Research map / Large language models / Language, safety and interpretability

Decomposing Uncertainty for Large Language Models through Input Clarification Ensembling

ICML 2024 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes input clarification ensembling, which decomposes an LLM's total uncertainty into aleatoric (input ambiguity) and epistemic parts by ensembling predictions over LLM-generated clarifications of the input, without modifying model weights.

Problem

Uncertainty decomposition for black-box LLMs: separating total predictive uncertainty into aleatoric (input ambiguity) and epistemic (missing knowledge) components, which BNNs and Deep Ensembles cannot do cheaply for LLMs.

Approach

Symmetric analogue of Deep Ensembles: instead of ensembling models to capture epistemic uncertainty, ensemble predictions over generated input clarifications so that the mutual information between outputs and clarifications approximates aleatoric uncertainty from input ambiguity, while the average entropy under clarified inputs approximates the remaining (epistemic) uncertainty.

Key results

• On AmbigQA, the method reaches 71.7 AUROC (GPT clarifier) and 67.1 (fine-tuned Llama-3-8B clarifier) for ambiguity detection, versus 55.0 for Ask4Conf-d and 54.9 for Semantic Entropy.
• On AmbigInst, the method reaches 81.3 AUROC versus 66.0 for Semantic Entropy and 57.9 for Ask4Conf-d.
• For mistake detection on GSM8K, the method reaches 89.7 AUROC, versus 88.2 for Semantic Entropy and 58.1 for Ask4Conf.

Measured effects reported in the paper

Surprising findings

• Ask4Conf performs well on NQ factual QA (70.4 AUROC) but poorly on GSM8K (58.1 AUROC).
• Deep Ensembles built from varying in-context examples yield aleatoric uncertainty with almost no gap between ambiguous and unambiguous AmbigQA questions (0.13 vs 0.13), so the BNN/Deep Ensemble decomposition does not transfer to black-box LLMs.

Limitations

• Relies on LLM calibration for factoid QA and math reasoning; performance on other tasks may degrade.
• Ambiguity detection is weak on AmbigQA where ambiguities are subtle, so clarification quality limits the method.
• The epistemic term is only approximate, since it assumes input ambiguity is the sole source of aleatoric uncertainty, and the paper focuses on aleatoric decomposition.

Experimental setup

Datasets: AmbigQA Natural Questions GSM8K Instruction induction AmbigInst

Benchmarks: Natural Questions GSM8K AmbigQA AmbigInst

Metrics: AUROC F1 score average aleatoric uncertainty average entropy recall of correct answers

Baselines: Semantic Entropy Ask4Conf Ensembles* SelfConsistency Sample Repetition Sample Diversity

Builds on: Bayesian neural networks Deep Ensembles Semantic Entropy Ask4Conf

Closest papers by meaning

• Distinguishing the Knowable from the Unknowable with Language Models (ICML 2024 · similarity 0.91)
• Variational Uncertainty Decomposition for In-Context Learning (NeurIPS 2025 · similarity 0.91)
• Benchmarking LLMs via Uncertainty Quantification (NeurIPS 2024 · similarity 0.90)
• Conformal Prediction with Paraphrase-Aware Scoring for LLM Uncertainty Quantification (NeurIPS 2026 · similarity 0.90)
• Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similarities (NeurIPS 2024 · similarity 0.90)
• Inv-Entropy: A Fully Probabilistic Framework for Uncertainty Quantification in Language Models (NeurIPS 2025 · similarity 0.89)