Research map / Large language models / Retrieval and knowledge

Is Factuality Enhancement a Free Lunch For LLMs? Better Factuality Can Lead to Worse Context-Faithfulness

ICLR 2025 · Large language models › Retrieval and knowledge · approach: Knowledge-augmented question answering · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that factuality enhancement methods (DoLa, CD, ICD, ITI, TruthX) applied to LLaMA2 chat models markedly reduce context-faithfulness on knowledge editing (MQuAKE ICE and MeLLo), with declines up to 60-70%.

Problem

Factuality enhancement methods that improve how LLMs convey parametric knowledge may make them overconfident in parametric memory, hindering their ability to follow counterfactual or updated context.

Approach

Uses knowledge editing accuracy (in-context editing and MeLLo on MQuAKE) as a proxy for context-faithfulness and tests training-free factuality enhancement methods against it, then analyzes logits and hidden states of new-knowledge versus parametric-knowledge tokens.

Key results

• TruthX on LLaMA2-7B-chat drops ICE editing accuracy on MQuAKE by up to 67.9% relative to baseline (e.g., 3-shot 78.3 vs 26.7 range), with all factuality methods declining across settings.
• Factuality methods yield unstable factuality gains: e.g., DoLa and ITI reduce TruthfulQA MC1 and FActScore on some models, while TruthX raises TruthfulQA MC2 to 73.9 on LLaMA2-7B-chat.
• With MeLLo, most factuality-enhanced models show declines exceeding 40% relative to baseline, with larger models showing larger drops.

Measured effects reported in the paper

Surprising findings

• Factuality enhancement methods sometimes fail to improve factuality consistently, yet still cause larger drops in context-faithfulness.
• Chain-of-thought reasoning tends to hurt factuality-enhanced models on the editing task, unlike the baseline where it helps.

Limitations

• Evaluation limited to LLaMA2-7B/13B chat models and a few factuality enhancement methods.
• Context-faithfulness proxied solely by knowledge editing accuracy on MQuAKE.
• Knowledge token capturing relies on string matching and a specific stubborn subset of 2,000 instances.

Experimental setup

Datasets: MQuAKE-CF-3k-v2 MQuAKE-stubborn HaluEval TruthfulQA FActScore

Benchmarks: TruthfulQA FActScore MQuAKE-CF-3k-v2

Metrics: MC1 MC2 MC3 FActScore editing accuracy relative decrease

Baselines: DoLa CD ICD ITI TruthX

Builds on: DoLa ITI TruthX MeLLo MQuAKE

Closest papers by meaning

• Fine-Tuning Language Models for Factuality (ICLR 2024 · similarity 0.89)
• Long-form factuality in large language models (NeurIPS 2024 · similarity 0.89)
• Do Automatic Factuality Metrics Measure Factuality? A Critical Evaluation (NeurIPS 2025 · similarity 0.89)
• FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows" (ICLR 2025 · similarity 0.88)
• ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
• Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.88)