• Multi-turn agent training is expensive because of decision density, not horizon length. Mila re-did the math: the real signal-killer is the flood of reward-equivalent routine actions, and signal-to-noise decays as ρ^(-1/2) — reproduced in a controlled setup at R²=0.999.

• RLVR moves to science, but higher scores aren't always real generalization. Mat-Pref splits its test set into in-distribution, unseen structure families, and cross-property transfer, and finds GRPO's gain over SFT looks more like reshaping the output distribution than learning new knowledge. A two-stage 8B model beats a 235B model zero-shot by 20-plus points on held-out families.

• Protein models predict accurately, but their attributions miss the real epitopes. Using real allergen epitopes as ground truth, ETH shows residue-level attributions align with true epitopes no better than chance. Treating attribution as biological evidence in high-stakes screening is a dangerous over-read.

• The old "BC pretrain, then AIL" trick finally has a theoretical guarantee. CoPT-AIL points out the real bottleneck is the error from learning the reward function from scratch, pretrains policy and reward together, and proves a tighter imitation-gap bound than standard AIL.

Also Notable

• Evaluating LLM Values Shouldn't Stop at Single-Question Behavior — This ACL work uses a symmetric Q-sort to measure how a model "structurally ranks" competing values, focusing on the internal consistency of its value system rather than item-by-item answers.
• Dynamic 3DGS Doesn't Have to Trade Motion Consistency for Visual Fidelity — Multi4D handles both at once with multi-level competitive allocation, giving anyone working on 4D reconstruction or dynamic scenes a new trade-off point.
• Zero-Shot Classification Shouldn't Contradict Itself Across Label Hierarchies — This CVPR work uses hierarchy-constrained contrastive learning to keep every classification level consistent and remove cross-level conflicting predictions.
• Infrared Small-Target Detection, Made Lightweight and Real-Time — A denoising-enhanced coarse-to-fine framework plus attention-prior-guided knowledge distillation, suited to anyone working on edge-deployed detection like drone surveillance.

Read the full edition →