Source-linked AI summary

Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback

Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, Christopher D. Manning

arXiv:2305.14975v2cs.CL

TL;DR

Poorly calibrated confidence can impede reliable prediction and human deferral, and RLHF-LMs may have degraded conditional-probability calibration. This paper evaluates ways to elicit confidence from RLHF-LMs and finds that verbalized probabilities are often better-calibrated than conditional probabilities, though results are mixed for Llama-2-70B-Chat. The study is limited mainly to factual-recall and short-form question-answering settings.

  • Problem

    Widely used RLHF-LMs may produce poorly calibrated conditional probabilities, while existing calibration studies have focused largely on unsupervised or maximum-likelihood models.

  • Method

    The paper evaluates numerical and linguistic verbalized-confidence prompts, including multi-hypothesis prompting and temperature scaling, across several RLHF-LMs.

  • Results

    Verbalized probabilities are often better-calibrated than conditional probabilities across several closed RLHF-LMs, with mixed results for Llama-2-70B-Chat.

  • Takeaways & Limitations

    Verbalizing confidence offers a practical way to elicit calibrated scores from RLHF-LMs when log probabilities are unavailable.

  • Takeaways & Limitations

    The experiments focus on factual-recall problems and short-form question answering, leaving reasoning-heavy and longer-form settings for future work.

Abstract

from arXiv · show

A trustworthy real-world prediction system should produce well-calibrated confidence scores; that is, its confidence in an answer should be indicative of the likelihood that the answer is correct, enabling deferral to an expert in cases of low-confidence predictions. Recent studies have shown that unsupervised pre-training produces large language models (LMs) whose conditional probabilities are remarkably well-calibrated. However, the most widely-used LMs are fine-tuned with reinforcement learning from human feedback (RLHF-LMs), and some studies have suggested that RLHF-LMs produce conditional probabilities that are very poorly calibrated. In light of this perceived weakness, we conduct a broad evaluation of methods for extracting confidence scores from RLHF-LMs. For RLHF-LMs such as ChatGPT, GPT-4, and Claude, we find that verbalized confidences emitted as output tokens are typically better-calibrated than the model's conditional probabilities on the TriviaQA, SciQ, and TruthfulQA benchmarks, often reducing the expected calibration error by a relative 50%.

1 Introduction

Calibration matters because confidence estimates can identify likely errors for human review, yet RLHF may worsen language models’ conditional-probability calibration. This paper evaluates verbalized confidence as an alternative and finds it can improve calibration for RLHF-LMs.

  • Motivation: Well-calibrated confidence estimates can identify predictions most likely to be incorrect, enabling human review or override.Poor calibration may contribute to confidently asserted incorrect facts or reasoning.
  • Motivation: Most widely used language models are fine-tuned with reinforcement learning from human feedback, whereas much prior calibration analysis studied maximum-likelihood or unsupervised models.Existing findings suggest RLHF may trade calibrated predictions for closer adherence to dialogue instructions.
  • Approach: The paper evaluates confidence-extraction methods for RLHF-LMs, emphasizing verbalized probabilities because per-token log probabilities are often unavailable.Verbalized confidence is expressed in token space as numerical probabilities or linguistic uncertainty expressions.
  • Findings: Figure 2 describes RLHF as generally worsening Llama-70B log-probability calibration, with reversal of some degradation strongest on TruthfulQA.ECE is lower-is-better, while AUC is higher-is-better.
  • Findings: Verbalized confidence is often better-calibrated than conditional probabilities for popular RLHF-LMs, without fine-tuning for verbalization.The study examines models including ChatGPT, GPT-4, Claude, and Llama-2-70B-Chat.
  • Findings: Prompting models to generate and evaluate multiple answer choices before assigning confidence further improves verbalized probability calibration.This strategy is inspired by psychological findings that considering alternatives can mitigate overconfidence; temperature scaling is also combined with the approach.

2 Evaluating Calibration in RLHF-LMs

The paper evaluates confidence extraction methods for RLHF-LMs, comparing conditional-probability approaches with verbalized numerical or linguistic confidences across multiple calibration measures and QA datasets.

  • The evaluation covers ChatGPT, GPT-4, Claude 1, Claude 2, and Llama-2-70B-Chat.
  • Calibration is measured with raw ECE, temperature-scaled ECE, temperature-scaled Brier Score, and selective-classification AUC.ECE compares confidence with average accuracy within confidence bins; Brier Score measures squared error against correctness labels.
  • On ChatGPT and GPT-4, directly verbalized confidences are better calibrated than model probabilities, including after temperature scaling for GPT-4.ChatGPT’s table contrasts Label prob. and Is True prob. with Verb. 1S, Verb. 2S, and Ling. 1S; GPT-4’s caption reports the same comparison after scaling.
  • The datasets assess factual question answering using TriviaQA, SciQ, and TruthfulQA, with correctness computation designed to reduce exact-match false negatives.TruthfulQA specifically tests whether models mimic human falsehoods.
  • Conditional-probability baselines estimate label probability or the probability that a sampled answer is true.Because many RLHF-LMs are closed-source, probabilities are estimated from samples, using n = 10 for the main protocol.
  • Verbalization methods elicit numerical or linguistic confidence, using one-stage or two-stage prompts and, in one variant, chain-of-thought context.Linguistic expressions are mapped either to human-survey probabilities or values optimized on held-out calibration questions.

3 Results

Across the evaluated RLHF-LMs, verbalized confidence is often better calibrated than conditional probabilities, although the consistency and strength of this pattern vary by prompting strategy and model.

  • Llama-2-70B-Chat shows improvement from verbalized calibration on some metrics, but less consistently than GPT and Claude models.
  • Large RLHF-LMs can often verbalize better-calibrated numerical or linguistic confidences than their conditional probabilities.
  • Generating and evaluating multiple hypotheses improves calibration among methods that verbalize probabilities directly.
  • Numerical probabilities calibrate as well as or better than linguistic uncertainty expressions, while chain-of-thought prompting does not improve verbalized calibration.
  • Claude 1 is less able to verbalize well-calibrated confidence, whereas Claude 2 is generally somewhat stronger than GPT-3.5-turbo at verbalization.

4 Discussion

The paper studies calibration in widely used RLHF language models and finds that prompting models to verbalize confidence can outperform conditional probabilities, while leaving important scope and model-understanding questions open.

  • The study examines calibration in widely used RLHF-LMs and methods for extracting their confidence scores.
  • Verbalized probabilities are better-calibrated than conditional probabilities across several closed models, but results are mixed for Llama-2-70B-Chat.
  • Future work should explain differences in confidence verbalization across GPT, Claude, and Llama-2 models and reduce sensitivity to prompting strategy.
  • The experiments focus on factual-recall problems, so whether the observations hold in reasoning-heavy settings remains open.
  • The study is limited to short-form question answering, leaving longer-form generation for future analysis.

A Additional Results

Additional results show that models express lower confidence on TruthfulQA than on standard factual-recall datasets, while chain-of-thought prompting does not noticeably improve calibration.

  • gpt-3.5 is systematically less confident on TruthfulQA than on TriviaQA and SciQ, with the contrast even more stark for gpt-4.
  • CoT does not noticeably improve calibration across the tested prompt variants or datasets.

B Fitting Procedure for Temperature and Probabilities for Linguistic Expressions

The appendix describes cross-validation procedures for fitting temperature-scaled metrics and linguistic-expression probabilities; CoT calibration results are not consistently improved.

  • Expected calibration error is not consistently improved for any CoT prompt variant on gpt-3.5-turbo.
  • Temperature-scaled ECE and BS use five-fold data splitting, fitting temperature on one fold and evaluating on the remaining folds.Fitting temperature on 20% of the data yields relatively stable temperatures across folds.
  • ECE-t and BS-t are reported as the average temperature-scaled ECE and BS across folds.
  • Ling. 1S-opt. fits probabilities for linguistic confidence expressions on four folds and evaluates on the held-out fold.
  • For Ling. 1S-opt., temperature fitting uses separate folds for expression probabilities, temperature estimation, and final evaluation.

C Prompt Templates

The appendix provides prompt templates for the evaluated methods and distinguishes sampling-based confidence estimation from verbalized confidence methods that require one sample.

  • Prompt templates substitute each evaluation question for the ${THE_QUESTION} variable.
  • Methods above the table’s double line use multiple samples to estimate confidence, whereas methods below it use verbalized confidences from a single sample.
  • Answer correctness is evaluated by asking a language model whether the ground-truth and generated answers are semantically equivalent.
Loading 2305.14975v2…