• Extra Sampling Steps Don't Guarantee Deeper Reasoning. Under specific conditions, constant-round parallel computation can simulate standard flow-matching trajectories. Self-conditioning lets later steps revise earlier estimates, making additional sampling useful.

• Generated Output Must Pass Checks Before It Changes World State. ATAS manages static scenes and time-indexed events separately. Its orchestration layer improved every reported cross-shot consistency metric across four generators.

• COUNT Moves Corpus Statistics Into Replaceable Host-Memory Tables. For a one-billion-parameter model, validation-loss gains matched up to 57% more compute. Swapping in a domain-specific table also enabled training-free adaptation.

Also Notable

• Two Axes Explain Multimodal Agent Memory — The taxonomy tracks information formation and representation alongside changes in memory management. Its three levels reuse fixed memory, update memory under fixed policies, or update both memory and policy.
• 16 Synthetic Game Environments Push Agents Toward Transferable Control — Player-1 starts from 6,044 demonstrations, then learns compound actions through verifiable rewards, randomized conditions, and PPO.
• Teachers Should Stop Following a Student's Wrong Turn — CriticOPD has the teacher critique an incorrect student rollout, then lets the student rewrite its solution after the error, restoring recoverability for on-policy distillation.
• Two Shared-Parameter Passes Need Unequal Compute — LatentLoop compresses the first pass, preserves full resolution for the second, and spends the savings on wider hidden states. It reduced validation NLL across all 15 configurations.
• Cortex Turns Context-Gained Capabilities Into Mergeable Updates — Low-rank or feed-forward updates absorb a target capability while measuring retention of unrelated capabilities.
• ActionOPD Distills Decisions, Not Tokens or Reasoning Traces — It compares actions in states the student actually encounters. Teacher and student need not share a model family or tokenizer.
• Finding Multi-Agent Conflicts Is Only the First Step — The 600 verifiable tasks also require systems to repair state, release shared locks, and complete handoffs.
• Verifier Diagnostics Can Become Layered Training Rewards — An integrity audit rejects altered specifications and invented assumptions, turning formal verification from an inference tool into a training signal.
• Query-Key Channel-Pair Norms Can Guide Precision — PN-Mix averages 2.26 bits for the key cache while matching the mean accuracy of existing methods using 2.32 to 2.39 bits.
• One-Step Generation Can Start With Reconstruction — Reconstruction followed by distribution post-training cuts training compute to roughly one-tenth at comparable quality. It reached a state-of-the-art FID of 1.19 among one-step generators on ImageNet 256×256.

Read the full edition →