• Leaderboard Rank Doesn't Predict Deployment. A position paper rebuilds one MCP agent benchmark 14 ways and finds aggregate scores rank unstably out of distribution; what you should measure for production agents is predictive validity.

• Test-Time Reasoning Is a Budget, Not a Switch. SEVRA decides per query whether to keep the initial answer or trigger verification, but the authors admit tuning the initial solve budget often saves more compute than fixing it afterward.

• Short on Training Data? Mine Community LoRAs. FreeStyle treats off-the-shelf style and content LoRAs as composable anchors to mass-produce triplets, turning the community's huge LoRA stockpile into a data source.

• Self-Improvement Moves to a Real Robot Arm. ENPIRE gives robots a repeatable physical feedback loop, letting a coding agent train dexterous tasks like pin insertion and zip-tie fastening to 99% success on its own.

Also Notable

• Camera, Object, and Weather Control Unified Into One Video World Model — Holo-World makes environment state controllably variable, and weather no longer depends on existing source video.
• Project-Level Code Engineering on a Professional Game Engine — JAMER pairs deterministic evaluation to fill the project-level gap beyond asset generation, gameplay design, and web-game coding.
• Web Services Are Mostly Static Endpoints, Bad at Long Flows With Loops — ToolPro writes an agent's tool-call intent as an executable program for a more flexible interface.
• Internalizing 3D Spatial Perception Into an MLLM — SpatialSV uses task-oriented visual supervision, avoiding both external-tool reasoning overhead and pure latent-feature distillation.
• Universal Multimodal Retrieval's Contrastive Paradigm Has a Granularity Blind Spot — ELVA switches to ranking-driven training to cover it.
• Structured Mamba Variants Break the Moment You Reduce Tokens — The root cause is that existing reduction ignores space; this proposes a spatially aware reduction framework.
• Pixel-Level Residual Diffusion Transformer for Scalable 3D CT Volume Generation — It targets the compute and optimization difficulty of high-resolution medical voxels.
• Camera-to-LiDAR Hierarchical Distillation Using Vision Foundation Models — HilDA eases the chronic scarcity of labeled LiDAR data in autonomous driving.
• Offline RL Rewards Often Go Missing Due to Sparse Logging — This work tackles off-policy evaluation under missing-not-at-random (MNAR) rewards.
• Triangular Consistency, Composing Two Flow Segments Into a Third, as a First Principle of Optical Flow — It's independent of architecture, supervision type, and dataset, working for single pairs and multi-frame alike.

Read the full edition →