Seeing through the Conflict: Transparent Knowledge Conflict Handling in Retrieval-Augmented Generation

TL;DR

The paper tackles hallucinations and inconsistent outputs in retrieval-augmented generation by introducing Transparent Conflict Resolution (TCR), a lightweight plug-and-play framework that disentangles semantic relevance from factual consistency, estimates self-answerability, and injects interpretable conflict signals into generation via soft prompts. TCR deploys dual encoders for semantic and factual information, trained with separate contrastive objectives, and uses three scalar signals to steer decoding through a soft-token prompt with dynamic, SNR-based weighting. Across diverse benchmarks, TCR improves conflict detection, knowledge-gap recovery, and resistance to misleading context while adding only a small parameter and computational footprint, with signals that align with human judgments. The approach enhances transparency and reliability in knowledge-intensive tasks and offers avenues for extension to multimodal and federated scenarios.

Abstract

Large language models (LLMs) equipped with retrieval--the Retrieval-Augmented Generation (RAG) paradigm--should combine their parametric knowledge with external evidence, yet in practice they often hallucinate, over-trust noisy snippets, or ignore vital context. We introduce TCR (Transparent Conflict Resolution), a plug-and-play framework that makes this decision process observable and controllable. TCR (i) disentangles semantic match and factual consistency via dual contrastive encoders, (ii) estimates self-answerability to gauge confidence in internal memory, and (iii) feeds the three scalar signals to the generator through a lightweight soft-prompt with SNR-based weighting. Across seven benchmarks TCR improves conflict detection (+5-18 F1), raises knowledge-gap recovery by +21.4 pp and cuts misleading-context overrides by -29.3 pp, while adding only 0.3% parameters. The signals align with human judgements and expose temporal decision patterns.

Figures (9)

• Figure 1: Knowledge conflicts in a typical RAG system. Our method disentangles semantic relevance from factual consistency, detects conflicts, and injects lightweight prompt signals to steer RAG models toward faithful, knowledge-aligned generation.
• Figure 2: Sentence‐embedding clusters for all-MiniLM-L6-v2 (top) vs. E5-large-v2 (bottom). Rows 1 & 3: two PCA views of 3-D space—paraphrase (blue), contradiction (red), unrelated (green). Rows 2 & 4: boundary regions (yellow) highlight transition zones. MiniLM (384 d) forms a single blob; the larger E5 (1024 d, 24 layers) begins to separate truth from meaning.
• Figure 3: 2D scatter plot of semantic-factual similarity space. Left: Sentence pairs show clear separation of paraphrases (top right), contradictions (bottom right), and unrelated sentences (bottom left) using just one linear projection layer. Contour lines show KDE boundaries. Right: Example of annotated transitions (other types perform similarly well).
• Figure 4: Overview of our Transparent Conflict Resolution (TCR) framework for Retrieval-Augmented Generation (RAG). Given a query, TCR identifies knowledge conflicts through semantic and factual similarity vectors, generating explicit conflict signals. These signals, along with self-answerability estimation, are integrated using soft prompt tuning to guide the Language Model's generation toward enhanced factual consistency and interpretability.
• Figure 5: Residual conflict-detection errors for TCR. Dot colour follows a blue - white - red scale : warmer hues denote denser error pockets. Missed conflicts cluster just below the factual threshold ($y\!\approx\!0.35$); a thin false-positive band appears near $y\!\approx\!0.45$. Dashed lines mark the current decision boundary $x{>}0.65,\;y{<}0.40$.

Theorems & Definitions (2)

• Theorem 1: Noise-Robustness Error Bound
• proof : Proof sketch