• The real lever for a deep-research agent is the gap between verifying and finding, not a stronger base model. AREX keeps partly-verified intermediate states and patches only the constraints that didn't hold. A 4B model matches rivals many times its size.

• Force a model to write spatial answers as text and you may be scoring translation, not spatial skill. ProVisE lets an image generator answer directly in pixels. Freed from the text interface, it catches up to text-output VLMs.

• How much of a coding-agent leaderboard score is memorized? Tencent's WorkBuddy reverse-engineers every task from a real commit to resist contamination. It refuses to publish one overall average, because its four subsets score on different standards.

• Mostly linear attention plus a quarter softmax makes single-GPU video generation 3.2x faster. SANA-Video 2.0 uses that softmax anchor to restore global dependencies, reaching 720p on one H100. The "matches full-softmax quality" claim still needs real testing.

Also Notable

• Users Refine Requirements Mid-Conversation, but Models Train and Test on One-Shot Prompts — This paper quantifies how they lose the thread step by step as intent drifts.
• Evaluate a Coding Agent as an Interactive Project Builder, Not a Spec-to-Code Filler — Starting from vague product intent, it tests planning, requirement clarification, debugging, and repo-level construction.
• NVIDIA Writes an Agent as an Ordinary Python Object — Methods are actions, fields are state, docstrings are prompts, type hints are contracts; a method body of "..." gets filled in by the model at runtime.
• On-Policy Self-Distillation Drops the External Teacher but Still Needs Teacher-Student Asymmetry — It removes even privileged answers and visual evidence, manufacturing the asymmetry from the input conditions alone.
• Camera Motion and Object Motion Are Tangled in Video; Self-Supervision Pulls Them Apart — The result is a steadier, more structured motion representation.
• Sinusoidal Activations Induce a Harmonic Line Spectrum, and Recurrent Unrolling Enriches Its Support — A shared sine block iteratively refines the latent for high-fidelity implicit neural representations (accepted at ECCV).

Read the full edition →