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

Efficient semantic uncertainty quantification in language models via diversity-steered sampling

NeurIPS 2025 · Large language models › Language, safety and interpretability · approach: Statistical frameworks for LLM evaluation · 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 a diversity-steered sampler that adds an NLI-based semantic-similarity penalty to ARM and masked diffusion decoding, corrected with importance weights and control variates, to estimate semantic entropy and mutual-information uncertainty more sample-efficiently.

Problem

Estimating semantic aleatoric and epistemic uncertainty in free-form LLM question answering requires many expensive IID samples, many of which are semantically redundant.

Approach

Inject a continuous bidirectional-entailment repulsion penalty into the token-level (or denoising-step) proposal distribution using an NLI model lightly finetuned on truncated prefixes ([TRUNC]) or masked spans ([MASK]); then correct the induced bias with self-normalized importance weights and reduce variance with adaptive control variates.

Key results

• On CoQA, the method achieves AUROC of .77 for LLaMA 3 8B-Instruct versus .72 for temperature 2 and .68 for vanilla sampling, and .81 for LLaDA 8B-Instruct.
• Across 16 cases in the SE AUROC table, the proposed method is best or tied in most settings, e.g. TruthfulQA LLaDA 8B: .94 vs .89 for temperature 2.
• For MI, gains over baselines are modest and high-variance (e.g. CoQA LLaMA 3 .68 vs .66 for temperature 2).

Measured effects reported in the paper

Surprising findings

• Gains for MI estimation are modest and high-variance, and the authors attribute this partly to noise in NLI-based clustering.
• The method's advantage is more evident on free-form datasets (CoQA, AmbigQA) than on TriviaQA with short unambiguous answers.

Limitations

• Sequential diversity steering is slower than parallel sampling, with cost growing linearly with N.
• NLI clustering is noisy, and the single-prompt conditioning ignores prompt-wording robustness.
• MI gains are modest and high-variance across datasets.

Experimental setup

Datasets: GLUE MNLI CoQA TriviaQA TruthfulQA AmbigQA

Benchmarks: CoQA TriviaQA AmbigQA TruthfulQA

Metrics: AUROC effective sample size average number of semantic clusters Spearman rho

Baselines: vanilla sampling with temperature 1 temperature 2 diverse beam search SDLG

Builds on: semantic entropy semantically diverse language generation diverse beam search DeBERTa-large-MNLI

Closest papers by meaning

• FLUE: Streamlined Uncertainty Estimation for Large Language Models (AAAI 2025 · similarity 0.92)
• Semantic Density: Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space (NeurIPS 2024 · similarity 0.91)
• Inv-Entropy: A Fully Probabilistic Framework for Uncertainty Quantification in Language Models (NeurIPS 2025 · similarity 0.90)
• Distinguishing the Knowable from the Unknowable with Language Models (ICML 2024 · similarity 0.90)
• Improving Uncertainty Estimation through Semantically Diverse Language Generation (ICLR 2025 · similarity 0.90)
• Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similarities (NeurIPS 2024 · similarity 0.90)