Stack layer / Contrast
Two-gap framework recasts reward hacking and hallucination as symptoms of requirement and model gaps
arXiv
Stack layer / Contrast
Reflexion-Style Verbal Memory Sometimes Lowers Success Versus Plain Retry, and Replay Experiments Show Why
arXiv 2609.12404
Stack layer / Threat pattern
787,562 Function Pairs Show AI Code Is Half the Size of Human Code With Different Defect Classes, Not Fewer
arXiv 2609.12708
Stack layer / Threat pattern
Agent Frameworks Detect Dangerous Plan Steps and Then Execute Them Anyway; Fewer Than 20 Lines Closes the Gap
arXiv 2609.15293
Stack layer / Contrast
Splitting Decode by Attention Type Instead of by Operator Buys 31-56% More Tokens Per Joule
arXiv 2609.13134
Stack layer / Update thread
Splitting a CTF Task Into Isolated Sub-Contexts Lets a Local gemma-4 Solve 18.52% of Challenges Standard Agent Loops Fail
arXiv 2609.12839
Stack layer / Contrast
Distilled Byte Models Match a Token Model's Accuracy on One-Sixth the Data and Cut Logit Storage to a Fifth
arXiv 2609.12303
Policy dependency / Stack layer
RIPPLE: an edit confined to one prompt-policy segment changes downstream behavior, so replay candidate edits after previously accepted ones before persisting
arXiv 2609.12127