• Agent Capability Is Becoming Sustained Delivery: Apodex 1.1 builds state retention, failure recovery, multi-agent coordination, and result verification into complete work trajectories.

• Complex Retrieval Amplifies Speech Recognition Errors: Entity graphs and multi-turn rewriting widen F1 declines by 36%–67%, with incorrect entities driving most failures.

• Environment-Side Regularization Frees the Exploration Budget: ERPO constrains query distributions with Query-KL, letting PPO and GRPO tune training stability separately from response exploration.

• A Few Global Attention Heads Can Handle Long-Term Memory: ReWorld uses local heads for real-time control, plus a fixed KV cache and landmark store for persistent spatial state.

Also Notable

• Persistent REPLs Turn Agent Improvement Into a Cumulative System. Prime Agent preserves history, memory, and skills across long-running trajectories through an open-source RLM runtime.
• RAG Answers Can Flip Even When Aggregate Accuracy Stays Flat. Corpus growth can silently change answers to individual questions, so production audits should track sample-level answer churn.
• Evicted Audio KV States Should Not Be Gone Forever. WnW lets speech LLMs restore historical positions on demand during decoding instead of relying on static compression.
• Step-Level Tests Expose Failures in Rule Execution. A new benchmark checks whether models can reliably apply reusable external rules instead of merely recalling facts.
• Static Math Sets Lose Signal as Preference Optimization Improves. DIAG diagnoses current weaknesses and generates samples near the model’s ability boundary, improving data efficiency for preference distillation.
• Motion Paths Beat Fixed Scan Orders for Video Interpolation. A state-space model follows motion trajectories to capture long-range dependencies between frames with linear complexity.
• Entropy-Guided Cropping Preserves Objects Across Patch Boundaries. ENCORE constrains attention to protect complete objects in lightweight vision-language models.
• The Moderator May Be the Bottleneck in Multi-Agent Debate. Meta-Moderator detects redundant discussion, manages the budget, and decides when the debate should stop.
• Cultural Knowledge Does Not Guarantee Visual Grounding. A Southeast Asian video benchmark separates cultural understanding, visual recognition, and temporal localization.
• LoRA Attention Patterns Reveal Where Ranking Behavior Forms. The study traces task-related correlation patterns through the network to support adapter diagnosis and pruning.
• Self-Reflection Can Become a Credit-Assignment Signal. SRPO uses reflection during post-training to connect sparse final rewards with specific decisions in long reasoning traces.

Read the full edition →