• SAS Trains Sparse Selectors on the Final Task Loss: Continuous scores enter attention logits, bypassing hard Top-K's gradient block. Its advantage grows as attention budgets shrink.

• AIM Treats Shared Memory as an Access-Control System First: Visibility classification reaches 96.0%, but strict operation accuracy is only 58.8%. Full-lifecycle reliability remains a deployment weakness.

• High Medical Rubric Scores Cannot Gate Deployment Alone: Clinically relevant hallucinations may leave aggregate scores unchanged. Evaluations must separate error types and their consequences.

Also Notable

• Feed Costly Device Failures Back Into Training — BlueLM-GUI sends real-device failures into a data, training, and deployment loop to narrow the gap between sandboxes and production.
• Model Learning Processes, Not Static Student IDs — Language representations built from learning histories may better track changes in cognitive state.
• Policy Self-Distillation Adds a Second Supervision Channel — SCOPE-OPSD uses final-layer differences while exploring how to constrain this privileged signal.
• Judge Bias May Shift With Model Capability — Multi-judge ensembles can support capability-specific calibration, not just voting.
• Consonants Contribute Unequally to Word Intelligibility — Acoustic masking measurements can help therapists prioritize pronunciation targets with the greatest impact.
• SWARM Tests Propaganda Detection in Multilingual Search — It evaluates webpage content itself instead of relying only on source labels.

Read the full edition →