• Privileged Information Can Make Self-Distillation Worse: DAPD constrains irreproducible teacher behavior with dual-path and dual-source anchoring, averaging a two-point gain across Qwen3-4B tasks.

• Early Access to Future Trajectories Pollutes Driving Supervision: Deferred Exposure makes the model select a candidate trajectory before using the actual trajectory for verification.

• Mature Autoregressive Models Can Move to Diffusion Cheaply: DiffusionGemma uses less than 10% of the original model’s training-token budget and reaches about 1,500 tokens/s on one H100.

• Semantic RGBA Layers Keep Generated Designs Editable: UniWorld-Design delivers objects as separate assets and cuts per-layer RGB error by 37% on Crello.

Also Notable

• JoyAI-Video-Edit Recasts Real-Time Video Editing as Causal Generation — It edits long videos without future frames or a preset duration while preserving low latency, source fidelity, and temporal consistency.
• AURORA-LM Stops Overcompressing Text Latents for Diffusion — It jointly learns continuous language representations that support both generation and exact decoding.
• AgentStream Tests Agents Against Changing Task Streams — Its streaming evaluation asks whether experience accumulated on isolated tasks still transfers.
• OmniPack Preserves Evidence Under Tiny Token Budgets — Its query-aware compressor decides which audio and video evidence cannot be discarded.
• TurnSight Assigns Turn-Level Credit With Hindsight — It replaces a single outcome reward with finer training signals across long tool-use chains.
• PCSD Judges Teacher Signals Through Cross-Position Consistency — It avoids frequently switching supervision sources because of accidental disagreement on one token.
• DocNavRAG Builds Evidence Along Document Structure — It offers a stateful alternative between fixed graph traversal and unstructured tool calls.
• RING Injects Retrieved Knowledge Into Generative Models — It trades training and architecture costs for lower serving latency and less system complexity.
• Principles of Robot Autonomy Maps the Full Autonomy Stack — This deployment-focused survey offers a practical entry point to established methods and tools.
• Letter Case May Act Like a Visual Salience Cue — Attention Is Case-Sensitive suggests that prompt formatting itself can change model attention.
• SEER Uses Local Evidence to Avoid Instance Mix-Ups — Spatial reasoning can fail because the model selects the wrong referent, not because it misses the object.

Read the full edition →