• Recursive Feedback Improves Paper-to-Poster Workflows. After training DesignHarness, all seven code-agent and model setups improved. The average score rose from 54.99 to 67.39, though current evidence only covers paper-to-poster tasks.

• Evoke Moves Persistent World State Outside the Generator. An external state store, long-horizon teacher supervision, and a three-step student model combine consistency with lower generation latency.

• Compound-Option Errors May Come From Composition, Not Missing Knowledge. The model judges atomic claims while integer programming enforces AND, OR, and NOR constraints. Macro-F1 rose by about 29 points on both datasets.

• Intern-S2-Preview Unifies the Scientific Agent Training Pipeline. A 397B backbone combines multimodal training with a 4B memory decoder for long task chains and lower-cost domain adaptation.

Also Notable

• Make Agent Players Pursue Long-Term Goals to Test Whether World Models Stay Consistent and Follow Actions. PlayWorld shifts evaluation from short-video quality to long-horizon goal completion.
• Delay Long-Term Memory Consolidation at the Semantic-Segment Level. LycheeMemory V2 aims to reduce continuous consolidation costs while preserving fine-grained evidence for retrieval.
• Extend AI Scientists to Cross-Disciplinary, Multimodal Evidence Access. OmniScientist's real test is whether unified capabilities reach the underlying research evidence.
• Estimate Contact and 3D Forces From One Egocentric RGB Image and Object Geometry. EgoPHI advances hand-object understanding from contact localization to physical interaction reasoning.
• Explain How Traffic Events Happened, Why They Happened, and What Comes Next. UniTraffic-Agent uses two out-of-domain evaluations to test transfer beyond fixed cities and datasets.
• Coordinate Infrared Heat and Visible-Light Texture With Progressive Fusion and Dual-Prior Distillation. P2Fusion uses prompt-driven fusion to prevent static priors from conflicting.
• Preserve Spatial Evidence as Well as Diversity When Pruning 3D VLM Tokens. CoverPrune uses optimal transport to select tokens for scene coverage.
• Capture Physical Defects Such as Unstable Support and Contact Errors in Human Motion Tracking. HumanTracker targets flaws that frame-level pose errors often miss but people readily notice.

Read the full edition →