When a request changes, what should an AI keep?
DHAIM studies how AI systems handle changing user goals in multi-turn dialogue. The question is not only whether a system follows the latest instruction, but whether it updates the right parts of the task while preserving what remains valid.
Early-stage research · No benchmark or dataset releasedChanging user/task state
We investigate whether a system applies a correction or new target, preserves earlier requirements that still hold, and propagates changes through relevant dependencies. A response may satisfy the new target while losing another part of the active task.
Traceable research integration
The project is developing a way to keep claims linked to their sources, evidence strength, scope, conflicts, and revision history. Whether that structure offers a measurable advantage over a strong simpler approach is still an open question.
What this page does—and does not—claim
This is a public introduction, not a finished theory, a model ranking, or a validated evaluation set. A recent internal development screen passed a technical provenance check; scientific interpretation and any result release require separate review. No empirical data or model outputs are published here.