Word-recall pauses now carry their place in the semantic hierarchy: TrajLift reaches 80.38% accuracy on 30-minute delayed recall
Synopsis
In verbal memory tests, the order of recalled words and the pauses between them carry diagnostic information, but a pause means different things inside a semantic category and at a switch between categories. TrajLift treats a recall episode as a timed trajectory on a semantic hierarchy, uses heat diffusion to lift each word onto several hierarchical scales, and lets the inter-item latency modulate the diffusion time so structure and timing enter one operator. On Hong Kong List Learning Test data from 387 participants after excluding 15 with annotation issues, accuracy on the 30-minute delayed-recall trial was 80.38%, 4.5 percentage points above a handcrafted baseline using recall count alone; the paired difference on the 10-minute trial was not significant.
Interpretation
A verbal recall sequence is now modeled as a timed trajectory on a semantic hierarchy, so each recalled word enters the representation together with the pause before it rather than being reduced to a total score or a flat string. Earlier practice either handed the recall sequence to a text encoder as a flat string or, following clinical habit, summarized the whole trajectory with recall count, category transitions, semantic clustering and mean latency. On the 30-minute delayed-recall HKLLT-5 trial, TrajLift reached 80.38% accuracy, 4.5 percentage points above both logistic-regression baselines built on recall count and on a seven-dimensional low-level feature set, while adding mean latency, semantic clustering, category transitions and demographic features to recall count brought no improvement.
Each word is first placed at several scales of the semantic hierarchy, and the following pause then writes timing into the same operator by changing the diffusion time. Heat diffusion on a hierarchically weighted graph has a known property: the diffusion time, a continuous scalar, controls the structural resolution of the resulting representation, and longer diffusion keeps coarser levels. The hierarchy has three levels, items, categories and domains, matched by micro, meso and macro diffusion scales; on the three variants of the synthetic ST-XOR benchmark that require structure-time interaction, TrajLift held 98% to 99% accuracy, including the three-level ST-XOR-wh variant where all baselines fell to 47% to 50%.
The advantage over conventional measures grows as retrieval timing becomes more informative. The two delayed-recall trials share the word list, the participants and the hierarchy, and differ in how strongly latency separates the two groups: the effect size of mean inter-item latency on the 30-minute trial is roughly twice that on the 10-minute trial. Setting all latencies to zero or to their mean at test time lowered accuracy by 6.8 and 17.3 percentage points on the 30-minute trial, against 3.9 and 7.2 points on the 10-minute trial, mirroring the difference in latency effect sizes.
Perspective
The demonstrated scope is three variants of the authors' synthetic ST-XOR benchmark and Trials 4 and 5 of a single-center Hong Kong List Learning Test corpus of 387 participants aged 60 and over, labeled normal cognition versus neurocognitive disorder; the authors state the work is at the research stage and has not been deployed in any clinical or real-world setting, and call for validation on larger multi-site cohorts and other domains involving timed trajectories on hierarchical structures.
The spectral analysis assumes a balanced hierarchical partition, which the authors say holds for the test protocol used and appears sufficient rather than necessary in synthetic experiments; each synthetic variant places the structure-time interaction at a single hierarchical level, so the benchmark does not test whether combining multiple scales is necessary; the clinical data are small, the labels are syndromic neurocognitive disorder rather than Alzheimer's-specific, and the cohort is local, so distributional shifts across populations have not been evaluated.