• Agentic Retrieval Trades 160× Latency for 8.7 nDCG@10 Points. With the same embedding model, average query latency rose from 0.67 to 107.4 seconds while consuming 764,100 input tokens. It may be best reserved for high-value, low-frequency work.

• OnePO Turns the Teacher Into Temporary Guidance. It first reinforces unlikely but informative teacher tokens, then removes guidance once the current policy beats the teacher. Training on 20K medical samples produced a 67.2 HealthBench Total score. That beat SFT+RL by 2.7 points and pure RL by 7.4 points. The reported evidence is limited to medical adaptation.

• ARRO Penalizes Both Missed and Unwanted Edits. A vision-language reward model audits candidate edits and generates rewards. On FLUX.1 Kontext-dev, average EditScore rose from 5.21 to 5.88, while unwanted pixel changes fell 8.4% across 600 samples.

Also Notable

• DP-ES Concentrates Privacy Budget on Whole-Prompt Evaluation — Population search avoids the drift of greedy token-level decisions under heavy noise, producing steadier results. Sensitive real-world deployments still need validation.
• Formal Verification Makes Algorithmic Stability Testable — The framework reduces dependence on analytical assumptions and computes reliable upper bounds on expected generalization gaps with a constant number of algorithm runs.
• RoSA Freezes Early Layers and Rotates Later Blocks — It reduces optimizer-state memory, shortens backpropagation, and works alongside existing PEFT methods. Forward-pass savings require caching activations from the final frozen layer.
• BrainTRACE Builds Auditable Evidence Chains for Brain MRI — Tests of longitudinal comparison, multi-sequence integration, and 3D localization show that current VLMs detect isolated clues but rarely combine them into grounded longitudinal judgments.
• Wander Unifies Three Graph Tasks With Random Walks — It frames node classification, homogeneous-graph link prediction, and knowledge-graph link prediction as partially observed graph completion. One checkpoint can expand its structural context at inference time.
• EpicWorldModel Uses Predictive Variance to Explore Occlusion — A stochastic JEPA represents several possible futures. The planner then searches unknown areas that are more likely to hide the target.
• MoCAR Encodes Trajectories and Coordinate Transforms Together — Continuous motion codes let autoregressive prediction extend latent states directly, avoiding repeated tokenization in trajectory space.
• TrustmeWatcher Connects State Predictions With Participant Explanations — The system combines activity logs, self-reported questionnaires, and privacy controls, but the current study includes only 17 participants.

Read the full edition →