IC-338Factuality enhancement methods (DoLa, ICD, ITI, TruthX, CD) cause large and consistent declines in context-faithfulness of LLaMA2-7B-Chat and LLaMA2-13B-Chat

Baolong Bi, Shenghua Liu, Yiwei Wang, Lingrui Mei, Junfeng Fang, Hongcheng Gao, Shiyu Ni, Xueqi Cheng

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

The paper applies five factuality enhancement methods to LLaMA2-Chat and evaluates their effect on the model's ability to follow counterfactual contextual knowledge via in-context editing (ICE) and MELLO on MQuAKE. All methods produce a marked decline in editing accuracy compared to the unmodified baseline, with no exception across any method, shot setting, or model size. The largest single decrease is 67.9% (TruthX on LLaMA2-7B-Chat, CoT 10-shot: 26.5 vs. 82.5 baseline). Under MELLO, declines are even more severe, with most exceeding 40% relative to baseline. The paper explains this via logit and hidden-state analysis: factuality enhancement shifts probability mass toward parametric knowledge tokens and constrains hidden-state dispersion, making the model resistant to incorporating new contextual information.

Evidence
correlational
Key metric
LLaMA2-7B-Chat ICE 3-shot: DoLa 52.3 (↓22.4), ICD 50.1 (↓26.9), ITI 51.9 (↓21.8), TruthX 26.7 (↓61.1); CoT 10-shot: TruthX 26.5 (↓67.9). LLaMA2-13B-Chat ICE 3-shot: DoLa 58.3 (↓20.0), ICD 59.5 (↓23.2), ITI 64.3 (↓14.0), CD 55.2 (↓29.5). MELLO: LLaMA2-7B-Chat TruthX 22.6 (↓58.5), ITI 31.5 (↓42.2); LLaMA2-13B-Chat ITI 34.5 (↓45.9), DoLa 35.6 (↓44.2)
Caveat
The paper uses only LLaMA2-Chat models; generalization to other model families is not tested. The MQuAKE-cf-3k-v2 dataset was specifically chosen to address internal knowledge conflicts, which may not represent all context-faithfulness scenarios.
Model
Llama 2 / Llama 2 base Llama 2 7B Chat / Llama-2-chat-7b, Llama-2-13B-Chat
Concepts
Failure mode
Datasets
MQuAKE [eval]
Methods
DoLa [primary], ICD [primary], Inference-Time Intervention / ITI / Neuron intervention (pinning activation) [primary], TruthX [primary], ICE [eval]
Related work
DoLa [builds-on], Inference-Time Intervention / ITI / Neuron intervention (pinning activation) [builds-on], TruthX [builds-on]
Related findings
IC-339
Extraction
automatic-extraction