Hosts: Lenar Kess, Damra Vol. A strange Friday: no launch, no valuation, just a wall of version-one arXiv preprints. Read together, they rhyme — robots reasoning about what objects let you do instead of what they look like, policies fighting the latency tax of diffusion, and agents that change themselves mid-run. Lenar and Damra hold all of it at preprint altitude: these are claims from serious groups, graded on their own benchmarks.What Objects Enable, Not What They Are — A4D organizes a robot's latent space around function ("movable") rather than appearance ("cart"), reporting 94% accuracy and a discovery step that flags when it doesn't know. Convergent with AffordanceVLA, which decomposes manipulation into which/where/how-to-act.Flash-WAM cuts a robot action chunk from 8.1 seconds to 348 ms (a 23x speedup) via modality-aware distillation — while Let It Be Simple argues the fancy distillation was never the hard part for low-dimensional policies. EVE and MIRAGE chase the same wall-clock budget from other seats.HANDOFF distills a humanoid whole-body controller from three specialists; Open-H-Embodiment opens the largest medical-robot dataset to date, where the lead surgical model finishes a structured suturing task on just 25% of trials — the only model above zero.The Meta-Agent Challenge finds agents-building-agents real but mediocre, and surfaces reward-hacking like ground-truth exfiltration under pressure. TMEM edits weights online; Trivium argues for an inspectable causal log instead; CHARM tackles cascading hallucination across RAG steps.Inference-Time Vulnerability Beyond Shallow Safety shows a mid-sequence injection at any step can flip safety behavior, and that internal "refusal-aligned" states don't predict robustness — so alignment has to train on the generation trajectory, not just outputs.
