• DiffusionOPSD Leads 19 of 20 Settings. It turns final image rewards into stepwise denoising targets on current-policy trajectories. It beats the strongest competitor by up to 44.0% while cutting training GPU hours by 40% and 63%.

• Better Memory QA Does Not Improve Satisfaction. Direct QA accuracy rose from 19.7% to 70.1%. User experience depended instead on whether the system cited memories naturally and at the right time.

• OPDVR Makes Token-Level Advice Pass Task-Level Verification. Online distillation provides dense guidance, while verifiable rewards protect correctness. Both signals stay aligned on current-policy trajectories.

• TorchMorph Closes a Gap in GPU Pipelines. Its 22 batched operators support up to eight spatial dimensions. They avoid CPU round trips that can quietly erase end-to-end speed gains.

Also Notable

• Separate Working Memory From Experience Memory So Long-Horizon Agents Can Invoke Skills From Current State Instead of Reprocessing Full Histories. Recuris explores recursive experience development..
• Search Agents That Receive Only Final Success Signals May Accumulate Errors Long Before Feedback Arrives. CAFE co-develops within-trajectory feedback and search policies..
• Test-Time Scaling for Multi-Turn Search Should Not Trust Model Confidence Alone. Aggregation should also account for whether a trajectory found real evidence..
• Sparsity, Quantization, and Low-Rank Approximation Each Hit Their Own Accuracy-Efficiency Wall. Compression Trinity tests whether they complement one another without compounding errors..
• Short Counseling Responses May Convey Better Listening Than Information-Dense Answers. LLM evaluation should distinguish saying less from helping less..
• Generating Tests From Model-Written Bugs Can Improve Both Verifiers and Dense Rewards. The approach targets test coverage as a limit on reward quality in code RL..
• Citation Errors in Deep-Research Reports May Originate Before the Final Writer. Tracing faithfulness and citation recall across agent stages makes system repair more targeted..
• Low-Rank Adaptation for Ternary Weights Tries to Avoid Dequantizing Before Merging. The goal is to preserve efficiency when fine-tuning extremely low-bit models..
• Improved Randomly Masked Images Can Carry Knowledge From Multiple Teachers Without Original Training Data. IDeaL tests how much real visual structure data-free distillation needs..
• Moving Guardrails Before Tool Execution and Applying Them Step by Step Better Matches Deployment Needs. StepGuard explicitly balances safety with task utility..
• Uncertainty Covariances in Geometric Learning Must Satisfy Both Equivariance and Positive-Definiteness. This keeps confidence estimates consistent with the geometry of tensor predictions..

Read the full edition →