• Pruning is not an unconditional shortcut. Princeton compares pruning against training from scratch under matched token budgets. Pruning wins reliably only when the training budget is tight. Loosen it and coarse-grained pruning gets caught and passed — only fine-grained pruning holds its edge.

• When a GRPO group is all right or all wrong, that batch of gradients is wasted. VISTA changes no algorithm. It builds comparison groups from multiple views of the same GUI instance, rescuing the collapsed relative signal and lifting ScreenSpot-Pro grounding accuracy from 55 to 63-67.

• Give a medical model's hallucinations a CT scan, not just a temperature reading. ClinHallu locates each hallucination at one of three stages — visual misperception, medical-knowledge recall, reasoning integration — so you know which part to fix instead of getting one more leaderboard.

• The bottleneck in audio-visual QA is data construction, not the model. OmniVideo-100K rebuilds sound-image links with entity-anchored scripts and writes questions backward from evidence, and its roughly 12% cross-benchmark gain shows the recipe transfers to your own pipeline.

Also Notable

• LLMs Are Turning From "Chat Generators" Into "Digital Coworkers" — today's top HF paper (16 upvotes) frames the shift along two axes, cognitive core plus continuous work; strong as a frame but macro-level, a quick scan for anyone wanting industry narrative coordinates.
• Agent Runtime Harnesses Are Still Mostly Hand-Built and Static — HarnessX tries to make prompt, tools, memory, and control flow into a composable, self-adapting, evolvable "foundry."
• Deployed Skill Packs Always Break on Edge Cases and API Changes — SkillAudit uses paired-trajectory auditing without ground truth to keep skills evolving, turning "where did the skill fail" into a locatable problem.
• A Full VLA Learning Stack From Data Collection to Real-Robot Deployment — HyVLA-0.5 covers model design, continual pretraining/SFT, RL post-training, and real-robot deployment end to end, for anyone wanting to see the full engineering assembly.
• One Layer's Junk Visual Token Can Be Another Layer's Treasure — existing token pruning mostly works at fixed layers; this one switches to layer-wise adaptive selection to lighten LVLMs.
• Split Learning Hides Two-Way Leakage — both the prompt side and the response side can be recovered; this paper offers attack and defense, accepted at ICML.
• Concept Erasure in Text-to-Image Often Overshoots and Deletes Normal Content — ForceForget uses reinforcement learning for concept removal, aiming for a tighter balance between safety and fidelity.
• Helpfulness Post-Training Flattens LLM Simulators Into Uniform "Obedient Assistants" — creating a Sim2Real behavior gap, which OdysSim narrows with large-scale data.
• Neural Networks Haven't Really Solved Systematicity — MIT pushes back, challenging the optimistic story that the cognitive puzzle is already solved.

Read the full edition →