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

Conformal Prediction with Paraphrase-Aware Scoring for LLM Uncertainty Quantification

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 paraphrase-aware nonconformity scores that aggregate a learned hidden-state proxy's probabilities across paraphrases, integrated with split CP and QCCP to yield compact, near-nominal-coverage prediction sets under rewording.

Problem

LLM uncertainty estimates and conformal prediction sets are unstable under meaning-preserving paraphrases of the input, producing inflated or unreliable prediction sets even when marginal coverage is guaranteed.

Approach

Label-wise nonconformity scores computed by a lightweight MLP proxy on LLM last-layer hidden states, averaged (or weighted/worst-cased) over a set of LLM-generated paraphrases, then calibrated with split CP or QCCP; theory shows score exchangeability under a distributional closedness assumption for test-only rewording.

Key results

• Under QCCP on Qwen3-8B normal/fully reworded settings, PA achieves set sizes roughly 2x or more smaller than TRON-adapt. and SPUQ-adapt. (e.g., DS 1.24 vs 2.79-2.85 normal) with coverage 89.9-95.6%.
• In the semi-reworded setting, PA attains the smallest set sizes in most cases (e.g., DS SS 1.00 vs 2.61 for LAC under split CP) while coverage ranges 92-96% for split CP.
• Ablation: hidden-state proxy reduces Semi SS from 2.8929 (raw probabilities) to 1.3725; six-view aggregation further reduces set change by 15.8%.

Surprising findings

• Raw-probability averaging across paraphrases increases set change (0.3561 to 0.5785), while the learned proxy accounts for most of the compactness gain.
• APS overshoots coverage substantially (up to 100%) across settings, and LAC/SPUQ/TRON can be unstable under semi-reworded test shift.

Limitations

• Semi-reworded coverage guarantee relies on an approximate distributional closedness assumption that is not guaranteed by the stochastic paraphrase pipeline; deviation bound scales with m epsilon.
• Requires hidden-state access and labeled proxy-training data; black-box API variant only in appendix.
• Coverage is sometimes conservative (up to 95.6%) and fixed-alpha comparisons do not match achieved coverage across methods.

Experimental setup

Datasets: MMLU CosmosQA HellaSwag HaluDial HaluSum MedMCQA MedQA

Benchmarks: MMLU CosmosQA HellaSwag HaluDial HaluSum MedMCQA MedQA

Metrics: coverage rate set size accuracy Brier score NLL

Baselines: LAC APS TRON-adapt. SPUQ-adapt.

Builds on: split conformal prediction quasi-conditional conformal prediction SPUQ TRON LAC APS

Closest papers by meaning

• Conformal Prediction Beyond the Seen: A Missing Mass Perspective for Uncertainty Quantification in Generative Models (NeurIPS 2025 · similarity 0.91)
• Inv-Entropy: A Fully Probabilistic Framework for Uncertainty Quantification in Language Models (NeurIPS 2025 · similarity 0.90)
• SPANUQ: Span-Level Uncertainty Quantification for Large Language Model Generation (NeurIPS 2026 · similarity 0.90)
• Benchmarking LLMs via Uncertainty Quantification (NeurIPS 2024 · similarity 0.90)
• Decomposing Uncertainty for Large Language Models through Input Clarification Ensembling (ICML 2024 · similarity 0.90)
• Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similarities (NeurIPS 2024 · similarity 0.89)