• Dynamic Gating Separates Reasoning Flaws From Ordinary Language. NSD identifies reasoning-critical tokens automatically, then limits negative-training updates to faulty behavior while preserving language priors.

• Page Vectors Can Be Rebuilt on Demand. GLIE stores just four vectors per page while retaining nearly 80% of the uncompressed system’s nDCG@5 on ViDoRe v1.

• NCP Expands Prediction Beyond Individual Tokens. An 8.9B model matches OLMo-3-7B’s final pretraining loss using 51.3% of its training tokens while preserving standard autoregressive generation.

• T1 Replays Sampled Tokens and MoE Routing Exactly. TITO with per-token expert replay cuts the train-inference log-probability gap from 0.021 to 0.013. Terminal-Bench 2.1 success rises from 43.8% to 64.0%.

Also Notable

• Separate Periodic Grids From Content-Bound Grain Before Repairing Either. Mi-Ripple removes traces of repeated AI edits while trying to preserve the source image’s structure.
• An 8B MoT Unifies Vision Without a Separate Encoder or VAE. SenseNova-U1.5 covers visual understanding, reasoning, and generation through one native interface.
• A General Hierarchical Transformer Replaces Optical-Flow Components. FreeFlow removes correlation volumes, feature warping, and iterative refinement to test a simpler feed-forward pipeline.
• Scene Programs and Solvers Recursively Build Complex 3D Worlds. RCWM processes compositional structures from one reference image instead of generating an entire scene program at once.
• An Open Math-Olympiad Gold Recipe Skips Formal Provers. The system combines specialized checkpoints, verification, and answer refinement in one test-time compute pipeline.
• Shared Anatomy Supports General Medical Image Restoration. UniH³ jointly models modality differences, degradation types, and structural commonalities to simplify training and improve generalization.
• Role Graphs Organize Multi-Agent Responsibilities and Relationships. DRG-MAPPO provides hierarchical decision representations for autonomous systems that require sustained tactical coordination.
• Solver Approval Can Still Hide Semantic Drift. A generative reward model detects formalizations whose conclusions pass despite unfaithful translations.

Read the full edition →