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

Geometry-Calibrated Conformal Abstention for Language Models

NeurIPS 2026 · 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 Conformal Abstention, a post-hoc method that uses conformal-style thresholds on perplexity-style uncertainty plus a representation-geometry calibrator (MLP knowledge contribution, embedding rotation, anisotropy alignment) to give participation and conditional-correctness guarantees for LLM abstention.

Problem

LLMs often produce confident hallucinations when they lack relevant knowledge instead of abstaining; training-based abstention rewards are fragile, and conformal prediction non-conformity scores are intractable for open-ended generation.

Approach

Apply conformal thresholding to a response-level uncertainty score (perplexity over geometry-calibrated token confidences) so abstention depends on confidence rather than nonconformity scores, and calibrate confidence with features from MLP knowledge contribution, embedding rotation, and anisotropic alignment via a Mahalanobis-distance-based XGBoost calibrator.

Key results

• Proposed method reaches an average conditional correctness of 75.0% under conformal abstention across six datasets (Gemma-3-4B-Instruct).
• With Gemma-3-4B-Instruct, the proposed uncertainty score achieves AUROC 0.83 on NQ and 0.86 on TruthfulQA, outperforming baselines such as SAR (0.80 on NQ).
• Empirical participation and conditional correctness exceed theoretical guarantees across thresholds on Simple-Questions-Wikidata with Gemma-3-4B-Instruct.

Surprising findings

• Calibrating baseline uncertainty scores with the same learned calibrator often degrades their performance, suggesting overfitting on one-dimensional signals.
• Knowledge contribution and embedding rotation are not necessarily tightly coupled, so similar contribution can yield different rotations.

Limitations

• Correctness labels depend on an evaluation pipeline combining HHEM, GPT-4o and GPT-5.1, which may misjudge responses.
• Calibrator requires a labeled reference set and training split, so the method is not fully training-free.
• Theoretical conditional guarantee depends on exchangeability assumptions.

Experimental setup

Datasets: Natural Questions TruthfulQA Simple Questions Wiki SciQ GSM8K CommonsenseQA

Benchmarks: Natural Questions TruthfulQA Simple Questions Wiki SciQ GSM8K CommonsenseQA

Metrics: conditional correctness participation rate AUROC AUPRC

Baselines: Perplexity SAR Semantic Entropy Focus Attention Score Eigen Score ATRMD P(True)

Builds on: conformal prediction llm-polygraph Mahalanobis distance XGBoost

Closest papers by meaning

• Large language model validity via enhanced conformal prediction methods (NeurIPS 2024 · similarity 0.90)
• Conformal Linguistic Calibration: Trading-off between Factuality and Specificity (NeurIPS 2025 · similarity 0.90)
• Conformal Language Modeling (ICLR 2024 · similarity 0.89)
• Conformal Prediction Beyond the Seen: A Missing Mass Perspective for Uncertainty Quantification in Generative Models (NeurIPS 2025 · similarity 0.88)
• Language Models with Conformal Factuality Guarantees (ICML 2024 · similarity 0.88)
• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)