• Legal RAG Should Prioritize Retrieval: An Uzbek-language study used 178 retrieval queries and 504 question-answer pairs. Targeted retriever tuning could narrow model gaps cheaply while supporting different cloud and local constraints.

• Recognition Does Not Trigger Refusal: Models from 1.7B to 70B parameters show detectable signals for impossible questions. Those signals sit almost orthogonal to refusal, pointing to response routing rather than missing knowledge.

• CoVA-SFT Supervises Visual Intermediate Representations: Its 51.9K samples contain over 222K multimodal reasoning steps. It more than doubles other interleaved chain-of-thought baselines but still trails a strong text-only baseline.

Also Notable

• Use Explicit Rules for Emotion Inference in Long Dialogues — EmoLASP combines language models with answer set programming. It explores a modular route with less dependence on long prompts and frequent fine-tuning.
• Anchor Low-Resource Speech Decoding With Translation Semantics — SAMA-ASR combines semantic and acoustic information through lightweight adaptation. This helps compensate for scarce transcription supervision.
• Search for Prompts in Natural Language — CoEvoAD selects cross-category transfer objectives to improve generalization in industrial zero-shot anomaly detection.
• Diagnose Local Prosody Before Fixing TTS — A screen-diagnose-correct pipeline targets stress, pause, and intonation errors that aggregate metrics often miss.
• Map Text Steganography Across 148 Methods — This survey covers linguistic steganography, detection, evaluation, and open challenges. It offers a structured entry point for generative steganography research.
• Remove Satellite Trails From Adjacent Frames — STARLINC uses correlations between neighboring frames. It turns hard-to-scale manual contamination screening in astronomy surveys into a workable data pipeline.
• Separate User Identity From Expressed Opinions — The study tests whether personalized models understand political positions or merely mirror personal details disclosed by users.
• Control Audience-Specific Explanations Through Internal Representations — Attribute-related activations offer another control interface when prompts cannot produce stable personalization.

Read the full edition →