• RL Helps Verifiers Avoid Dead Ends. RSB reranks branching heuristics with reinforcement learning. Across 600 hard instances, it solves 11% more while exploring 50% fewer branches than state-of-the-art branching heuristics. This does not mean models now carry safety guarantees.

• Visual Evidence Can Stay While Compute Shrinks. δ-Vision reconstructs layer-wise visual states with lightweight MLPs. It beats visual-token pruning methods at comparable or lower compute.

• Training Order Leaves an Actionable Weight Trace. Targeted intervention narrows the loss gap in a Qwen-3-4B experiment. A separate four-LLM study identifies training order with 92% accuracy.

Also Notable

• More Changes Do Not Always Fix Agent Failures. ControlScope finds that the returns from continuing execution, editing only the next tool call's data arguments, or replacing the unfinished workflow vary by task, model sampling, and review timing.
• Model Merging Should Respect Transformer Structure. CASS selects attention heads and FFN neurons by task contribution, then uses structured masks to reduce interference between task vectors.
• A Full Persona Can Teach Its Summarized Self. OSPD switches divergence constraints based on sharply peaked teacher confidence at character-critical tokens and diffuse confidence around generic wording.
• Aggregate Accuracy Hides How Agents Search and Answer. AgentHop separates performance into four diagnostic axes: retrieval, synthesis, tool use, and resource management.
• Personality May Not Live in a Flat Space. In highly curved regions, steering along manifold geodesics is more coherent than Euclidean linear interpolation.
• Frozen VLMs Can Guide Medical Time-Series Retrieval. ViRe uses visual shape priors to locate temporal and channel evidence in raw numerical sequences. Across six public benchmarks, it improves overall performance by 6.42% relative to the previous best methods.
• 3D Anomaly Detection Can Learn Broken Local Geometry. Controlled synthetic anomalies help the model extract defect signals that transfer across categories.
• Whole-Slide Images Are More Than Isolated Patches. TMEvolve models signal propagation between stable tissue regions and heterogeneous boundaries through a reaction-diffusion process.
• Prompt Disagreement Can Locate Uncertain Segmentation Regions. Binary preferences can replace expensive pixel-level mask supervision in specialized domains.
• Newer ACL-Family Publications Have Similar GitHub Unavailability Rates. Empty and placeholder repositories have increased, creating a new reproducibility risk.

Read the full edition →