Clarification Irrelevant

Goal Conversation Quality Frequency Common Category Agent Interaction Published View source on GitHub ↗

Issue

The agent correctly recognizes that a request is ambiguous and asks a clarifying question, but the question it asks targets the wrong axis of ambiguity — it doesn’t actually narrow down the interpretation that matters. The user answers the question, the agent proceeds, and the output is still wrong because the real ambiguity was never resolved. This is distinct from over-clarification (asking when nothing needed clarifying) and under-clarification (not asking at all): here the agent’s instinct to ask was correct, but its question-selection logic picked a low-information question over the high-information one.

Frequency: Common

Symptoms

  • User answers the clarifying question, agent proceeds, output is still corrected on a different dimension
  • Clarifying question addresses a minor stylistic detail while the actual blocking ambiguity (scope, target system, time range) goes unasked
  • Multiple rounds of Q&A occur before the agent stumbles onto the question that actually mattered
  • Users report the clarifying question felt “beside the point” or unrelated to their confusion
  • Time-to-resolution for ambiguous requests is high despite the agent asking questions promptly

Root Cause

Generating a plausible clarifying question is easier for a language model than identifying which source of ambiguity has the highest expected impact on the outcome. Without an explicit ambiguity-ranking step — enumerating candidate interpretations and asking specifically about the split that most changes the output — the model tends to ask about whichever ambiguous term appeared most recently or most saliently in the prompt, which is not necessarily the one that determines correctness. The question-generation step and the interpretation-space analysis are typically the same undifferentiated generation pass, so there’s no mechanism forcing the question to target the highest-variance unknown.

Example

User: "Pull last quarter's numbers for the report."

Ambiguity actually present: which report template (there are two: Board
deck vs. Investor update, with different metrics and different quarter
boundaries), and which entity ("last quarter" could mean calendar Q2 or
fiscal Q1, since the company's fiscal year starts in April).

Agent asks: "Would you like the numbers formatted as a table or as
bullet points?"

User: "Table is fine."

Agent pulls calendar-Q2 numbers into a Board-deck-style table. The user
actually needed fiscal-Q1 numbers for the Investor update — the
formatting question resolved nothing that mattered, and the fiscal-vs-
calendar and template ambiguities were never raised.

User: "This is the wrong quarter and wrong format entirely."

Statistics

FindingContext
An estimated 30-45% of single clarifying questions in ambiguous-request flows fail to resolve the interpretation that actually determined the final correctionEstimated from review of agent Q&A transcripts followed by user corrections
Sessions where the first clarifying question misses the key ambiguity average roughly 1.8 additional Q&A rounds versus 1.0 when it hitsTypical range observed in multi-turn agent logs
Ranking candidate ambiguities by output-impact before asking reduces irrelevant-question rate by around a thirdReported range across teams that added interpretation-space enumeration

Mitigations

  1. Interpretation-space enumeration: Before generating a clarifying question, have the agent explicitly list 2-4 candidate interpretations of the request and identify which unresolved variable would change the output most, then ask about that one.
  2. Impact-ranked question selection: Score candidate clarifying questions by how much they narrow the space of valid outputs, and always ask the highest-impact one first rather than the most fluent-sounding one.
  3. Multi-slot single question: When more than one ambiguity exists, combine them into one structured question (e.g. “Which quarter — calendar Q2 or fiscal Q1 — and which template — Board or Investor?”) rather than resolving them one irrelevant question at a time.
  4. Post-answer re-check: After the user answers, have the agent re-verify whether the original ambiguity is actually resolved before proceeding, rather than assuming any answered question clears it to proceed.
  5. Irrelevant-question review loop: Log cases where output was corrected despite a prior clarifying question having been asked and answered, and use them to retrain or reprompt the question-selection step.

Production Signals

Key Metrics

MetricDescriptionAlert Threshold
post_clarification_correction_rateRate of outputs still corrected after a clarifying question was asked and answeredAlert if > 25%
clarification_rounds_to_resolutionAverage number of Q&A rounds needed before output is acceptedAlert if > 1.5
question_output_impact_scoreAverage estimated output-narrowing impact of clarifying questions askedAlert if trending down

Alerts

AlertConditionSeverityResponse
High post-clarification correction ratepost_clarification_correction_rate exceeds threshold over a rolling windowMediumReview question-selection prompt, add interpretation-ranking step
Repeated multi-round ambiguity resolutionA session requires 3+ clarification rounds for one requestLowFlag for manual review of question quality
  • Under-Clarification - the opposite failure to ask at all; both leave the real ambiguity unresolved, just via different paths
  • Over-Clarification - both involve poorly targeted clarification behavior, one asking when unneeded, the other asking the wrong thing
  • Disambiguation Strategy Ineffective - the broader pattern of a resolution approach that doesn’t narrow ambiguity, of which irrelevant questioning is one specific cause