Reward-guided topology design cuts multi-agent communication tokens 20.5% at equal accuracy.

Same source domain / Shared topic / Tension

Memory system serving cost is unpredictable from conversation length by 18-69%.

Reported by the same outlet (arxiv.org); overlapping topics (accuracy, against, cost, dominat); pushes against this story (against).

Same source domain / Shared topic / Tension / Downstream implication

Test your RAG under corpus mutation, not against a static snapshot.

Reported by the same outlet (arxiv.org); overlapping topics (against, correctness); pushes against this story (against).

Same source domain / Shared topic / Tension

Let the model pick its reasoning mode at the first [redacted] fewer tokens for 1.4 points of accuracy.

Reported by the same outlet (arxiv.org); overlapping topics (accuracy, against, design); pushes against this story (against).

Best models hit 70.4% on single-hop API calls and 2.4% on knowing a question is unanswerable.

Reported by the same outlet (arxiv.org); overlapping topics (accuracy, against, agent); pushes against this story (against).

Continuity Kernel model-checked a commit protocol for agent memory across 2.8M states.

Reported by the same outlet (arxiv.org); overlapping topics (against, agent, design); pushes against this story (against).

Test your agent against messy requirement histories, not the consolidated ticket.

Reported by the same outlet (arxiv.org); overlapping topics (accuracy, against, agent); pushes against this story (against).

Early cache compaction costs AI agents accuracy

Reported by the same outlet (arxiv.org); overlapping topics (accuracy, against, agent); pushes against this story (against).

SWE-NFI grades AI refactors on structure, not tests alone

Reported by the same outlet (arxiv.org); overlapping topics (against, agent, correctness); pushes against this story (against).