LLM interpretation errors

Type: index · Status: current

LLM output deviates from what the user intended for three distinct reasons, each a property of a different part of the system and each requiring different remedies:

  • Underspecification — the prompt admits multiple valid interpretations. A property of the specification language. Even a perfect interpreter faces this. Remedy: narrow the spec.
  • Indeterminism — the same prompt produces different outputs across runs. A property of the sampling process. Theoretically eliminable. Remedy: sampling control.
  • Interpretation error — the LLM's output distribution is biased away from the valid space (for a theoretical deterministic LLM: simply the wrong output). Remedy: error detection and correction.

Ma et al.'s prompt stability study empirically separates all three: temperature+sampling measures indeterminism within each prompt variant, cross-variant comparison measures underspecification, and systematic degradation under emotional prompts reveals bias. Performance and stability are decoupled (Spearman rho = -0.433), confirming these are independent phenomena.

Conflating the three leads to misdiagnosis — e.g. narrowing the spec (underspecification remedy) when the LLM is ignoring constraints it already has (interpretation error), or lowering temperature (indeterminism remedy) when the spec genuinely admits the unwanted output (underspecification). This area covers the taxonomy, the detection and correction machinery (oracles, voting, verification), and architectural responses (separation, bounded context) for managing all three.

Error Correction Theory

Oracle Theory

Aggregation & Correction

  • synthesis-is-not-error-correction — merging agent outputs propagates errors; voting discards minorities and corrects them; the aggregation operation must match the decomposition structure

Architectural Responses

Sources

  • learning-theory — oracle and verification theory originated there; this area applies it specifically to LLM interpretation errors
  • computational-model — the scheduling architecture that separation notes describe; error correction explains why it works

Other tagged notes