Clarification Is Not Correction: LLMs Fail to Let Go
Jianzhe Lin and colleagues at Meta AI argue that many multi-turn failures come from early commitment rather than forgetting: an ambiguous first turn becomes a fixed task state that later clarification only patches.
Ask this paper
Order changes outcomes. With equivalent final information, ambiguous-first dialogues lower success from 50.0% to 42.8% for Gemini-2.5-Pro and from 55.4% to 45.7% for Gemini-2.5-Flash compared with clarified-first.
Coding is most affected. In ambiguous-first runs coding reaches 24.4% success with contamination 0.256, about double writing (69.3%, 0.125), because early assumptions end up in interfaces and control flow.
Acknowledging is not revising. In coding, models acknowledge a correction 93.5% of the time but revise the artifact 86.4% of the time.
Common memory fixes do not help. Summary memory raises success but also raises contamination (0.228 vs 0.177 for raw history); 62.5% of summaries collapse the ambiguity. A single-state ledger is worst at 0.295.
Rebuild instead of patch. Explicit rollback lifts success from 17.5% to 27.5% and removes measured contamination; ask-first and two-phase policies drop wrong commitments from 40% to 0%. Setup: 290 generated tasks, about 7,160 trials.
Abstract
Dialogue failures in language models are usually framed as memory failures: context too long, summaries lossy, a constraint forgotten. We argue this misses a deeper problem: in many conversations the model does not forget, it commits too early. An ambiguous early turn collapses into a single hidden interpretation, and later clarification is filtered through that commitment. We call this early posterior collapse: unresolved user intent collapsing into a committed task state before ambiguity is resolved. We study it with controlled dialogue tasks in writing, planning, and coding using Gemini-2.5-Pro and Gemini-2.5-Flash. Across thousands of trials, the same information in different orders yields different outcomes, even when the final dialogue contains equivalent task-relevant information. This order effect suggests later clarification is treated as extra context rather than a corrective signal: it refines a stale task state without invalidating it. Coding tasks are especially vulnerable, suggesting early assumptions get embedded in structured artifacts such as interfaces and control flow. Standard prompting and memory strategies do not reliably help: summaries can collapse ambiguity, and chain-of-thought can reduce explicit wrong commitment in reasoning traces without improving final task success. These findings motivate uncertainty-preserving state management. If assistants cannot let go of early interpretations, robustness cannot rely on post hoc correction alone; it must keep ambiguous early turns from hardening into one task state. Assistants should hold tentative hypotheses while ambiguity remains, ask before executing when high-impact ambiguity persists, and rebuild from a revised state when later evidence invalidates an earlier reading. Rather than one prompting fix, we aim to redirect research for interactive LLMs from retaining more context toward preserving uncertainty.