• Stretch a task to 1.6 hours and frontier agents finish only a fifth of it. OSWorld 2.0 moves the computer-use yardstick from 30 tool calls to real workflows that take humans a median 1.6 hours and 318 calls. Claude Opus 4.8 finishes 20.6%, and the failures are all long-horizon context management.

• Policy checks should read the whole conversation, not block a single argument. PolicyGuard turns the verifier into a sub-agent that reads the dialogue, lifting pass rates 6–12 points on tau²-Bench airline while blocking half as often as argument-level guards.

• Video evals are shifting from "can it see" to "can it follow along and do the task." VG-GUIBench asks GUI agents to operate an interface by following a tutorial video, paired with TASKER, which picks keyframes using both task and scene signals.

• Portrait retouching switches from instructions to examples. MirrorPPR gives the model one before/after pair, infers the edit, and applies it to a new photo — trained on 47M pairs in a two-stage curriculum.

• In transparent scenes, monocular depth "ground truth" is just an annotation convention. One ray passes through glass and both the foreground and background depths hold geometrically. MD-3k finds depth models pick different layers, and a training-free spectral transform flips a frozen model to the other one.

Also Notable

• Apple Evicts KV Cache by Coverage, Not Raw Attention Score. Cheaper long-context inference with no quality drop.
• Scale Activations of Selected Attention Heads Without Touching Pretrained Weights. Near-zero extra parameters add spatial reasoning to a VLM.
• Does a Model Actually Read Back the Intermediate State It Wrote on Its Scratchpad? Stanford asks whether those variables feed the computation or just look nice.
• Let Multiple LLMs Invent a Compact Symbolic Code to Collaborate. It replaces verbose natural-language CoT for higher reasoning efficiency.
• Use "Bayesian Surprise" as the Reward Signal. Allen Institute steers an LLM through long-horizon hypothesis search-and-verify loops to pick what to test next.
• Reclaim the Idle Compute Edge AI Chips Reserve for Peaks. Huawei uses general approximation to harvest capacity that mostly sits idle.
• VLMs Inherit Relational Priors From Language but Can't Apply Them to Images. This work uses Gromov-Wasserstein to align semantic relations across language and vision.
• Face Video Super-Resolution Doesn't Have to Be Full Generation. Initialize from dynamic trajectories and skip the expensive fixed-sampling inference.
• 3D Scene Layout Stops Converting Assets and Coordinates to Text First. Working natively in 3D lets the LLM arrange more sensible layouts.
• Whether a Model Can "See" an Object Is Capped by the Description System It Learned. Microsoft borrows Wittgenstein to draw a boundary around "seeing."

Read the full edition →