Capability Learns When to Stop

npj Metamaterials published an algorithmic route to laser-print fully connected nickel-titanium braids in one build, then tune local fiber density, thickness, and discontinuities for vascular-device concepts. Chemical etching followed by electropolishing reduced mean strut thickness from 318 to 170 micrometers and improved uniformity; radial tests placed the braided W9 and W12 samples within or below reported commercial ranges for chronic outward force and stiffness. The team demonstrated stents, valve frames, clot retrievers, flow diverters, bifurcations, and fenestrated structures. These are mechanical proofs, not implants: testing used one crimp cycle, while fatigue, corrosion, nickel release, biocompatibility, and in-vivo performance remain open. The useful expansion sits in geometry—compliance can change along one continuous structure.

FORM identifies a deformable material law from one tracked interaction, then reuses that model for planning. Across four material classes, the preprint reports identification in two to five seconds, versus roughly 10 to 25 minutes for iterative baselines; elastic properties stayed within 3.4 percent of reference values and elastoplastic properties within 2 percent. Hardware trials extended the same estimate to rod insertion, golf, shaping, and pouring, where glycerol targets showed a maximum mean error of 3.8 milliliters. The evidence remains narrow: two cameras observe prepared geometries, the planner chooses among assumed material-law families, and execution is open-loop after identification. Water pouring received only one trial per target. Restraint here is epistemic—probe once, model only what the next action needs.

CommSketch treats speaking while drawing as one design action instead of forcing intent into a separate prompt box. In a randomized study, 24 participants completed a 30-minute toaster-design task with sketch-only or concurrent speech-plus-sketch support. The multimodal condition produced higher perceived alignment after correction for multiple comparisons, and nine of 12 participants expressed experience-related intentions, versus two of 12 in the sketch-only group. The system still accumulates transcripts and sends them with the current sketch at user-triggered checkpoints; it does not align utterances to individual strokes. This is work in progress with a small, mostly design-trained sample, one simple task, and no measure of downstream product quality. Its design move is continuity: language carries changing intent without interrupting the hand.

A virtual-reality study with 34 adults with multiple sclerosis compared eight gait-feedback conditions across three walking trials each. Spatial auditory feedback closed 72.7 percent of the velocity gap while remaining below average in reported burden; multimodal feedback closed 95.6 percent but carried the highest burden. Five conditions sat on the study's benefit-burden Pareto frontier, while spatial vibrotactile and vibrotactile-plus-visual combinations were dominated. The reference speed came from 19 non-fallers in a separate 41-person dataset and weakly distinguished fallers from non-fallers; burden came from two single-item ratings. All effects are single-session and diagnosis-specific. The result is not that more cues work better. It is that rehabilitation feedback has a dose, and one channel may preserve most of the benefit.

REVOIR frames clarification as a value-of-information decision: ask only when the expected improvement outweighs interruption cost, action reversibility, and the chance to correct later. On the ADAPT benchmark, the preprint reports 58.8 percent preference satisfaction for seen personas and 56.0 percent for unseen personas—13 to 15 points above a fine-tuned clarification policy—while asking about five times fewer questions. ADAPT contains 16 personas and eight breakfast tasks, and the broader evaluation still relies mainly on simulated users. The policy is myopic, considering one clarification step rather than a full interleaving of questions and actions; real-user deployment remains future work. The restraint is explicit: uncertainty alone does not buy the right to interrupt.