• Production Traffic Should Set Post-Training Priorities: One self-hosted model now handles 116 million monthly requests after training against demand from over 200 internal applications.

• Action-Free Video Can Sharply Improve Robot Generalization: Training on 120,000 hours of egocentric video raised zero-shot success from 36.1% to 77.8%.

• Pixel-Based Text Encoding Needs Four Coordinated Choices: Adaptive resolution, real image-text pairs, preserved layouts, and multilingual training sustain performance after compressing visual tokens by 80%.

• Mid-Training Distillation Splits Reasoning From Recall: Switching objectives based on teacher entropy delivers 1.61–1.71× stronger reasoning while retaining 96.7%–96.8% of factual memory.

• GUI Agent Competition Now Includes Infrastructure: Supporting phones, browsers, and desktops requires coordinated investment in environments, task creation, and reward verification.

Also Notable

• Make Safety State Part of Native Memory: Safin-1 studies whether safety state can accumulate and adapt across long interactions instead of relying only on external guardrails.
• Extend Pretrained VLMs Into Driving: Qwen-Drive-1.0 tests whether one visual representation can support 3D perception, question answering, and motion planning.
• Let Error Memory Choose the Next Lesson: DiagEvo creates problems around unresolved weaknesses instead of selecting training data only for difficulty and diversity.
• Jointly Train Query and Product Generators: Generative components move beyond query expansion and into candidate matching itself.
• Train on Histories Split by Context Compression: MemoryWalker drops the assumption that production agents always see complete, linear execution traces.
• Teach Streaming QA When to Search Deeper: StreamScout allocates retrieval cost by question instead of repeatedly scanning limited memory.
• Test Agents on Failed Power-Flow Solves: RestoreBench measures whether systems can interpret intermediate results, call tools, and repair tightly constrained engineering failures.
• Generate PDE-Constrained Designs With Conditional Flow Matching: The method aims to reduce expensive iterative solving and sensitivity to initialization.
• Retrieve Ideas for Their Future Influence: MUSES challenges retrieval systems that favor popular, central papers over earlier ideas that may shape later research.
• Intervention Does Not Prove Causal Use: Steering a model along one direction does not mean the model originally relied on that direction when contextual and parametric memories conflicted.
• Separate Geometry and Appearance Priors for 3D Repair: DualDiff3D addresses missing geometry and novel-view artifacts with distinct diffusion priors.

Read the full edition →