• TRM Defines Rubrics Before Assigning Fine-Grained Scores. PD-GRPO reduces the score polarization caused by conventional pairwise preference optimization, giving visual generation RL more specific reward signals.

• SAKI Switches Supervision Based on Teacher-Student Disagreement. It preserves reverse KL for accepted positions and directly learns the teacher’s preferred token at correction points. Generation throughput reaches 4.22× on matching workloads.

• GEB Tracks Long-Video Memory Through Entity Identity. It improves results across four day-long to week-long video benchmarks. EgoLifeQA accuracy reaches 72.0%, beating the previous best by 4.4 percentage points.

Also Notable

• Complementary Trajectories Can Repair All-Failure Training Groups. GRAFT replaces all-failure groups, then controls cross-model bias through sequence-compatibility weighting and token-level importance-ratio clipping..
• Fuse Multi-Level Visual Detail Without Expanding Latents. HiRAE anchors on deep representations and adds residual information from each layer under a norm budget, improving reconstruction and text-to-image alignment..
• Validate Actions Before Teaching the Robot. Real2Gym verifies demonstrations or retargeted actions through native physics execution. It converts successes and failures into procedures, relative object motions, and recovery strategies without updating base-model weights..
• Explore Where Models Might Change Their Minds. HDL identifies branch points through token-likelihood changes caused by hindsight feedback. It samples alternative continuations only from those positions, focusing exploration while generating fewer tokens..
• Agents Can Act Before Finishing Their Reasoning. ActFirst-OPD separates environment actions from full reasoning generation. It advances the interaction first, then produces thought-action responses for teacher supervision asynchronously..
• Distill Each Video Capability Once Per Backbone Family. LongLive-Plug distills single-pass CFG, few-step sampling, and long-context correction into reusable LoRAs for compatible downstream models..
• Geometry Code Can Scale Spatial Query Templates. Exemplar2VQA lets collaborating agents call geometry tools, turning varied static and object-centric query templates into large synthetic 3D question-answering datasets..

Read the full edition →