• Visual Reasoning Becomes Trainable and Verifiable. VBVR-Pro compares image, video, and interleaved reasoning across 300 procedural tasks with task-specific graders.

• Document Retrieval Can Route Each Query Differently. RetrievalRouter improves accuracy by 2.5% over the best static baseline while running 12.4× faster.

• Code Can Carry World Rules and State Transitions. Code World Model separates executable dynamics from visual rendering, though current evidence covers only simple interactive settings.

Also Notable

• Targeted Bit Flips Can Trap MoE Models in Endless Generation — Groundhog Bit-Flip Attack exposes an exploitable attack surface between hardware faults and model routing..
• Generative VLM Debiasing Is Moving Into Inference — GGSS studies generation under changes in demographic attributes, rather than only classification representations..
• Parameter Distance Cannot Predict Rapid Relearning — The alignment gap between forgotten targets and retained abilities may be a more useful measure..
• Long-Narrative RAG Suffers From Evidence Islands — PonsRAG coordinates scattered clues and broken connections across narrative layers instead of simply expanding context..
• First-Person Assistants Must Resolve Conflicting Evidence — EgoArgus tests what models should believe when visual evidence conflicts with user claims, and what support they should provide..
• Multi-Object Images Make Test-Time Adaptation Noisier — This work extends adaptation from single-label recognition to realistic multilabel settings..
• Hidden States Can Reveal Novelty Judgments — Think-Probe-Respond probes internal states and calibrates final answers to reduce the model's bias toward “moderately novel.”.
• Shared Pinyin and Glyph Encoding Can Correct Chinese Queries — GUIDE targets search systems where new terms change quickly and labeled correction pairs are hard to maintain..
• More Camera Views Do Not Guarantee Better Action Assessment — Fusing redundant perception may matter more than stacking first- and third-person views..
• Inference-Time Sampling May Improve Protein Language Models — Generation limits depend on sequence and structure sampling, not only model training..

Read the full edition →