• A Terminal-Only Agent Matches Fully Equipped MCP Setups. 72 HF upvotes confirm practitioners' collective anxiety about agent over-engineering is real — but whether the benchmark tasks cover true enterprise complexity still deserves scrutiny.

• On-Demand Reasoning Tokens During Code Generation Hit SOTA Across Four Benchmarks. Think-Anywhere triggers reasoning at high-entropy positions, matching how complexity actually unfolds when you write code.

• Three-Layer Agent Collaboration Turns Hours of Footage Into Music-Synced Short Videos. Understanding and editing existing material delivers far more practical value to creators than text-to-video generation.

• Image Generation Shifts From "Memorize Everything" to "Retrieve on Demand." Unify-Agent uses an agentic pipeline to break through the knowledge ceiling on long-tail concepts, approaching top closed-source models after training on 143K trajectories.

Also Notable

• MCTS-Driven Literature Exploration and Idea Co-Evolution — Research ideation moves from static retrieval to dynamic search trees.
• Town-Scale 3D Scenes From a Single Image — No training required; extends object-centric model latent spaces via composition.
• Diffusion Models Generate Synthetic Training Data in RAW Domain — Tackles the long-standing data scarcity bottleneck for low-level vision on camera RAW.
• Privacy Sensitivity Judgments Distilled From 675B to Lightweight Models — Targeting privacy compliance assessment at scale for large text corpora.
• Semantic-Geometric Joint Pruning for 3D QA Visual Tokens — Multi-view tokens are massively redundant; joint pruning delivers major speedups under token budgets.
• Panoramic Video-Driven Controllable Long-Range Scene Exploration — Exploits panoramic footage's natural full-scene coverage for long-range generation.
• Structured Intermediate Representations Before Reasoning for Long-Document QA — More stable than end-to-end generation (ICLR).
• Vector-Granularity Sparse Attention — Finer-grained compute reduction than existing coarse attention patterns for long-context video Transformers.
• All TTS Conditioning Paths Replaced With SSM — Fully removes attention and RNN layers at inference (ICLR).
• Are Multimodal Models Fusing Cross-Modal Information or Exploiting Unimodal Priors? — Information decomposition provides a quantitative answer (ICLR).

Read the full edition →