A one-second window as a behavioral fingerprint
The paper asks whether very short slices of movement, rather than task-level scores such as success rate, contain enough information to identify an individual animal and to track its learning. The authors' hypothesis is that behavioral characteristics tied to identity and learning stay relatively stable across nearby time intervals, so two moments close in time should have similar representations.
That hypothesis sets the method's supervision: instead of labeling behavioral states or hand-designing kinematic features, TRAIL treats temporal proximity as the learning signal. A trajectory and a nearby trajectory are pulled together in a compact representation, and the resulting vector is then tested on separate prediction targets. The targets are identity, recording day, proficiency, success rate, and reward, which are different quantities and are reported separately throughout.
Temporal proximity replaces handcrafted features
TRAIL learns directly from movement trajectories. The encoder maps a short history of body keypoints into a compact representation, and contrastive training rewards similarity between temporally neighboring windows. No manual definition of behavioral states or kinematic features is required, which is the paper's main methodological departure from feature-engineering approaches.
The default representation has 16 dimensions, and the paper reports that an eight-dimensional version also performs well across identity, day, and proficiency. Ablations show that adding a reconstruction loss raises identity accuracy from 81.1% to 84.0% and lowers day error from 1.32 to 1.26 and proficiency error from .017 to .016, while multi-frequency modeling gives only modest further gains. Narrower positive neighborhoods, spanning 10% rather than 100% of a recording partition, lower day error from 1.465 to 1.258 and proficiency error from .0211 to .0156 in all six animals.
Identity is easier to decode than learning progress
On the alternation cohort, the full model reaches 84.0% identity accuracy, with day error 1.2579 and proficiency error .0156 on 105,882 test windows. Per-animal identity accuracy ranges from 72.70% for Rat A to 95.92% for Rat B, so the pooled figure hides substantial variation between animals.
The learning-related readouts are regression errors, not accuracies, and should not be read as the same kind of score as identity. In the held-out-day test, where days 1–3 train the encoders and probes and day 4 selects regularization, TRAIL reaches 74.87% identity accuracy on 139,875 day-5 windows, against 38.89% for BAMS with local coordinates and maze context, 25.90% for kinematic features, and 16.67% chance. On the same held-out day, rewarded-count error is 0.948 for TRAIL versus 1.608 for BAMS, with five of six animals improving.
Local pose and global motion carry different signals
The paper separates two views of the same movement. Local pose keeps the body keypoint at the origin and the body–neck segment vertically aligned, removing global translation and rotation while retaining relative body configuration. Global trajectory keeps only body position and the body–neck orientation over time.
Local pose alone reaches 77.1% identity accuracy, while global trajectory alone reaches 72.1%; the full movement representation reaches 84.0%. For learning-related variation the ordering reverses: global trajectory gives day error 1.4733 and proficiency error .0172, better than local pose at 1.6047 and .0189. The authors therefore describe egocentric pose dynamics as more predictive of identity and allocentric motion as more predictive of learning-related variation, and the full representation as combining both.
Decision junctions and turning motifs concentrate the signal
In the alternation cohort, the signal is not uniform across the maze. At the junction where the three arms intersect, identity accuracy is 88.7% and proficiency error is .0135, compared with 81.7% and .0211 on arm segments. The junction has only 326 scored windows, so its interval is wider than the arm-endpoint group's 2,377 windows.
Movement motifs show a similar concentration. Turning windows give 98.0% identity accuracy and proficiency error .0083 across 151 windows, while scanning gives 76.6% and .0195 across 1,080 windows. The probabilistic-reward cohort repeats the pattern: turning gives 96.63% identity accuracy and the lowest day, proficiency, and success-rate errors among motifs, while reward error is lowest during straight movement. These are subgroup estimates from the same fitted probes, not independent replications.
Gains persist across sessions, days, and new animals
The comparisons are not confined to a single split. On held-out sessions, TRAIL lowers day error to 1.765 versus 1.932 for BAMS, proficiency error to .02032 versus .02460, and reward error to 0.760 versus 1.179, with five, five, and six of six animals improving respectively. On a held-out day, identity accuracy is 74.87% for TRAIL against 38.89% for BAMS.
Transfer is tested by training an encoder on ten animals and freezing it before evaluation on four different animals, each supplying its own chronological calibration partition. TRAIL has lower error than kinematic features on every target in all four animals: day 1.621 versus 1.861, proficiency .00944 versus .01112, reward 0.860 versus 1.025, and success rate 0.05 versus 0.07. The paper also reports a within-clip identity comparison on MABe22 mouse triplets, where TRAIL reaches 73.65% against 63.08% for BAMS, an improvement of 10.56 percentage points.
Conditional generation makes the differences inspectable
Beyond prediction, the authors train a representation-conditioned diffusion model on 3-second trajectories of nose, neck, and body coordinates from the six alternation animals, with the movement encoder frozen. By holding the diffusion latent fixed and varying the conditioning representation, they generate trajectories that keep the broad motion structure while changing stopping, accelerating, and stretching.
These generated trajectories are qualitative illustrations, not measured outcomes. The paper presents them as a way to make behavioral differences associated with the learned representations directly inspectable and to suggest hypotheses for later experimental validation. They do not establish that the varied movement causes the identity or learning difference.
A predictive tool, not a causal account of learning
The evidence supports prediction from short movement windows in the studied settings: freely behaving rats in a Y-maze under alternation and probabilistic reward rules, plus benchmark mouse data. The paper does not show that movement patterns cause learning progress, and the readouts are behavioral proxies such as proficiency, success rate, and rewarded count rather than direct measures of a cognitive process.
The most decision-relevant split is between what each view and location buys. If the goal is identifying individuals, local pose and junction or turning windows carry the strongest signal; if the goal is tracking learning-related variation, global motion carries more, and the full representation outperforms either view alone. The paper's own framing is that short-term movement is an informative behavioral fingerprint and a hypothesis-generating tool, which is a narrower claim than a general theory of how learning unfolds.
Sources used in this answer
Representing short-term movement to reveal animal identity and learning progress
TRAIL, a self-supervised framework using temporal proximity as supervision, learns compact representations from short rat movement trajectories that predict identity and learning-related readouts better than the compared baselines, with trajectories as short as one second carrying informative signal.