• Structural Equivalence Does Not Guarantee Equal Performance. A model can solve the original yet depend on its wording. Grouped isomorphic variants test whether rule reasoning is systematic.

• Agent Training Budgets Should Follow Learnability and Transfer. HarnessBandit prioritizes updates with strong learning signals that also carry across execution harnesses.

• A Correct Diagnosis Still Carries Verification Obligations. VeriDx checks evidence, alternative explanations, and contradictions to expose process failures hidden by the final answer.

• Continual-Learning Routers Must Track Task Distributions and Modality Reliability. Hyper-LLaVA uses uncertainty weighting to reduce modality imbalance and parameter-selection errors.

Also Notable

• Harmful Memes Hide Risk in Text-Image Contrast and Cultural Cues. Can Symbolic Language and Associative Retrieval Supply the Missing Context for Zero-Shot Detection? — SyRHM combines symbolic language reasoning with associative retrieval to detect implicit harmful intent.
• Aggregate Medical-Coding Metrics Can Hide Failures on Complex Cases. Which ICD Scenarios Expose the Limits of Agent Workflows? — This study examines the capability limits of agent-based ICD-10-CM coding and the errors concealed by aggregate metrics.
• A 218B-Parameter MoE Translation Model Activates 25B Parameters. Watch the Tradeoffs Among Cost, Throughput, and Instruction Following. — North Small Translate builds an instruction-tuned translation model on the architecture behind Command A Plus.
• Can a Lightweight Quantum-Classical Detector Turn Parameter Efficiency Into Real Gains on Embedded Underwater Hardware? — Quantum-Gated LiteSSD targets low-parameter object detection for forward-looking sonar.

Read the full edition →