Early Posterior Collapse: Gemini 2.5 Treats Later Clarifications as Context, Not Corrections, and Coding Tasks Suffer Most

arXiv·medium signalXBlueskyLinkedInCopy link

Across thousands of controlled writing, planning and coding dialogues with Gemini-2.5-Pro and Flash, the same information in a different order produced different outcomes. The model commits early to one reading of an ambiguous turn, and later clarification only refines that stale task state. Coding is especially vulnerable because early assumptions get baked into interfaces and control flow, and summaries and chain-of-thought did not reliably fix it. The practical takeaway is to resolve ambiguity before the agent writes structure, or restart the context.