PaperlayerView exact revision on arXiv

Source-linked AI summary

TRACE: Single-Pass Decoding-Trace Risk Localization for Generation Calibration

Yuebin Xu, Xuemei Peng, Junlan Chen, Zhiyi Chen, Zeyi Wen

arXiv:2609.35387v1cs.CL

TL;DR

Answer-level calibration is challenging because global confidence summaries can miss localized errors in generated responses. TRACE preserves token-level decoding uncertainty as localized risk, and TRACE+ calibrates trace features into probabilities. Across evaluated tasks and models, TRACE+ improves Brier score and AUROC over non-TRACE baselines.

  • Problem

    Global sequence-level summaries can dilute uncertainty concentrated at a critical number, entity, or factual claim, complicating confidence estimation for a decoded answer.

    Source: §1 · p. 1

  • Method

    TRACE records token-level surprisal and predictive entropy during the original generation pass and applies local risk operators; TRACE+ calibrates trace-only features on a held-out split.

    Source: §1 · p. 2

  • Results

    Across four tasks, 19 baselines, and eight LLMs, TRACE+ achieves the best overall Brier score and AUROC.

    Source: §5 · p. 9

  • Takeaways & Limitations

    TRACE provides a label-free confidence score, while TRACE+ produces calibrated correctness probabilities from trace features without extra generations or external verifiers.

    Source: §1 · p. 2

    TRACE requires token-level probabilities or entropy during decoding, which may be unavailable from closed-source APIs; TRACE+ also requires a representative held-out calibration split.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Reliable confidence estimation is essential for large language model deployment. However, answer-level calibration remains challenging because generation errors are often localized: a response may be fluent and high-probability overall while still failing at a critical number, entity, or factual claim. Existing estimators compress token probabilities, sequence likelihoods, entropy, or beam statistics into a global score, which can dilute such local risk signals. We propose TRACE, a single-pass, decoded-answer-preserving confidence estimator that treats decoding-time uncertainty as a trajectory through three steps: (i) recording token-level surprisal and predictive entropy during decoding, (ii) applying local risk operators to preserve uncertainty spikes, and (iii) converting localized trace risk into answer-level confidence. TRACE produces a label-free risk score, while TRACE+ calibrates trace-only features into probabilities using a held-out split, without extra generations or external verifiers. We evaluate four tasks against 19 calibration baselines, and TRACE+ reduces Brier from 0.149 to 0.137 and improves AUROC from 0.758 to 0.792 over the strongest likelihood baseline. Across seven LLMs, TRACE+ improves over the best non-TRACE baseline pool from 0.136 to 0.120 Brier and from 0.764 to 0.817 AUROC. Results show that localizing decoding-time risk provides a general approach to calibration.

1 Introduction

Answer-level confidence is difficult because a response may be fluent and confident overall yet wrong at a critical number, entity, or factual claim. TRACE preserves uncertainty along the decoding trajectory, and TRACE+ calibrates its trace features; the paper evaluates both in a single-pass setting across tasks and models.

1 Introduction

  • Generation calibration estimates the answer-level correctness probability of a decoded response, which matters as LLMs are increasingly deployed where users need confidence in answers.Correctness may hinge on a single number, entity, or factual claim despite a fluent response.Source: §1 · p. 1
  • Global sequence-level confidence summaries can conceal where uncertainty occurs, even when a wrong digit, entity, or factual attribution is produced at a locally unstable step.Local uncertainty peaks are informative indicators of answer-level risk, not direct labels of token-level error.Source: §1 · pp. 1–2
  • TRACE records token-level surprisal and predictive entropy during generation, then applies local risk operators rather than immediately averaging away position-sensitive uncertainty spikes.TRACE is label-free; TRACE+ uses a held-out calibration split to map trace-only features to calibrated probabilities.Source: §1 · p. 2
  • All methods score the same decoded answer without extra samples, retrieval, or external verifiers under the single-pass protocol.The setting also excludes LLM-as-a-judge labels at inference time.Source: §1 · pp. 1–2
  • Across four tasks, 19 confidence estimators, and seven LLMs, TRACE+ achieves the best average Brier score and AUROC across diverse models.The reported evaluation also finds that TRACE+ transfers across model families and scales.Source: §1 · p. 2

2 Related Work

Prior generation-confidence methods derive answer-level scores from token probabilities, likelihood, entropy, beam statistics, or semantic comparisons. TRACE instead focuses on localized decoding-time risk while preserving the decoded answer and avoiding additional sampling or inference procedures.

Calibration and single-pass confidence

  • Many decoding-statistic estimators compress a generation into a global sequence-level score, whereas TRACE preserves localized surprisal and entropy patterns along the decoding trajectory.The surveyed estimators include token-probability and likelihood summaries, entropy-based methods, first-token confidence, entropy dynamics, and beam-distribution statistics.Source: §2 · p. 3

Multi-sample and semantic uncertainty

  • Sampling-based agreement and semantic entropy assess uncertainty across completions or meaning-level equivalence classes, but may require extra samples or inference procedures.The cited approaches include self-consistency, sampling-based methods, semantic entropy, verbalized confidence, and self-evaluation.Source: §2 · p. 3
  • SAR and MARS reweight tokens or sentences by semantic relevance, while TRACE localizes where decoding-time risk concentrates.TRACE asks where risk is concentrated rather than which generated components are semantically important.Source: §2 · p. 3

3 Method

TRACE scores a fixed decoded answer from token-level uncertainty recorded during generation, preserving localized risk instead of reducing the trace to a global average. TRACE+ maps trace-derived features to calibrated probabilities using a held-out split. [p12, p20, p34, p42]

3.1 Problem Formulation

  • For a fixed decoded answer, the task is to estimate answer-level correctness confidence; TRACE is label-free, while TRACE+ uses held-out labels to calibrate probabilities.Source: §3.1 · p. 3

3.3 Decoding Traces as Local Risk Signals

  • At comparable average-risk levels, high-spike samples have higher empirical error rates, and low-average-risk examples can conceal sharp local risk that global scores miss.These patterns support treating decoding uncertainty as an ordered trajectory rather than a single sequence-level statistic.Source: Figure 3 · p. 4, §3.3 · p. 4

3.2 Method Overview

  • TRACE records token-level surprisal and predictive entropy during the original generation pass, then applies local-risk operators and produces a label-free confidence score.TRACE+ calibrates trace-derived features into answer-level probabilities using a held-out split.Source: §3.1 · p. 3, §3.2 · p. 4

3.4 TRACE: From Local Risk to Confidence

  • TRACE combines position-decayed entropy, maximum local-window entropy, and length-normalized selected-token surprisal to preserve distinct localized risks.One fixed hyperparameter configuration is used across tasks and models; the resulting monotone score is for ranking and is not a calibrated probability before post-hoc calibration.Source: §3.4 · pp. 4–5

3.5 TRACE+: Calibrated Trace Confidence

  • TRACE+ trains a lightweight logistic calibrator on trace-only features from a held-out split, without extra decoding passes or external verifiers.The feature set includes entropy confidence, answer length, entropy slope, and position-decayed entropy and surprisal confidence.Source: §3.5 · p. 5

4 Experiments

TRACE+ is evaluated across generation tasks and models against diverse confidence estimators, with analyses testing cross-model transfer, localized-risk aggregation, calibration robustness, and feature contributions.

4.1 Experimental Setup

  • Across four tasks, setup compares 19 confidence estimators on shared decoded answers, using beam statistics from the same prompt for beam-based methods.The evaluation uses MLQA, SVAMP, TriviaQA, and TruthfulQA, with Brier, ECE, and AUROC reported over 20 random calibration/evaluation splits.Source: §4.1 · p. 5, Table 1 · p. 6

4.2 Experimental Results

  • TRACE+ reduces average Brier from 0.149 to 0.137 and improves average AUROC from 0.758 to 0.792 over SeqLogP / Total NLL across four tasks.It achieves the best Brier score on all four tasks; its AUROC is best on TriviaQA, second-best on MLQA and TruthfulQA, and below beam methods on SVAMP.Source: §4.2 · p. 5, Table 1 · p. 6
  • TRACE+ reduces average Brier from 0.136 to 0.120 and improves AUROC from 0.764 to 0.817 against the per-metric Best Non-TRACE oracle across seven LLMs.Task-level gains occur in 21/21 Brier and 17/21 AUROC settings; the estimator landscape places TRACE+ near the upper-left frontier across model families.Source: §4.2 · pp. 6–7, Figure 4 · p. 6, Table 2 · p. 7

4.3 Analysis and Ablations

  • On incorrect generations with localized risk peaks, whole-answer mean aggregation often assigns modest risk, while TRACE local-peak risk remains high.The analysis attributes the difference to sparse peaks being diluted by many low-risk tokens; preserving local risk structure provides information beyond global likelihood or entropy summaries.Source: §4.3 · p. 7, Figure 5 · p. 7
  • Combining position-decayed entropy, local-window entropy, and length-normalized surprisal yields average AUROC 0.772, above each operator alone and near the learned combination's 0.776.Disrupting risk positions reduces ranking quality, while performance remains stable across broad operator-parameter variations.Source: §4.3 · pp. 7–8, Table 4 · p. 8
  • TRACE+ reaches 0.137 Brier and 0.792 AUROC, while adding conventional selected-token likelihood features slightly hurts performance and trace signals are complementary.Calibrating the TRACE scalar improves Brier from 0.182 to 0.154 but leaves AUROC at 0.772; learning operators reaches 0.140 Brier and 0.778 AUROC. Adding TRACE improves MARS from 0.173/0.754 to 0.141/0.775 and TokenSAR from 0.174/0.738 to 0.141/0.781 Brier/AUROC.Source: Table 3 · p. 7, §4.3 · p. 8, Table 4 · p. 8
  • From 20% calibration data onward, TRACE+ has the lowest Brier and highest AUROC; it also leads both metrics across short, medium, and long answers.At 5% calibration data it matches the strongest baseline in Brier, though AUROC is less stable; its reliability and risk-coverage analyses also show close confidence-accuracy alignment and lower risk over most coverage levels.Source: §4.3 · pp. 8–9, Figure 6 · p. 9
  • Without target-task calibration labels, TRACE+ retains competitive AUROC but Brier worsens on tasks with substantially different correctness priors.For example, MLQA Brier is 0.347 versus 0.264 and TruthfulQA Brier is 0.212 versus 0.137; ranking transfers better than calibrated probabilities.Source: §4.3 · p. 9, Table 6 · p. 9

5 Conclusion

The paper studies general generation calibration and proposes TRACE, which models decoding-time uncertainty as a token-level trajectory. Across baselines and LLMs, TRACE improves correctness ranking, while TRACE+ further improves calibration and ranking.

5 Conclusion

  • Across 19 baselines and eight LLMs, TRACE+ achieves the best overall Brier score and AUROC, further improving calibration and ranking beyond TRACE.Ablations attribute gains to localized decoding-trace evidence and report robustness across models, answer lengths, and calibration sizes.Source: §5 · p. 9
  • TRACE combines position-sensitive, local-peak, and length-normalized local risk operators into a label-free confidence score.TRACE+ learns richer trace features with a lightweight held-out calibrator to produce calibrated correctness probabilities.Source: §5 · p. 9
  • TRACE consistently improves correctness ranking over standard likelihood and entropy estimators, while TRACE+ further improves calibration and ranking.The reported gains remain robust across models, answer lengths, and calibration sizes.Source: §5 · p. 9

Limitations

TRACE needs token-level probabilities or entropy during decoding, which may be unavailable through closed-source APIs; TRACE+ also requires a representative held-out calibration split. The appendix contents list additional analyses and experimental results.

Limitations

  • TRACE requires token-level probabilities or entropy during decoding, which may be unavailable from closed-source APIs.Source: Limitations · p. 9
  • TRACE+ additionally requires a representative held-out calibration split.Source: Limitations · p. 9
  • The stated scope is single-answer confidence estimation; TRACE is described as complementing retrieval, multi-sample consistency, and external verification for factual errors.Source: Limitations · p. 9

Appendix Contents

  • Appendix contents list localized-risk analysis, feature definitions, protocol and baseline details, and supplementary experimental results.Source: Appendix Contents · p. 13

A Localized Risk Spike Analysis

The localized-risk analysis finds that uncertainty spikes add predictive information beyond global mean risk, while the surrounding sections define TRACE features and compare inference protocols and baselines.

A Localized Risk Spike Analysis

B TRACE Feature Definitions

  • TRACE combines position-decayed entropy, maximum local-window entropy, and length-normalized surprisal; TRACE+ uses trace-derived features without conventional selected-token likelihood summaries in its final variant.Source: B TRACE Feature Definitions · p. 13, Table 9 · p. 14

C Protocol comparison

D Baseline and Protocol Details

  • TRACE uses only the original answer’s token-level trace, while beam methods require beam statistics and semantic-reweighting baselines use semantic weights; probability metrics calibrate scalar baselines on the same held-out split.Source: D Baseline and Protocol Details · pp. 14–15

E.1 Comparison with Broader UQ Methods

The broader comparison reports TRACE’s strongest AUROC without extra inference or training, and TRACE+’s best overall Brier and AUROC with held-out calibration labels. The following section details cross-model and task-level comparisons.

E.1 Comparison with Broader UQ Methods

  • TRACE achieves the best AUROC without extra inference or training, while TRACE+ achieves the best overall Brier and AUROC with held-out calibration labels.Source: §E.1 · p. 15

E.2 Detailed Cross-Model Results

  • Best non-TRACE Brier and AUROC estimators are selected independently for each target model after averaging over SVAMP, TriviaQA, and TruthfulQA.Source: §E.2 · p. 15

E.3 Localized-Error Analysis

TRACE and TRACE+ outperform the strongest non-TRACE baseline on localized-error AUROC, with the largest subtype gain for number and arithmetic errors.

E.3 Localized-Error Analysis

  • Localized errors are failures determined by a compact answer part, whereas global errors include off-topic, incoherent, abstaining, or multi-span failures.The comparison matches correct answers by task and answer length.Source: §E.3 · p. 15
  • TRACE and TRACE+ raise localized-error AUROC from 0.778 for the strongest non-TRACE baseline to 0.796 and 0.805, respectively.TRACE performs best at 5% FPR; TRACE+ has the highest TPR at 10% and 20% FPR and strongest partial AUROC below 10% FPR.Source: §E.3 · p. 15
  • TRACE+ achieves AUROC of 0.927, 0.836, and 0.607 for number/arithmetic, entity/span, and factual-claim errors, respectively.The gains over the strongest non-TRACE baseline are 0.048, 0.035, and 0.018, respectively.Source: §E.3 · p. 16

E.4 TRACE Operator Analysis

This analysis examines TRACE’s three fixed operators, whether risk position matters, and sensitivity to the fixed hyperparameters. The combined score captures most of the benefit of learned operator fusion, and preserving risk positions improves discrimination.

Position sensitivity

  • Perturbing risk-signal positions lowers average AUROC from 0.772 to 0.755 across four tasks.The result indicates that uncertainty ordering and location provide information beyond the marginal magnitude of token-level risk.Source: §E.4 · p. 16

Operator contributions

  • TRACE’s three operators yield average AUROC of 0.751–0.758 individually; fixed fusion reaches 0.772, versus 0.776 for learned fusion.The fixed aggregation captures most of the benefit of combining the three risk signals without task-specific fitting.Source: §E.4 · p. 16, Table 15 · p. 17

Hyperparameter sensitivity

  • Across tested hyperparameter settings, Brier ranges from 0.152 to 0.160 and AUROC from 0.763 to 0.775.When one mixture weight is varied, the other two are renormalized according to their original ratio; performance remains stable across broad parameter ranges.Source: Table 16 · p. 17, §E.4 · p. 17

E.5 Robustness and Reliability Analyses

The analysis tests TRACE+ across calibration-set sizes, confidence reliability and selective prediction, and answer lengths. It reports robustness across the evaluated sizes and lengths, with reliable confidence scores after calibration.

Evaluation scope

  • The evaluation averages results over MLQA, SVAMP, TriviaQA, and TruthfulQA using the main evaluation split protocol.Calibration-size robustness is reported in Brier and AUROC; length-stratified robustness is also reported in Brier and AUROC.Source: §E.5 · p. 17, Table 19 · p. 18, Table 21 · p. 18

Reliability and selective prediction

  • Reliability analysis reports ECE, Brier, AUROC, and selective-prediction risk at coverage levels, pooling predictions across tasks within each split.The analysis evaluates confidence reliability and its usefulness for deciding which answers to trust.Source: §E.5 · p. 17, Table 20 · p. 18

Robustness and reliability

  • TRACE+ is robust across evaluated calibration sizes, produces reliable confidence scores after calibration, and remains effective across answer-length strata.The analyses cover calibration splits from 5% to 50% and short, medium, and long answers.Source: §E.5 · pp. 17–18, Table 19 · p. 18, Table 21 · p. 18

Related papers

Loading 2609.35387v1…