• A fuzzy function no longer needs a live API call. Program-as-Weights compiles a natural-language spec straight into a local weight file. A 0.6B interpreter matches a 32B model prompted directly, at 1/50th the VRAM, with a 10M-sample FuzzyBench released alongside.

• Training-free image generation, 10x faster, 1% quality drop. MrFlow fixes the artifacts that have dogged multi-resolution acceleration. No custom kernels, stacks on existing distilled models, near-zero adoption cost.

• An agent's memory should be an auditable contract. AgenticSTS builds a clean context for each decision through typed retrieval, so any memory component can be ablated on its own.

• Denser self-distillation is easier to break, not harder. SDPO throws cold water: the denser the self-distillation signal, the larger the drift, and out-of-distribution it forgets worse than GRPO. Stop treating it as the default stabilizer for continual training.

Also Notable

• Running SWE-Bench/GAIA Costs Thousands and Takes Days — PACE Predicts Agentic Scores From Cheap Non-Agentic Probes — swaps expensive end-to-end evals for proxyable single-capability tests.
• Autonomous Policy Evolution Gets Tangled Up in Final Scores — EvoPolicyGym Isolates Self-Editing Under a Fixed Interaction Budget — makes "does the agent improve itself" a separately quantifiable dimension.
• Overlapping Skills Make Reliable Calls Hard and Final Verifiers Too Coarse — SkillCoach Scores Skill-Use Step by Step With a Self-Evolving Rubric — moves evaluation from "right at the end" down to "how each step was called."
• Decoupling Semantic Motion From Pixel Rendering, WorldDirector Uses an LLM to Coordinate 3D Trajectories and Camera — a controllable world model whose objects remember and steer around each other.
• Video Grounding Benchmarks All Test Zero-Shot on Everyday Scenes — AnyGroundBench Probes VLMs on Rare Concepts and Complex Spatiotemporal Dynamics — pushes evaluation from "recognizes common things" into the long tail.
• Augmentations Built for Natural Images Destroy the Fine-Grained Topology Vein Recognition Relies On — AGVBench Runs a Reliability Check — a systematic evaluation of 30 augmentation strategies across five datasets.
• Diffusion Language Models Don't Condition on Timestep — So Do They Secretly Represent Denoising Progress Anyway — probes find a latent time representation hidden in DLM layers.

Read the full edition →