• Active Learning Needs Stage-Aware Training: Retrain from scratch early, then fine-tune from checkpoints after stabilization. HybridAL cuts retraining time by up to 49% while preserving performance.

• ICL's Scarce Resource Is Now Selection Budget: DearICL ranks demonstration sets per query, improving accuracy by 8.08%–15.9% over a strong linear bandit baseline.

• Agent Consistency Depends on Relational World States: MARBO tracks roles, factions, and relationships so actions and language reflect updateable beliefs.

• Multi-Step History Could Protect Unedited Regions: MIEdit combines predictor-corrector sampling with automatic semantic angle masks to improve fidelity, stability, and efficiency.

Also Notable

• Model Event Interactions Through a Low-Rank Bottleneck — For extracting before, after, and simultaneous event relations, test whether a lightweight structure can replace simple concatenation.
• Move Cross-Lingual Alignment Down to Tokens — Token-level optimal transport offers finer control than sentence-level alignment while preserving language-specific information.

Read the full edition →