16752 (2026) 2. Adeli, V., Klabučar, I., Rajabi, J., Filtjens, B., Mehraban, S., Wang, D., Seo, H., Hoang, T.H., Do, M.N., Muller, C., de Oliveira, C.N., Coelho, D.B., Ginis, P., Gilat, M., Nieuwboer, A., Spildooren, J., McKay, J.L., Kwon, H., Clifford, G., Esper, C.D., Factor, S.A., Genias, I., Dadashzadeh, A., Shum, L., Whone, A., Mirmehdi, M., Iaboni, A., Taati, B.: CARE-PD: A multi-site anonymized clinical dataset for Parkinson’s disease gait assessment. In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 3. Adeli, V., Mehraban, S., Mirmehdi, M., Whone, A., Filtjens, B., Dadashzadeh, A., Fasano, A., Iaboni, A., Taati, B.: GAITGen: Disentangled motion-pathology im- paired gait generative model — bringing motion generation to the clinical domain. arXiv preprint arXiv:2503.22397 (2025) 4. Dadashzadeh, A., Duan, S., Whone, A., Mirmehdi, M.: PECoP: Parameter effi- cient continual pretraining for action quality assessment. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). p. 42–52 (2024) 5. Filtjens, B., Ginis, P., Nieuwboer, A., Slaets, P., Vanrumste, B.: Automated freez- ing of gait assessment with marker-based motion capture and multi-stage spatial- temporal graph convolutional neural networks. Journal of NeuroEngineering and Rehabilitation 19(1), 48 (2022). https://doi.org/10.1186/s12984-022-01025- 3 6. Goetz, C.G., Tilley, B.C., Shaftman, S.R., Stebbins, G.T., Fahn, S., Martinez- Martin, P., et al.: Movement disorder society-sponsored revision of the unified Parkinson’s disease rating scale (MDS-UPDRS): Scale presentation and clinimetric testing results. Movement Disorders 23(15), 2129–2170 (2008) 7. Hoang, T.H., Zallek, C., Do, M.N.: Smartphone-based digitized neurological ex- amination toolbox for multi-test neurological abnormality detection and docu- mentation. IEEE Journal of Biomedical and Health Informatics (2024). https: //doi.org/10.1109/JBHI.2024.3439492 8. Hoang, T.H., Zehni, M., Xu, H., Heintz, G., Zallek, C., Do, M.N.: Towards a comprehensive solution for a vision-based digitized neurological examination. IEEE Journal of Biomedical and Health Informatics 26(8), 4020–4031 (2022). https: //doi.org/10.1109/JBHI.2022.3167927 Aggregate, Don’t Adapt13 9. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence 36(7), 1325–1339 (2014) 10. Kirichenko, P., Izmailov, P., Wilson, A.G.: Last layer re-training is sufficient for robustness to spurious correlations. In: International Conference on Learning Rep- resentations (ICLR) (2023) 11. Kwon, H., Clifford, G.D., Genias, I., Bernhard, D., Esper, C.D., Factor, S.A., McKay, J.L.: An explainable spatial-temporal graphical convolutional network to score freezing of gait in Parkinsonian patients. Sensors 23(4), 1766 (2023). https://doi.org/10.3390/s23041766 12. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection. In: IEEE International Conference on Computer Vision (ICCV) (2017) 13. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics 34(6), 248:1– 248:16 (2015) 14. McKay, J.L., Goldstein, F.C., Sommerfeld, B., Bernhard, D., Perez Parra, S., Fac- tor, S.A.: Freezing of gait can persist after an acute levodopa challenge in Parkin- son’s disease. npj Parkinson’s Disease 5, 25 (2019). https://doi.org/10.1038/ s41531-019-0099-z 15. Mehraban, S., Adeli, V., Taati, B.: MotionAGFormer: Enhancing 3D human pose estimation with a transformer-GCNFormer network. In: IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV) (2024) 16. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: International Conference on Learning Represen- tations (ICLR) (2021) 17. Saerens, M., Latinne, P., Decaestecker, C.: Adjusting the outputs of a classifier to new a priori probabilities: A simple procedure. Neural Computation 14(1), 21–41 (2002) 18. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neu- ral networks for group shifts: On the importance of regularization for worst-case generalization. In: International Conference on Learning Representations (ICLR) (2020) 19. Shida, T.K.F., Costa, T.M., de Oliveira, C.E.N., de Castro Treza, R., Hondo, S.M., Los Angeles, E., Bernardo, C., Dos Santos de Oliveira, L., de Jesus Carvalho, M., Coelho, D.B.: A public data set of walking full-body kinematics and kinetics in individuals with Parkinson’s disease. Frontiers in Neuroscience 17, 992585 (2023). https://doi.org/10.3389/fnins.2023.992585 20. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: FixMatch: Simplifying semi-supervised learning with consistency and confidence. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) 21. Spildooren, J., Vercruysse, S., Desloovere, K., Vandenberghe, W., Kerckhofs, E., Nieuwboer, A.: Freezing of gait in Parkinson’s disease: The impact of dual-tasking and turning. Movement Disorders 25(15), 2563–2570 (2010). https://doi.org/ 10.1002/mds.23327 22. Sun, B., Saenko, K.: Deep CORAL: Correlation alignment for deep domain adap- tation. In: European Conference on Computer Vision (ECCV) Workshops (2016) 23. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test- time adaptation by entropy minimization. In: International Conference on Learning Representations (ICLR) (2021) 14Junlong Shen 24. Wang, D., Zouaoui, C., Jang, J., Drira, H., Seo, H.: Video-based gait analysis for assessing Alzheimer’s disease and dementia with Lewy bodies. In: Applications of Medical Artificial Intelligence (AMAI), MICCAI Workshop. p. 72–82. Springer (2023). https://doi.org/10.1007/978-3-031-47076-9_8 25. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: MixSTE: Seq2seq mixed spatio- temporal encoder for 3D human pose estimation in video. In: IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) (2022) 26. Zhou, D., Bousquet, O., Lal, T.N., Weston, J., Schölkopf, B.: Learning with local and global consistency. In: Advances in Neural Information Processing Systems (NeurIPS) (2003) 27. Zhu, W., Ma, X., Liu, Z., Liu, L., Wu, W., Wang, Y.: MotionBERT: A unified per- spective on learning human motion representations. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023) Aggregate, Don’t Adapt15 Table 4: Rejected families, each closed by measurement. “Server” marks a paid read of the organizers’ scorer on the hidden test; the remaining entries are local cross- validation, which Sec. 4 shows is a weak selector on decision and aggregation axes but a usable one for gross representation failure. familyoutcome Encoder fine-tuning Dead in four forms: unsupervised 2D→3D pretext fine- tuning (server 0.3605, far below frozen), anti-site-latch low- rank adaptation (LoRA) with gradient reversal and a bottle- neck, supervised LoRA, and contrastive self-supervision with a gradient reversal layer. Alternative encoders All below MotionAGFormer-S: MixSTE [25] (server 0.4081, including in its own native normalization pipeline), MotionAGFormer-B/L, MotionBERT [27], PoseFormerV2, MotionCLIP, TMR, an NTU-pretrained skeleton trans- former, MoMask, a probe on the Riemannian manifold of symmetric positive-definite matrices, and a time-series foun- dation model. Alignment beyond the first moment Only the mean transfers. CORAL [22] and diagonal- std alignment, median centering, PCA denoising, itera- tive nullspace projection, transductive batch normalization, quantile-CDF matching, and per-site clustered centering all lost. The last lost even with oracle site labels, so it is not a clustering-quality problem. Decision rules and objectives Ordinal and CORAL-style ordinal heads (server 0.4558, −0.23, bracketed over τ so not a mistuning artifact), thresh- old tuning, linear discriminant analysis, Saerens EM label shift [17], class-balanced centering, prototype and nearest- class-mean rules, a soft macro-F1 surrogate objective, the invariance objectives V-REx, GroupDRO [18] and IRM, and training-time de-confidencing. A parameter-free joint macro- F1 decoder returned predictions byte-identical to argmax on the server, so argmax is already F1-optimal here, and the recall gap is separability rather than decision-rule mis- calibration. Per-class decision scaling also lost in both direc- tions: suppressing class 3 gave 0.64882 and boosting class 2 gave 0.5715. EnsemblesServer-negative in every form: diverse-encoder blends, Mo- tionAGFormer ⊕ MixSTE (locally positive, server collapse to 0.3720), fine-tuned ⊕ frozen, and two-view side and back (cross-validation +0.032 inverted to −0.042 on the server). See the note below. Transductivein- ferencebeyond centering Graph label propagation [26], Nyström-RBF and metric learning, density-ratio importance weighting, FixMatch [20] over the five unlabeled cohorts, Deep Feature Reweight- ing [10], attention pooling over clips, bag-of-windows pool- ing, and self-training test-time adaptation [23]. The last was an exact no-op, changing 0 of 284 probe items. Hand-crafted clinical gait features Cadence, stride, asymmetry and variability features are 96– 98% cohort-predictable: they act as site detectors on this data, and a random-forest late blend improved the pooled mean while hurting held-out cohorts. A useful by-product is that k-means on those same features recovers the cohorts at ≈ 1.00 purity, so they are an excellent site signal and a poor severity signal.