Resolving Knowledge Conflicts in Large Language Models

1 Introduction

• Knowledge conflicts are discrepancies between internal parametric knowledge and non-parametric knowledge supplied through external sources in prompt context.The paper identifies misinformation, varying perspectives, time-sensitive information, and knowledge updates as possible sources of conflict.Source: §1 · p. 1
• Prior studies examined how LLMs choose between parametric and non-parametric knowledge, with choices depending on external-context coherence and model size.This work asks whether models can acknowledge conflicts and how they should respond.Source: §1 · pp. 1–2
• The authors argue that LLMs should provide distinct answers so users can make informed decisions rather than rely solely on either knowledge source.They posit that identifying conflicts, pinpointing their locations, and responding distinctly can support more accurate, comprehensive, and trustworthy responses.Source: §1 · p. 2
• The paper evaluates whether LLMs identify conflicts, pinpoint the conflicting segments, and generate distinct responses from the conflicting information.The framework simulates contextual conflicts and includes three tasks of increasing complexity.Source: §1 · p. 2, Figure 1 · p. 2
• LLMs perform well above random at detecting conflicts, but struggle to pinpoint conflicting information and provide distinct answers.Instruction-based approaches improve GPT-3.5-TURBO on Tasks 1 and 3, while robust conflict handling remains open.Source: §1 · p. 2

2 Framework

• The framework uses diverse knowledge sources, conflict-creation methods, and progressively challenging settings to assess LLMs’ ability to address contextual conflicts.It is designed to reflect real-world knowledge conflicts.Source: §2 · p. 2

2.1 Knowledge Scope

• The generated entity list contains 9,083 unique entities spanning diverse knowledge areas and domains.The framework is independent of this entity list and can be extended to other subject areas.Source: §2.1 · p. 2

2.2 Knowledge Conflict Generation

• For each entity, the framework elicits the LLM’s parametric knowledge and factually modifies it to create conflicting information for the prompt context.Parametric knowledge is elicited in a closed-book setting, with a separate elicitation for each experimental LLM.Source: §2.2 · p. 3

Conflicting Knowledge Creation

• Conflicts are generated by substituting a named entity with a same-type, in-domain entity or shuffling the main entity with one from another same-domain context.Entity substitution uses named-entity recognition; shuffling is used when those models do not identify an entity in the specified categories.Source: Conflicting Knowledge Creation · p. 3

2.3 Tasks

• The three tasks test conflict recognition, question-relevant conflict detection, and generation of distinct answers based on conflicting context and the model’s own knowledge.Task 1 is binary classification; Task 2 uses questions about conflicting or non-conflicting segments; Task 3 evaluates each answer and both answers together.Source: §2.3 · pp. 3–4

2.4 Dataset Analysis

• 96% of Task 1 and 83.5% of Task 2 contain perfectly clear knowledge conflicts; Fleiss’ Kappa among five annotators is 0.51, indicating moderate agreement.Conflict-generation methods may fail when context is not unique to a specific entity, such as when multiple people hold the same title.Source: §2.4 · p. 4
• Human evaluation assesses whether generated Task 1 and Task 2 examples contain clear knowledge conflicts.The methods may fail when context is not unique to a specific entity.Source: §2.4 · p. 4

3.1 Baselines

• The evaluation includes zero-shot, few-shot, Chain-of-Thought, Generated Knowledge Prompting, Self-ask, Break-down, and Self-Consistency approaches.Self-Consistency uses five responses at temperature 0.7; other experiments use temperature 0.Source: §3.1 · pp. 4–5, §3.2 · p. 5
• The dataset contains 9,083 distinct entities, approximately evenly distributed across 20 domains; about one-third of conflicts use named-entity substitution and two-thirds use entity shuffling.Source: Table 1 · p. 4

3.2 Models and Settings

• GPT-3.5-TURBO is the main experimental LLM; few-shot prompts use four exemplars for Tasks 1 and 2 and two for Task 3.The exemplars come from different primary fields and conflict-generation methods and are balanced between positive and negative samples.Source: §3.2 · p. 5

Task 1: Contextual Knowledge Conflict

• The proposed approach outperforms all baselines on F1, although LLMs’ overall contextual conflict identification is only above random.LLMs tend to declare “no conflict,” producing nearly perfect precision but low recall; obligatory reasoning with CoT alleviates concerns about negative predictions.Source: Table 2 · p. 5, §4 · p. 5

Task 2: QA-Span Knowledge Conflict

• LLMs struggle to pinpoint conflicting information, with a considerable performance decline on QA-span detection.Self-ask is the strongest baseline, and its intermediate answers help narrow the conflict scope and encourage localization.Source: Table 3 · p. 5, §4 · p. 5

Task 3: Distinct Answers Generation

• Both-answer accuracy hovers around 50%, and LLMs struggle to provide distinct answers simultaneously.Except in zero-shot prompting, answers based on conflicting knowledge are more accurate than parametric-based answers.Source: Table 4 · p. 5, §4 · p. 5

• The authors explore conflict-specific instructions because prior work found instruction-based methods can induce new LLM abilities (Wang et al., 2023a).Source: §5 · p. 6

Task 1: Contextual Knowledge Conflict

• For contextual conflict detection, a four-step, sentence-level procedure with localization achieves higher F1 than all baselines, with slightly lower precision.The approach elicits entity knowledge, breaks context into sentences, checks conflicts sentence by sentence, then evaluates remaining sentences against parametric knowledge.Source: §5 · p. 6

Task 2: QA-Span Knowledge Conflict

• For QA-span detection, the approach elicits a parametric answer, finds answer-bearing context sentences, and compares their answers, but falls short of all baselines.The authors identify improving conflict localization as an opportunity for future research.Source: §5 · p. 6

Task 3: Distinct Answers Generation

• For distinct-answer generation, separating sources with “solely” and “disregard” and repeating the question improves performance across all three metrics.The question is repeated because LLMs have limited ability to retain information across extended contextual spans (Chen et al., 2023; Liu et al., 2023).Source: §5 · p. 6

Knowledge Domain

• Knowledge domain substantially affects performance across all three tasks; History and Health Sciences lead on Tasks 1 and 2, while Biology leads on Task 3.Art history and anthropology pose greater challenges; the authors hypothesize that domain differences may relate to conflicting information in pretraining data.Source: §6 · p. 6, Figure 3 · p. 7

Prompt Consistency Check

• Performance is generally stable across instruction phrasings, and CoT has lower standard deviations than few-shot prompting.Source: Table 5 · p. 7, §6 · p. 7

Conflict Generation Method

• Entity shuffling yields better performance than in-domain named-entity substitution in the reported conflict-generation comparison.The authors suggest conflicts about the main entity may be easier for LLMs to identify and specify.Source: Figure 4 · p. 7, §6 · p. 7

Number of In-context Exemplars

• F1 generally plateaus with as few as two exemplars, though more help CoT on Task 2 and few-shot prompting on Task 3; exemplar effects saturate at eight.Increasing exemplars may help, but their effect is limited and does not induce perfectly robust approaches.Source: §6 · pp. 7–8, Figure 5 · p. 8

More Capable LLMs

• GPT-4 improves accuracy for both parametric-based and conflicting-knowledge-based answers on Task 3, but scaling alone does not solve the challenge.The study assesses GPT-4 only on the most challenging task, due to budget constraints.Source: Table 6 · p. 8, §6 · p. 8

Finetuning

• Fine-tuning yields F1 0.953 on Task 1 and F1 0.745 on Task 2; on Task 2, recall exceeds precision.Training data are from Political Science and test data from Geology to avoid the chance of overfitting.Source: §6 · p. 8

• Fine-tuning results surpass baselines on Tasks 1 and 2 but remain imperfect.Source: Table 7 · p. 8

Understanding and expanding the knowledge abilities of LLMs

• LLMs encode factual knowledge but face outdated knowledge, factuality problems, and hallucination; prior responses include retrieval, search, and other neural LMs.The paper defines its focus as conflicts between internal parametric and external non-parametric knowledge, which can call response reliability into question.Source: §7 · pp. 8–9

Knowledge Conflict in LLMs

• Prior conflict studies examine source choice, confirmation bias, training augmentation, and distractors; the paper identifies conflict detection, pinpointing, and distinct answers as underexplored desiderata.The cited studies include Mallen et al. (2023), Longpre et al. (2021), Xie et al. (2023), Neeman et al. (2023), and Qian et al. (2023).Source: §7 · p. 9

8 Conclusion

• LLMs excel at detecting knowledge conflicts but struggle to pinpoint conflicting segments and provide distinct responses; instruction-based approaches improve Tasks 1 and 3, while Task 2 remains especially challenging.Further analyses find that exemplars, domains, and conflict simulation methods affect performance.Source: §8 · p. 9
• The framework simulates contextual knowledge conflicts and evaluates LLMs on detection, pinpointing conflicting segments, and generating distinct answers or viewpoints.Source: §8 · p. 9
• The framework and study provide quantitative avenues for evaluating and improving LLMs’ ability to respond to knowledge conflicts.Source: §8 · p. 9

Limitations

• Synthetic conflicting knowledge may be factually incorrect because it is created by replacing facts with randomly related entities; the study emphasizes conflict presence, not factual correctness.The authors identify more plausible alternatives as a direction for future work.Source: Limitations · p. 9
• The experiments treat the top answer from zero-shot prompting as the LLM’s unique parametric knowledge for an entity or question.The authors note that multiple parametric answers are explored in Appendix F.Source: Limitations · p. 9
• Generated contexts may contain hallucinations, and probing LLMs to build the entity list biases it toward entities well known to those models.The authors argue this bias reduces the likelihood of hallucination; their focus is not whether parametric knowledge is factually accurate.Source: Limitations · p. 10
• The synthetically generated dataset is limited to word-level edits, whereas real-world conflicts may be more complex and introduce new entities that increase hallucination risk.The paper also points to multi-hop questions and subjective datasets as additional settings explored in appendices.Source: Limitations · p. 10

Ethics Statement

• Knowledge conflicts raise ethical concerns because trust in external or internal knowledge depends on context, and some conflicts involving opinions may have no unequivocally correct answer.The paper contrasts knowledge updates, where external information may take precedence, with possible malicious interference, where internal beliefs may be preferable.Source: Ethics Statement · p. 10

A Discussion and Future Work

• Presenting distinct answers can give users agency, reduce reliance on one information source, and help LLMs reflect plural opinions in social and political issues.Source: A Discussion and Future Work · p. 14
• Future work could test multiple valid answers, combine the three tasks, evaluate opinions as conflicts, and assess different models, especially open-source LLMs.The current setup has one context-derived answer and one answer from parametric knowledge.Source: A Discussion and Future Work · pp. 14–15

B Prompting Methods

• Zero-shot prompting requests a direct answer, few-shot prompting supplies exemplars, and Chain-of-Thought prompting guides reasoning; for Tasks 2 and 3, it elicits parametric and context-based answers separately.For Task 1, Chain-of-Thought prompting asks models to break context into atomic facts and check for inconsistencies.Source: B Prompting Methods · p. 15
• Generated Knowledge Prompting adds elicited parametric knowledge as input, Self-ask elicits follow-up questions, and Break-down prompting classifies sentences before consolidating responses.Source: B Prompting Methods · p. 15
• Self-Consistency samples diverse reasoning paths and selects the most consistent answer by marginalizing over those paths; experiments combine it with Chain-of-Thought and Generated Knowledge Prompting.Source: B Prompting Methods · p. 15

C Experiment Details

• For Task 3, correct responses must follow the specified format; parametric ground truth comes from answering the question alone, and cases where the two ground truths coincide are removed.The context-based ground truth uses the substituted or shuffled entity.Source: C Experiment Details · p. 15
• The additional figures present Task 3 accuracy by conflict-creation strategy and prompting performance across the other ten knowledge domains.Source: Figure 6 · p. 16, Figure 7 · p. 16

D Additional Analysis

• LLMs perform better on Task 3 when conflicts are introduced through entity shuffling than through named-entity substitution, consistent with Task 1 and Task 2 findings.Source: D Additional Analysis · p. 16
• Performance in identifying conflicts, locating conflicting information, and generating distinct responses varies across knowledge domains.Source: D Additional Analysis · p. 16
• Task 3 performance remains consistent across additional instruction variations, with Chain-of-Thought showing greater prompt robustness.Table 8 covers Task 3 accuracy under different instruction texts.Source: D Additional Analysis · p. 16, Table 8 · p. 17
• In the qualitative example, the same description yields “Pareto efficiency” as parametric knowledge and “Trade balance” as conflicting knowledge; the proposed approach significantly enhances performance.Source: D Additional Analysis · p. 17

E Prompt Text

• Tables 13–19 provide the adopted prompt texts to facilitate reproducibility.Source: E Prompt Text · p. 17

F Multiple Parametric Knowledge Answers

• 25% of additional parametric answers matched the top answer exactly and 65% were synonyms; the remaining answers tended to be invalid or incorrect.The study sampled 600 questions with conflicts and generated three additional answers per question; the authors conclude that relaxing the single-answer assumption would not significantly change findings.Source: F Multiple Parametric Knowledge Answers · p. 17

G Multi-hop Questions

• 0.795 precision, 0.995 recall, and 0.884 F1 were obtained for Task 1 with Chain-of-Thought prompting on the two-hop dataset.The dataset evaluation used 400 positive and 400 negative examples and supports the conclusion that LLMs identify conflicts above random in this multi-hop setting.Source: G Multi-hop Questions · p. 17

H Subjective Datasets

• The evaluation dataset sampled 350 controversial situations and paired parametric viewpoints with opposite viewpoints, yielding 350 positive and 350 negative samples.Examples include giving money to a son, accusing a murderer, giving blood, and protesting for Black Lives Matter.Source: H Subjective Datasets · p. 18
• CoT prompting on PrimeValue achieved 0.98 precision, 0.78 recall, and 0.87 F1 for Task 1, with results above random.The authors note that this setting assumes one unique firm viewpoint for each situation.Source: H Subjective Datasets · p. 18, Table 22 · p. 30
• On 600 sampled questions, three GPT-3.5-TURBO parametric answers were usually nearly the same, and the authors say conclusions would not change significantly without the unique-answer assumption.Source: Table 20 · p. 30
• Task 1 results on Compositional Celebrities also indicate conflict detection beyond random chance.Source: Table 21 · p. 30