PaperlayerView exact revision on arXiv

Source-linked AI summary

Micro-Act: Mitigating Knowledge Conflict in LLM-based RAG via Actionable Self-Reasoning

Nan Huo, Jinyang Li, Bowen Qin, Ge Qu, Xiaolong Li, Xiaodong Li, Chenhao Ma, Reynold Cheng

arXiv:2506.05278v2cs.CLcs.AI

TL;DR

Retrieved knowledge can contradict LLM parametric knowledge, creating conflicts that undermine question answering. MICRO-ACT uses a hierarchical action space to perceive context complexity and adaptively decompose comparisons. Across five benchmark datasets and three conflict types, it outperforms state-of-the-art baselines while maintaining competitive performance on conflict-free questions.

  • Problem

    Retrieved information can contradict an LLM’s parametric knowledge, undermining RAG’s benefits in downstream tasks such as question answering.

    Source: §1 · pp. 1–2

  • Method

    MICRO-ACT uses a hierarchical action space to perceive context complexity and adaptively decompose knowledge comparisons into finer-grained steps.

    Source: §1 · p. 2

  • Results

    Across five benchmark datasets and three conflict types, MICRO-ACT consistently outperforms state-of-the-art baselines and maintains competitive performance on conflict-free questions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The reported findings support MICRO-ACT’s effectiveness and robustness for resolving knowledge conflicts in real-world RAG systems.

    Source: §1 · p. 2

    MICRO-ACT requires additional intermediate steps, which the authors characterize as modest overhead justified by improved conflict resolution.

    Source: §10 · p. 10

Abstract

from arXiv · show

Retrieval-Augmented Generation (RAG) systems commonly suffer from Knowledge Conflicts, where retrieved external knowledge contradicts the inherent, parametric knowledge of large language models (LLMs). It adversely affects performance on downstream tasks such as question answering (QA). Existing approaches often attempt to mitigate conflicts by directly comparing two knowledge sources in a side-by-side manner, but this can overwhelm LLMs with extraneous or lengthy contexts, ultimately hindering their ability to identify and mitigate inconsistencies. To address this issue, we propose Micro-Act a framework with a hierarchical action space that automatically perceives context complexity and adaptively decomposes each knowledge source into a sequence of fine-grained comparisons. These comparisons are represented as actionable steps, enabling reasoning beyond the superficial context. Through extensive experiments on five benchmark datasets, Micro-Act consistently achieves significant increase in QA accuracy over state-of-the-art baselines across all 5 datasets and 3 conflict types, especially in temporal and semantic types where all baselines fail significantly. More importantly, Micro-Act exhibits robust performance on non-conflict questions simultaneously, highlighting its practical value in real-world RAG applications.

1 Introduction

Knowledge conflicts between retrieved information and LLM parametric knowledge can undermine QA, while existing comparison methods face granularity and adaptability limitations. MICRO-ACT adaptively decomposes comparisons and outperforms state-of-the-art baselines across conflict benchmarks while remaining robust on conflict-free questions.

1 Introduction

  • Conflicts arise when retrieved information contradicts an LLM’s parametric knowledge, undermining RAG’s benefits and raising concerns about downstream QA reliability.Retrieval may introduce noisy, outdated, or incorrect information.Source: §1 · pp. 1–2
  • Prior in-context methods rely on manually crafted instructions and side-by-side comparisons that can miss different-granularity conflicts and impair conflict-free performance.The introduction identifies limited cross-domain adaptability and the assumption that a conflict is present as additional concerns.Source: §1 · p. 2
  • MICRO-ACT dynamically adjusts granularity by perceiving model-level complexity preferences and adapting the granularity of each action.This decomposition supports conflict detection beneath superficial context.Source: §1 · p. 2
  • Across five benchmark datasets and three conflict types, MICRO-ACT consistently outperforms state-of-the-art baselines and maintains competitive performance on conflict-free questions.The introduction also reports that MICRO-ACT mitigates over-rationalization by locating conflicts beneath superficial context.Source: §1 · p. 2

2 Related Work

Related work describes RAG development and knowledge conflicts involving retrieved evidence, model beliefs, or inconsistent documents. Existing conflict solutions modify models or compare sources, but can rely on ad hoc instructions and miss subtle discrepancies.

2 Related Work

  • RAG research has refined retrieval, model architectures, dynamic retrieval, domain adaptation, and efficient fine-tuning to improve adaptability and reliability.The cited examples span retrieval modules and generation architectures.Source: §2 · p. 2
  • Knowledge conflicts include disagreements between retrieved evidence and model beliefs, as well as inconsistencies among retrieved documents.The discussion notes their particular relevance in evolving domains where timely accuracy matters.Source: §2 · pp. 2–3
  • Conflict solutions either modify model parameters or architectures, or explicitly reconcile source discrepancies, but may assume external evidence should override internal knowledge or rely on ad hoc comparisons.The related-work discussion identifies a need for nuanced reasoning and validation.Source: §2 · p. 3

3 Preliminaries

The preliminaries describe RAG as retrieval followed by answer generation conditioned on evidence and parametric knowledge. They define knowledge conflict as factual or logical inconsistency between parametric knowledge and retrieved evidence.

3.1 RAG for QA

  • RAG retrieves potentially relevant textual fragments for a query, then generates an answer conditioned on retrieved and parametric knowledge.The process is described as a retrieval phase followed by a generation phase.Source: §3.1 · p. 3

3.2 Knowledge Conflict

  • A knowledge conflict occurs when parametric knowledge and at least one retrieved fragment are factually or logically inconsistent.The passage frames the conflict as an inconsistency between Kp(q) and some Kr(ei).Source: §3.2 · p. 3

4 Methodology

MICRO-ACT combines a hierarchical action space with a reasoning process that can adapt comparison granularity. Its actions gather information, check for conflicts, and decompose complex comparisons into finer steps when needed.

4 Methodology

  • MICRO-ACT combines a hierarchical action space, a Reasoning Body, and adaptive-granularity and optimization strategies.The action-space decomposition component refines context granularity.Source: §4 · p. 3

4.1 Hierarchical Action Space

  • Navigational actions elicit parametric knowledge and reasoning paths, while functional actions verify consistency between parametric knowledge and retrieved evidence.The action space also includes bridging actions that adjust granularity.Source: §4.1 · p. 4
  • When a side-by-side comparison of complex context may miss subtle conflicts, decomposition breaks the action into smaller steps and continues until confidence is sufficient or the turn limit is reached.New sub-actions examine further evidence fragments to increase the chance of revealing fine-grained conflicts.Source: §4.1 · p. 4

4.2 Reasoning Body

  • The Reasoning Body iterates through thoughts, selected actions, and resulting observations, adjusting granularity through decomposition when subtle conflicts require finer checks.After the step budget, the process generates a final answer.Source: §4.2 · pp. 4–5
  • Dynamically selecting navigational, functional, and decomposition actions is intended to detect and mitigate subtle conflicts and improve final-output reliability.Figure 2 illustrates the process.Source: §4.2 · p. 5

5 Understanding Complexity-Driven Knowledge Decomposition Dynamics

The analysis models how newly inferred knowledge and conflict verification evolve across decomposition turns, linking context complexity and model capability to proactive decomposition.

5 Understanding Complexity-Driven Knowledge Decomposition Dynamics

  • The paper characterizes the turn-t distribution of newly inferred knowledge representations using the trajectory over the preceding t −1 turns.The representation is often obtained by decomposing the current knowledge representation; the formulation distinguishes ground-truth conflicts from potentially incorrect conflicts identified by the model.Source: §5 · p. 5
  • Higher context complexity increases conflict-verification probabilities and the risk of inaccurate conflict detection.The analysis notes longer contexts and harder domains as examples of greater complexity.Source: §5 · p. 5
  • Less capable LLMs are more susceptible to erroneous conflicts, requiring further decomposition and increasing the probability of newly inferred knowledge.The section relates model compatibility to the likelihood of being influenced by an incorrect conflict.Source: §5 · p. 5

6 Preventing Infinite Decomposition

Micro-Act combines a complexity-aware trigger with safeguards against unbounded recursive decomposition, stopping further splits when the context is manageable or the turn bound is reached.

6 Preventing Infinite Decomposition

  • Micro-Act prevents infinite decomposition and supplements this theoretical safeguard with a hard maximum turn budget.The safeguard is described as building on the probabilistic dynamics in Eq. (12).Source: §6 · p. 5
  • A decomposition step is triggered only when the current context’s latent complexity score Ct exceeds threshold τ.τ is the minimum complexity the underlying LLM can handle confidently.Source: §6 · p. 5
  • At the finite bound Tτ, the DECOMPOSE branch has zero probability of further actions, while an LLM that understands the context simply refuses to split it further.The stopping criterion follows a strict inequality because each decomposition shortens the context and narrows its semantic scope.Source: §6 · p. 5

7 Experiments

MICRO-ACT is evaluated across five benchmark datasets and shows gains across conflict types and models while maintaining conflict-free robustness. Further analyses examine adaptive decomposition, reasoning models, action components, and a real conflict case.

7.1 Experiment Settings

  • The evaluation uses five ConflictBank and KRE datasets, QA accuracy, multiple-choice answers, and samples of 3,000 ConflictBank and 2,000 KRE examples.Compared methods include generic reasoning and generation-aided reasoning approaches.Source: §7.1 · pp. 5–6

7.2 Main Results

  • MICRO-ACT surpasses all baselines across tested LLMs and five datasets covering three conflict types; gains reach 9.40% on ConflictBank and 6.65% on KRE for GPT-4o-mini.For LLaMA-3.1-8B, the reported gains are 11.47% on ConflictBank and 13.85% on KRE.Source: §7.2 · p. 6, Table 1 · p. 6

7.3 Over-Rationalization Issue

  • Over-rationalization—supporting contradictory information as equally valid—is more frequent in temporal and semantic conflicts, where MICRO-ACT’s decomposition targets underlying conflicts missed by side-by-side comparisons.The paper also reports this behavior more frequently for GPT-4o than GPT-4o-mini, with GKP performance degradation.Source: §7.3 · p. 7

7.4 Robustness Under Conflict-Free Scenarios

  • MICRO-ACT gains over 24% in conflict scenarios while sacrificing less than 2% accuracy in conflict-free cases versus end-to-end or Self-Ask.Existing generic methods lose 70–95% when conflicts arise, while generation-aided methods such as GKP have lower conflict-free accuracy.Source: §7.4 · p. 7

7.5 Complexity Perception Analysis

  • Decomposition increases with complexity: its rate rises from 15% for 0–100-token contexts to 95% for 400+-token contexts, and GPT-4o-mini decomposes more often than GPT-4o across complexity dimensions.The analysis uses context length, domain difficulty, and perplexity as complexity measures; MICRO-ACT adapts without model-specific tuning.Source: §7.5 · pp. 7–8

7.6 General LLMs vs. Reasoning LLMs

  • Reasoning-oriented Gemini-2.5-flash-thinking and o3-mini outperform general-purpose GPT-4o and Llama-3.1-70B, with the largest gap on misinformation conflicts.The gap narrows for temporal and semantic conflicts, where over-rationalization is more frequent; MICRO-ACT mitigates it while still surpassing the general models.Source: §7.6 · p. 8, Figure 6 · p. 8

7.7 Ablation Study

  • Removing navigational actions drops performance 9.4%, functional actions 12.7%, and decomposition over 20%, with granularity guidance especially important for complex contexts.Iterative decomposition refines inputs until the other actions can operate at suitable granularity.Source: §7.7 · pp. 8–9, Table 2 · p. 8

7.8 Case Study

  • In the Paul Eugène Gillon case, MICRO-ACT distinguishes 2010 from recent appointment references and Norway from France, then reasons to the recent French-politician answer.The case illustrates fine-grained conflict points that simple side-by-side comparisons can miss.Source: §7.8 · p. 9, Figure 7 · p. 9

8 Cost Analysis

MICRO-ACT uses more tokens and slightly more time than GKP on ConflictBank, with additional overhead occurring when conflicts trigger deeper decomposition.

8 Cost Analysis

  • On ConflictBank, MICRO-ACT uses roughly 2.8 times more input tokens and 1.3 times more output tokens than GKP, with latency increases of 0.6 s for GPT-4o and 0.3 s for GPT-4o-mini.The reported extra cost per query is $0.008 for GPT-4o and $0.0005 for GPT-4o-mini.Source: §8 · p. 9
  • The additional overhead appears when genuine conflicts trigger deeper decomposition; conflict-free questions finish as quickly as the baseline.Source: §8 · p. 9
  • The authors characterize the added cost and delay as marginal relative to the reported conflict-resolution accuracy gains.Source: §8 · p. 9

9 Conclusion

MICRO-ACT addresses RAG knowledge conflicts through hierarchical action decomposition that perceives context complexity and breaks comparisons into fine-grained steps. The authors report effectiveness across datasets and conflict types alongside robustness in non-conflict scenarios.

9 Conclusion

  • MICRO-ACT addresses RAG knowledge conflicts through hierarchical action decomposition, automatically perceiving context complexity and breaking comparisons into fine-grained steps.The authors say this approach overcomes limits of simple side-by-side comparisons, including over-rationalization.Source: §9 · p. 9
  • The framework shows effectiveness across multiple datasets and conflict types while maintaining robustness in non-conflict scenarios.The authors identify this combination as valuable for real-world RAG.Source: §9 · p. 9
  • MICRO-ACT is presented as a foundation for future research in knowledge conflict resolution.Source: §9 · p. 9

10 Limitations

The authors identify additional intermediate reasoning steps and an English-focused evaluation as limitations of MICRO-ACT.

10 Limitations

  • MICRO-ACT requires additional intermediate steps to locate conflicts that simple side-by-side comparisons may miss.The authors argue that the modest overhead is justified by improved conflict resolution.Source: §10 · p. 10
  • The evaluation focuses primarily on English contexts, so effectiveness across other languages and cultural contexts remains uncertain.Source: §10 · p. 10
  • The authors frame performance as a priority before efficiency, while noting that the additional overhead is relatively small.Source: §10 · p. 10

12 Ethical Statement

These passages cover the study's ethical practices and supplementary analyses of MICRO-ACT's overhead, conflict reasoning, complexity measures, and errors.

12 Ethical Statement

  • The study filtered sensitive, biased, or potentially harmful examples, used publicly available datasets, and required no additional model training.The authors state that avoiding additional training reduces environmental impact compared with fine-tuning.Source: §12 · p. 10

A.1 Token Usage and Monetary Cost

  • MICRO-ACT uses about 2.8× more input tokens and roughly 1.3× more output tokens than GKP, adding $0.008 per GPT-4o query or $0.0005 per GPT-4o-mini query.Source: §A.1 · p. 15

A.2 Inference Time Overhead

  • GPT-4o averages 1.9s versus GKP's 1.3s, while GPT-4o-mini averages 0.7s versus 0.4s; conflict-free queries avoid the extra latency.The added overhead comes from decomposition steps when MICRO-ACT perceives a conflict.Source: §A.2 · p. 15

B Over-Rationalization Issue and Analysis

B.1 A More Detailed Case Study

  • In the Paul Eugéne Gillon example, iterative decomposition identifies conflicting dates and locations before arriving at a recent French-politician answer.The case contrasts a 2010 appointment and Norway with a more recent appointment and France.Source: §B.1 · p. 16

C Perplexity Calculation

D Mathematical Derivation of Knowledge Representation Transitions

  • The transition derivation models knowledge representations and conflict results probabilistically, then combines their transitions using Bayes' rule.Its components include the prior knowledge representation, verifier evaluation, and model probability of generating new knowledge.Source: D Mathematical Derivation of Knowledge… · p. 17, §D.4 · p. 18

E Error Analysis

  • In a 1,000-example error analysis, context distraction accounted for 63% of errors and over-reliance on retrieved evidence for 37%.The authors suggest improving reasoning clarity in complex scenarios and balancing retrieved and parametric knowledge.Source: E Error Analysis · p. 18

F More Complexity Aspect Analysis

The appendix analyzes how decomposition varies by conflict type and model, describes the action and baseline prompts, and reports human checks on the sampled evaluation data.

G.1 Action Details

F More Complexity Aspect Analysis

  • Temporal and semantic conflicts typically trigger more decomposition steps than misinformation conflicts, and GPT-4o-mini uses more steps across conflict types.The authors associate the higher counts with over-rationalization and the need to investigate conflicts beneath superficially reasonable context.Source: Table 5 · p. 19, F More Complexity Aspect Analysis · p. 19

G.2 Generic Reasoning Models

  • End-to-End QA requests an immediate answer, Few-Shot QA supplies example pairs, and Chain-of-Thought prompts explicit reasoning steps (Wei et al., 2022).Source: §G.2 · pp. 19–20

G.3 Generation-aided Reasoning Models

H Human Evaluation

  • Across datasets, sampled QA accuracy differed by less than 5% from full-dataset accuracy; Fleiss' kappa exceeded 0.80 and annotator accuracy standard deviations stayed below 2%.The authors interpret these checks as supporting the representativeness of the samples and reliability across annotators.Source: H Human Evaluation · p. 20

I Model Descriptions

  • The model descriptions cover GPT-4o, GPT-4o-mini, LLaMA-3.1-70B-Instruct, and LLaMA-3.1-8B-Instruct.GPT-4o is used to assess upper-bound adaptive complexity, while GPT-4o-mini examines how moderate capacity affects decomposition granularity.Source: I Model Descriptions · pp. 20–21

G.4 MICRO-ACT

Loading 2506.05278v2…