• Looping for deeper compute saturates at two passes. LoopCoder-v2 turns "how many loops" into an engineering knob, and two loops push SWE-bench Verified from 43.0 to 64.4 — but three or more regress.

• Make the agent ship a game you can actually play. GameCraft-Bench sets 140 tasks inside the Godot engine, scoring by running the game and replaying it rather than reading code.

• Put the teacher in the prompt, not the gradient. ZPPO uses the teacher as a hint to rescue all-fail problems in small-model post-training, beating distillation and GRPO across four sizes from 0.8B to 9B.

• One shared discrete tokenizer can work after all. UniAR bets against UniDDT, using bitwise quantization so understanding and generation share visual tokens, hitting dual SOTA on image generation and editing.

Also Notable

• Self-Evolving Agents Should Learn How to Use Experience, Not Just Store It — OPD-Evolver proposes a slow-fast on-policy self-distillation framework that separates "remember trajectories" from "learn to evolve."
• Looped Architectures Move Into World Models for the First Time — Looped World Models reuse depth to ease the tension between long-horizon simulation needing deep compute and deep models being expensive and error-prone.
• Most Transformers Make Every Layer the Same Width — This One Doesn't — Variable-Width Transformers use an ×-shaped structure to reshape width and compute along depth.
• Interactive World Models Have Shrunk Their Action Vocabulary to Navigation — ActWorld uses action-aware memory so you can actually pick up a plate or open a door, not just move the camera.
• Non-Differentiable Top-K Routing Has Long Plagued MoE Training — SoftMoE switches to soft differentiable routing so expert selection can be learned end-to-end (ICML).
• Reasoning Models Keep Grinding After the Answer Is Out — Dynamically pruning rollouts from a GRPO view cuts overthinking and the dead reasoning after the answer appears (Huawei).
• The "Dense Fourier Spectrum" of Transformers Doing Modular Multiplication Is an Artifact — Re-analyzed, it's a discrete logarithmic clock (Stanford).
• When Image and Text Conflict, MLLMs Always Side With Text — The root cause is a late-layer text bias, and it can be corrected directly (IJCAI).
• LLM Code Translation Fixates on Correctness and Ignores Runtime Efficiency — In the post-Moore era this starts to matter (ICML).
• Unsupervised Dense Retrieval Catches Semantic Similarity but Misses Temporal Relevance — Temporal preference optimization fills the gap (ICML).

Read the full edition →