• Open-Source 32B Reaches Top Tier for Hardware Code Debugging. InCoder distills reasoning chains from engineers' actual error-fix cycles. It ranks among the best open-source models on LiveCodeBench and CAD-Coder, though KernelBench at 38% shows GPU optimization is still far from production-ready.

• CLIP's Spatial Blindness Is Baked Into Its Training Objective. CoME-VL fuses CLIP with DINO at the representation level, lifting grounding tasks by 5.4%. The real value: systematic ablation data for anyone evaluating dual-encoder designs.

• Agents That "Got It Right" May Just Be Guessing. Agentic-MME evaluates multimodal agents on process, not just final answers. The strongest model manages only 23% on hard tasks. An overthinking metric exposes step-efficiency gaps hidden by accuracy scores.

• RAG Failures Are Multi-Dimensional; a Single Accuracy Number Can't Find the Bottleneck. This AAAI paper splits diagnosis into four axes: reasoning complexity, retrieval difficulty, document structure, and explainability. It moves teams from blanket tuning to targeted fixes.

Also Notable

• Computer-Use Agents Have Fundamentally Different Safety Failure Modes Than Chat — persistent state and cross-step side effects introduce new evaluation dimensions.
• Open-Vocabulary Detection Can Drop the Text Encoder at Inference — DeCo-DETR decouples visual-text cognition paths. ICLR accepted.
• GNN Surrogate Models Move Into Operational Flood Forecasting — NVIDIA team focuses on the speed-accuracy engineering trade-off.
• First Multi-Sensor Foundation Model for Mars Remote Sensing — uses model merging to integrate three sensor modalities at different resolutions.
• Lightweight Plug-and-Play Module Fixes Model Drift in Multi-Frame Tracking — CVPR accepted, a practical improvement for visual tracking pipelines.
• Membership Inference Attacks May Fail Under Adversarial Inputs — the "honest query" assumption in existing MIAs may be too optimistic.
• Probabilistic 3D Ocean Dynamics From Sparse Satellite Observations — Google team's depth-aware generative approach.
• Recognizing Unseen Defect Types in Industrial Inspection — visual prompting approach, CVPR accepted.
• Text-to-Physically-Plausible Hand-Object Interaction Meshes — targeting dexterous grasping and VR content generation.
• Millimeter-Wave Signals for High-Fidelity 3D Scene Imaging — a potential alternative to cameras and LiDAR in adverse weather.

Read the full edition →