• Skill Entropy Lifts Qwen3-4B From 34.4% to 68.4%: Long-horizon tasks depend on individual skills and reliable handoffs between them.

• Earlier Multimodal Unification Works Better: Controlled experiments suggest delayed fusion causes “visual laziness,” making evidence-based training decisions essential.

• Models Over-Infer 35%–49% of User Profile Statements: Self-monitoring can also fail, so personalization systems should track memories and their evidence separately.

Also Notable

• OneDayAgent Unifies Goal Drift, State Loss, and Context Overflow — It tests whether agent infrastructure works consistently across model backends.
• HelloWorld Makes Video Characters Respond Socially to Users — Interaction expands from the generated scene to characters inside it.
• FocusMem Splits GUI Memory Into Content, Retrieval, and Confidence — Fixed latent memory blocks no longer carry every decision requirement alone.
• ABSeeker Traces Useful Search Steps Backward From Final Answers — Long-chain retrieval training can distinguish correct actions from redundant and incorrect ones.
• ToolArtist Coordinates Reasoning, External Tools, and Image Generation — It explores open-world image tasks without predefined workflows.
• NOLLI Uses Calibrated Bilingual Riddles to Isolate Language Gaps — Reproducible questions reduce interference from mismatched difficulty across English and Korean.
• K-EXAONE 2.0 Expands an Existing Model to 750B Parameters — Its mixture-of-experts design activates about 37B parameters per token, offering a scaling path without training from scratch.
• ContextMaster Routes Information Across Multi-Shot Video Workflows — One framework supports generation, reference following, and editing within a fixed budget.

Read the full edition →