• Iris Turns Web Links Into Training Questions. Multi-hop entity chains and difficulty filters produce fresh supervision for iterative “SFT-RL climbing.”

• MaxKernel Puts Compiler Feedback Into The Code Generation Loop. It switches among human collaboration, automated iteration, and multi-agent graph search.

• Chain-Of-Thought Operations Are Separable In Hidden Representations. Task framing, goal decomposition, and deduction differ most clearly in middle layers, but classification does not prove causality.

• Layer Dropout Saves Up To 25% Of Training Compute. It also prepares models for layer removal, raising inference speed by up to 1.5×.

Also Notable

• Evaluate Complete Agent Deliveries, Not Single Tool Calls. τ^τ-Bench tests whether coding systems can complete end-to-end builds under constraints resembling real customer projects.
• Visual Safety Evaluation Needs Component-Level Responses. One request may contain both answerable and rejectable parts, while blanket refusal hides fine-grained judgment.
• Build Synthetic Teachers Through Self-Extrapolation. RISE aims to avoid external-teacher distribution mismatch and the capability ceiling of ordinary self-distillation.
• Response Structure May Reduce False Refusals. Studying the structure of safety-tuned answers could help separate harmful requests from benign prompts containing sensitive words.
• Put An Explicit Price On Avoiding Animal Harm. HarvestBench measures how agents trade goal rewards against harms that instructions never explicitly mention.
• Quantization Errors Accumulate In Recurrent Models. Low-precision evaluations must test errors across time, not only single-step outputs or static weights.

Read the full edition →