PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive Perspective

Zizhuo Zhang, Xiong Peng, Jingwei Sun, Rong Yao, Borui Jiang, Bo Han

arXiv:2610.07894v1cs.CLcs.LG

TL;DR

Instance-level faithfulness evaluations often vary questions and contexts together, leaving controlled assessment of response switching unclear. PFaithBench pairs sufficient and insufficient contexts for the same question and studies training-data composition; results show prevalent over-answering and trade-offs between answering and abstaining.

  • Problem

    Existing evaluations typically assess different questions under different contextual conditions, confounding question and context variation and obscuring whether models switch appropriately.

    Source: §1 · p. 2

  • Method

    PFaithBench pairs sufficient and insufficient contexts for the same question to evaluate answering when evidence is available and abstaining when it is not.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    35/39 models show substantially higher OverAns than OverAbs, while mismatched training sources induce shortcuts and supervision shifts errors in opposite directions.

    Source: §1 · p. 2, §3.4 · p. 7

  • Takeaways & Limitations

    Faithfulness training depends on context-coupled data and balanced answering and abstaining supervision to avoid dataset-specific shortcuts and skewed behavior.

    Source: §1 · p. 2

Abstract

from arXiv · show

Large language models (LLMs) are expected to answer questions faithfully based on the provided context, abstaining when the context information is insufficient to answer the questions. Existing faithfulness evaluations typically assess each question-context instance in isolation; however, such instance-level evaluation fails to capture a fundamental requirement of faithful behavior: the ability to adapt model responses to changes in available contexts. In particular, a model should provide correct answers when sufficient evidence is present and abstain when it is not. In this work, we propose a Pairwise Faithfulness Benchmark (PFaithBench) that evaluates whether a model can switch between answering and abstaining for the same question under supporting versus non-supporting contexts. Our evaluations across thirty-nine models with seven model families demonstrate that faithfulness fundamentally involves a trade-off between answering and abstaining, and that most current models exhibit a strong bias toward answering, with most faithfulness errors arising from over-answering, i.e., models tend to fabricate a response even when the provided context is insufficient. We further conduct a series of studies on faithfulness training under different data constructions. Our results show that training outcomes are highly sensitive to the specific composition of answering and abstaining data. Constructing answering and abstaining data from mismatched sources can cause models to rely on dataset-specific shortcuts rather than actual context sufficiency. Moreover, increasing answer-supervised data improves answering performance but exacerbates over-answering, while increasing abstaining data reduces hallucination but leads to over-abstention. The code and data are released at https://github.com/tmlr-group/PFaithBench.

1 INTRODUCTION

Instance-level evaluations vary questions and contexts together, obscuring whether models switch appropriately as context sufficiency changes. PFaithBench evaluates that switching for fixed questions and finds prevalent over-answering alongside trade-offs in faithfulness training.

1 INTRODUCTION

  • Existing evaluations often test different questions under different contextual conditions, confounding question variation with context variation.This prevents controlled assessment of whether models switch between answering and abstaining as contextual sufficiency changes.Source: §1 · p. 2
  • Across 39 models from seven families, most exhibit over-answering, with some errors associated with reliance on internal parametric knowledge.Over-answering occurs when models answer despite insufficient context; over-abstention occurs when they abstain despite sufficient context.Source: §1 · p. 2
  • PFaithBench keeps the question fixed while varying contextual sufficiency, enabling controlled assessment of answering-versus-abstaining behavior.The benchmark characterizes over-answering and over-abstention as two primary faithfulness errors.Source: §1 · p. 2
  • Mismatched answering and abstaining sources can induce dataset-specific shortcuts, while increasing either type of supervision shifts errors in opposite directions.More answering data improves answering but exacerbates over-answering; more abstaining data reduces unsupported answering but leads to over-abstention.Source: §1 · p. 2
  • The paper’s contributions are a pairwise benchmark, an analysis of faithfulness errors, and data-centric studies of training composition.The authors emphasize context coupling and balanced answering-abstention data for faithfulness training.Source: §1 · p. 2

2 RELATED WORK

Prior work studies faithfulness across contextual conditions and training approaches, but evaluations often confound question and context variation. This paper motivates pairwise evaluation and analyzes how training-data construction shapes context-sensitive behavior.

Faithfulness evaluation

Faithfulness training

  • Prior faithfulness training includes context-grounded fine-tuning, abstention-aware supervision, preference optimization, and reinforcement learning.The paper uses GRPO (Shao et al., 2024) to study how data construction and composition affect context-sensitive answering and abstaining.Source: §2 · pp. 3–4

3 PAIRWISE FAITHFULNESS EVALUATION

Pairwise faithfulness requires models to answer correctly with sufficient context and abstain when paired context is insufficient. The section defines metrics, constructs PFaithBench, and evaluates existing models’ context-sensitive behavior.

3.1 PROBLEM DEFINITION

  • Pairwise evaluation holds the question fixed while pairing sufficient and insufficient contexts, testing whether models answer correctly in the former and abstain in the latter.Insufficient contexts are constructed by minimally modifying key answer information, enabling comparison without changing the question.Source: §3.1 · p. 4

3.2 METRICS OF PAIRWISE FAITHFULNESS

  • DFAcc measures appropriate switching between answering and abstaining, while SFAcc additionally requires a correct answer under sufficient context.OverAns and OverAbs measure the opposing tendencies to answer under insufficient context and abstain under sufficient context.Source: §3.2 · pp. 4–5

3.3 PFAITHBENCH CONSTRUCTION

  • PFaithBench pairs sufficient and insufficient contexts across multi-hop, single-hop, and counterfactual QA, with insufficient contexts created by altering or removing key evidence and manually verified where applicable.Its four subsets contain 1,000, 762, 791, and 791 examples, respectively.Source: §3.3 · pp. 6–7

3.4 EVALUATION RESULTS

  • 35/39 models show substantially higher OverAns than OverAbs, indicating a prevalent bias to answer despite insufficient context.The authors suggest outcome-based post-training may contribute because it can reward correct answers while treating appropriate abstention as incorrect.Source: §3.4 · p. 7
  • All evaluated models show internal knowledge leakage, and leakage tends to increase with model size within most model families.The paper notes that reliance on internal knowledge can increase unsupported-answer risk when that knowledge is outdated or incorrect.Source: §3.4 · p. 7
  • Many models’ marginal AnsAcc and AbsAcc performance exceeds their joint DFAcc and SFAcc performance, so success on separate contexts may not transfer to pairwise faithfulness.Table 2 reports joint pairwise accuracy, while Table 3 reports marginal answer and abstention accuracy.Source: Table 2 · p. 5, Table 3 · p. 6, §3.4 · p. 8

4 DATA-CENTRIC ANALYSES FOR FAITHFULNESS TRAINING

The analyses test how context pairing and the answer/abstain data mix shape faithfulness training. They find that mismatched sources can cue behavior, while a balanced data mix supports switching between answering and abstaining.

4.1 FAITHFULNESS TRAINING SCHEME

  • The GRPO training scheme rewards correct answers under sufficient context and abstention under insufficient context, while penalizing over-abstention and over-answering.The authors also examine a symmetric reward that penalizes incorrect answers and abstention equally under sufficient context.Source: §4.1 · p. 8

4.2 CONFOUNDING EFFECTS FROM MISMATCHED CONTEXT COUPLING

  • MHSuff-SHInsuff yields OverAns 63.40 on PFB-MultiHop and OverAbs 61.68 on PFB-SingleHop for Llama-3.1-8B-Instruct; PairContext achieves the best pairwise faithfulness in most cases.The authors interpret the pattern as evidence that models may use source-specific cues to decide whether to answer.Source: §4.2 · p. 9
  • Under MHSuff-SHInsuff, Llama-3.1-8B-Instruct has AnsAcc/AbsAcc of 81.80/36.60 on PFB-MultiHop and 32.28/98.56 on PFB-SingleHop.The authors suggest mismatched training may induce a source-conditioned answer-abstention bias.Source: §4.2 · p. 9

4.3 IMPACT OF ANSWERING-ABSTENTION DATA PROPORTION

  • At α = 0, SFAcc falls to zero; at α = 100, OverAns reaches 100% for both Llama models, while DFAcc and SFAcc peak near α = 50 across all three models.Increasing α generally raises OverAns and Leakage while reducing OverAbs.Source: §4.3 · p. 9
  • Faithfulness-trained models retain performance comparable to their untrained base models across the reported reasoning, knowledge, and instruction-following benchmarks.The evaluated benchmarks include AMC, AIME24, AIME25, MATH500, MMLU-Pro, GPQA-Diamond, and IFEval.Source: §4.3 · p. 9

5 CONCLUSION

The conclusion presents PFaithBench as a pairwise test of context-sensitive answering and abstention, and reports over-answering and over-abstention as the main error types. Its construction draws on four public QA datasets, with evaluation samples human-verified.

5 CONCLUSION

  • PFaithBench evaluates whether models adapt between answering and abstaining for the same question, and the evaluation finds over-answering and over-abstention as main errors, with models biased toward over-answering.The conclusion associates over-answering closely with internal knowledge leakage and notes that context coupling and data balance matter for faithfulness training.Source: §5 · pp. 9–10

A.1 DETAILS OF DATA SOURCES

  • PFaithBench is built from HotpotQA, 2WikiMultiHopQA, SQuAD, and ConFiQA, covering multi-hop, single-hop, and counterfactual contexts.The paper describes these sources as supporting diverse contextual conditions for pairwise-faithfulness evaluation.Source: §A.1 · p. 14
  • For multi-hop c− contexts, one supporting passage is replaced; SQuAD contexts have answer spans removed or blurred, while ConFiQA contexts have key information removed or blurred.ConFiQA insufficiency construction applies to both original and counterfactual contexts.Source: §A.1 · p. 14

A.2 STATISTICS OF PFAITHBENCH

  • All evaluation samples undergo human verification, whereas training samples follow the same construction pipeline without manual verification; PFB-ConOri and PFB-ConCF are evaluation-only.Source: §A.2 · p. 14

A.3 RELATIONSHIP BETWEEN PFAITHBENCH AND EXISTING BENCHMARKS

PFaithBench differs from instance-level benchmarks by pairing sufficient and insufficient contexts for the same question across several contextual and reasoning settings. The appendix describes how insufficient contexts are constructed and filtered.

A.3 RELATIONSHIP BETWEEN PFAITHBENCH AND EXISTING BENCHMARKS

  • Unlike benchmarks that assess instances independently, PFaithBench pairs sufficient and insufficient contexts for each question across answerable, unanswerable, and counterfactual settings and single- and multi-hop reasoning.This enables pairwise evaluation of whether a model answers with sufficient context and abstains with insufficient context.Source: §A.3 · p. 14

A.4 MORE DETAILS OF CONSTRUCTING INSUFFICIENT CONTEXT c−

  • For HotpotQA and 2WikiMultiHopQA, c− is constructed by replacing a passage containing a key supporting fact with an irrelevant passage, breaking the evidence chain.Source: §A.4 · p. 15
  • For single-hop and counterfactual QA, Qwen3-32B rewrites c+ into c−, and postgraduate students retain only examples judged insufficient; 762/1,000 single-hop and 791/1,000 counterfactual examples remain.Source: §A.4 · p. 15
  • The rewriting instructions aim to keep contexts coherent and mostly unchanged while removing support for the original answer.Source: §A.4 · p. 15

A.5 MORE EVALUATION DETAILS

The evaluation instructions ask models to answer from supplied contexts or say “I don’t know” when evidence is insufficient. The appendix also specifies response judging, the evaluated model range, and a paired-context example.

A.5 MORE EVALUATION DETAILS

  • Models are instructed to answer only from supplied contexts and respond “I don’t know” when context is insufficient.Source: §A.5 · p. 16
  • The judge classifies responses by matching the final answer to the reference or identifying an explicit refusal, prioritizing an explicit final-answer section when present.A concrete answer after a disclaimer is judged against the reference rather than counted as a refusal.Source: §A.5 · pp. 16–17

A.6 EVALUATION MODELS

  • The evaluation covers 39 models from seven families, with parameter counts from 0.6 billion to 70 billion.Source: §A.6 · p. 17

A.7 CASE EXAMPLE

  • The case example contrasts paired contexts where a rewrite removes answer-critical information but preserves surrounding details and overall meaning.The resulting insufficient context remains highly similar to the sufficient context but no longer supports the answer.Source: §A.7 · p. 17
  • Table 8 defines imbalance as |AnsAcc − AbsAcc| and asymmetry as |max(AnsAcc, AbsAcc) − SFAcc|, with averages across four subsets.Bold marks the minimum within each model family and column.Source: Table 8 · p. 18

B BASE TRAINING METHOD: GRPO

The paper uses GRPO as its training algorithm for context-based question answering in its data-centric analysis.

B BASE TRAINING METHOD: GRPO

  • The GRPO objective is used for context-based question answering in the paper’s data-centric analysis.The passage introduces the objective before describing its components.Source: B BASE TRAINING METHOD: GRPO · p. 18
  • The policy ratio ρ_i,t(θ), normalized advantage Â_i,t, and KL-divergence term D_KL are components named in the objective.The passages identify these terms but do not give further operational details for the advantage or divergence.Source: B BASE TRAINING METHOD: GRPO · p. 18
  • GRPO is the training algorithm used in the paper’s data-centric analysis.Source: B BASE TRAINING METHOD: GRPO · p. 18

C.1 MORE METRICS OF IMBALANCE AND ASYMMETRY

The section uses imbalance and asymmetry to show that marginal answering and abstaining performance can conceal failures on paired contexts. These discrepancies are not reliably resolved by scaling or reasoning enhancements.

C.1 MORE METRICS OF IMBALANCE AND ASYMMETRY

  • Imbalance measures the absolute gap between AnsAcc and AbsAcc, while asymmetry measures the gap between the stronger marginal score and SFAcc.The metrics are defined as |AnsAcc − AbsAcc| and |max(AnsAcc, AbsAcc) − SFAcc|, respectively.Source: §C.1 · p. 18
  • In the pairwise setting, c+ and c− differ only in the key information needed to answer, so faithful behavior requires consistent answering and abstaining across the pair.High values of either diagnostic indicate a larger discrepancy in the corresponding capabilities.Source: §C.1 · p. 18
  • Substantial imbalance and asymmetry persist across model families and scales, and neither increasing scale nor enabling reasoning reliably reduces them.Asymmetry can remain pronounced even when imbalance is relatively small.Source: §C.1 · pp. 18–19

C.2 MORE RESULTS OF SYMMETRIC REWARD

Under symmetric rewards, answering–abstention data composition shows a pattern consistent with the asymmetric-reward results. A moderate mixture generally improves pairwise faithfulness, while answer-dominated training can worsen it.

C.2 MORE RESULTS OF SYMMETRIC REWARD

  • The symmetric reward assigns equal penalties to incorrect answers and abstention, unlike the asymmetric reward used in Section 4.1.The reward under insufficient context remains unchanged.Source: §C.2 · p. 19
  • A moderate mixture of answering and abstention data generally improves DFAcc and SFAcc, while answer-dominated training increases over-answering and degrades pairwise faithfulness.Exclusively abstention-based training instead leads to severe over-abstention and poor pairwise faithfulness.Source: §C.2 · p. 19
  • The composition trend under symmetric reward is qualitatively consistent with that under asymmetric reward, supporting the reported importance of balancing the two kinds of supervision.Source: §C.2 · p. 19

C.3 FAMILY-LEVEL RESULTS

Family-level results show over-answering as the prevalent pairwise faithfulness failure, with no clear positive relationship between model size and pairwise faithfulness. Additional analyses find benefits from balanced supervision and no general loss in instruction-following performance after faithfulness training.

C.3 FAMILY-LEVEL RESULTS

  • Most model families show higher OverAns than OverAbs, while Olmo is the main exception; pairwise faithfulness has no clear positive relationship with model size.Smaller Qwen3.5 and Phi families achieve SFAcc comparable to or higher than substantially larger families.Source: §C.3 · pp. 19–20

C.4 MORE RESULTS OF ANSWERING-ABSTENTION DATA COMPOSITION.

  • Across the four PFaithBench subsets, single-type training leads to severe over-answering or over-abstention, while moderate mixtures generally improve DFAcc and SFAcc.Increasing the answering-data proportion generally increases leakage.Source: §C.4 · p. 21

C.5 MORE RESULTS OF IFEVAL

  • Faithfulness-trained models generally maintain comparable or better IFEval performance than their base models, while improving their ability to distinguish sufficient from insufficient contexts.Source: §C.5 · p. 21

C.6 CASE STUDY

The case study shows that Llama-3.3-70B-Instruct answers from explicit counterfactual evidence but fails to abstain when that evidence is removed.

C.6 CASE STUDY

  • The case pairs a question with counterfactual sufficient, original, and insufficient contexts to compare the model’s responses.The question is “What is Camila Cabello’s native language?”Source: §C.6 · pp. 21–22
  • Under the counterfactual sufficient context, the model answers “Catalan” correctly despite recognizing a conflict with its internal knowledge; the question concerns Camila Cabello’s native language.Source: §C.6 · pp. 21–22
  • After key evidence is removed, the model answers “Spanish,” matching the answer supported by the original context rather than abstaining.The insufficient context retains her birthplace but omits the native-language statement.Source: §C.6 · pp. 21–22
  • The example illustrates internal leakage: the model follows counterfactual evidence when present but falls back on internal knowledge when it is removed.Jointly evaluating both contexts can reveal failures that may remain hidden in instance-level evaluation.Source: §C.6 · p. 21

Related papers

Loading 2610.07894v1…