Early posterior collapse: Gemini 2.5 treats later clarifications as context, not corrections.

arXiv 2609.25337 ran thousands of controlled writing, planning and coding dialogues and found the same information in a different order produces different outcomes. The model commits early to one reading of an ambiguous turn, and later clarification only refines the stale task state. Coding is worst hit, because early assumptions get baked into interfaces and control flow. Summaries and chain-of-thought didn't reliably fix it. Resolve ambiguity before the agent writes structure, or restart the context.