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

Localizing Input Uncertainty Quantification for Large Language Models via Shapley Values

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 0 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 ShaQ, a Shapley-value framework that localizes input-induced (aleatoric) uncertainty to ambiguous spans of an LLM input, using bottom-up marginalization over LLM-generated premises, and shows gains in ambiguity detection on AmbiEnt and AmbigQA.

Problem

Input-level uncertainty scores for LLMs do not reveal which parts of an ambiguous or under-specified input cause uncertainty, so users cannot target clarification; existing input-level methods give only a single aggregate score.

Approach

Treat ambiguous spans as players in a cooperative game with value = mutual information between the answer Y and span clarifications; Shapley values give per-span attributions that sum exactly to total input-induced uncertainty, and a bottom-up marginalization keeps marginal contributions non-negative.

Key results

• On multi-span AmbiEnt with GPT-5.4 mini, ShaQ reaches 86.56 F1, 80.05 AUROC and 91.27 AUPRC, versus 59.64 F1 / 53.40 AUROC for ICE.
• On the high-uncertainty AmbigQA superset in uncertainty-guided clarification, ShaQ gives the largest entropy reduction (0.0389) with the fewest word edits (5.52) versus Semantic Entropy (0.0330, 6.45).
• Across Gemini backbones on multi-span AmbiEnt, ShaQ reaches AUROC 85.85 with Gemini-2.5-flash, though it does not beat LOO on F1 for Gemini-3.1-flash-lite.

Surprising findings

• Baselines like ICE and Deep Ensembles sometimes increase predictive entropy after clarification (ΔH = -0.0633 for ICE), suggesting input-level uncertainty can fail to point to the right spans.
• On AmbiEnt single-label-space NLI, dropping semantic clustering still gives the same trend, reducing LLM calls.

Limitations

• No gold annotations for ambiguous spans or premises, so span-level localization is not directly evaluated; only input-level ambiguity detection is measured.
• Relies on several LLM calls and sampling (premises x answers per coalition), with cost growing with span count.
• Results depend on proprietary API backbones (GPT-4, GPT-5.4 mini, Gemini), limiting reproducibility; few-shot demonstrations were manually annotated by the authors.

Experimental setup

Datasets: AmbigQA AmbiEnt MediTOD

Benchmarks: AmbigQA AmbiEnt MediTOD

Metrics: best F1 AUROC AUPRC entropy reduction word-level edit distance

Baselines: Input Clarification Ensemble Deep Ensembles Ask4Conf-D Semantic Entropy Sample Diversity Sample Repetition Self-Consistency LOO

Builds on: Input Clarification Ensemble SHAP Semantic Entropy

Closest papers by meaning

• SPANUQ: Span-Level Uncertainty Quantification for Large Language Model Generation (NeurIPS 2026 · similarity 0.89)
• Inv-Entropy: A Fully Probabilistic Framework for Uncertainty Quantification in Language Models (NeurIPS 2025 · similarity 0.89)
• Large Language Models Must Be Taught to Know What They Don’t Know (NeurIPS 2024 · similarity 0.88)
• FLUE: Streamlined Uncertainty Estimation for Large Language Models (AAAI 2025 · similarity 0.88)
• Position: Uncertainty Quantification Needs Reassessment for Large Language Model Agents (ICML 2025 · similarity 0.88)
• CoMet: Context and Multiplicity Decomposition for Multimodal Uncertainty Estimation (NeurIPS 2026 · similarity 0.88)