IC-339Factuality enhancement methods produce inconsistent and modest improvements in factual accuracy on LLaMA2-Chat, with some metrics declining below baseline

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 evaluates five factuality enhancement methods on TruthfulQA (multiple-choice accuracy MC1/MC2/MC3) and FactScore (response ratio, number of facts, fact score) for LLaMA2-7B-Chat and LLaMA2-13B-Chat. While most methods show some improvement on at least one metric, the gains are unstable: several methods show declines on specific metrics compared to the unmodified baseline. For example, DoLa on LLaMA2-7B-Chat drops MC1 from 37.6 to 32.9, and CD on LLaMA2-13B-Chat drops MC1 from 37.7 to 28.2. The authors note that these methods 'do not show consistent enhancements across all metrics' and that 'some metrics exhibit declines compared to the baseline.'

Evidence
correlational
Key metric
LLaMA2-7B-Chat TruthfulQA MC1: baseline 37.6, DoLa 32.9, ICD 46.3, ITI 37.0, TruthX 54.2. LLaMA2-13B-Chat TruthfulQA MC1: baseline 37.7, DoLa 37.3, CD 28.2, ICD 45.6, ITI 38.9. FactScore score: LLaMA2-7B-Chat baseline 63.8, DoLa 61.3, ICD 66.3, ITI 62.4, TruthX 65.3.
Caveat
The paper notes that evaluation settings follow prior studies (Chuang et al. 2023; Li et al. 2024b; Zhang et al. 2023a), and the TruthfulQA multiple-choice format may not capture all aspects of factuality.
Model
Llama 2 / Llama 2 base Llama 2 7B Chat / Llama-2-chat-7b, Llama-2-13B-Chat
Datasets
TruthfulQA / TruthfulQA MC1 [eval], FactScore [eval]
Methods
DoLa [primary], ICD [primary], Inference-Time Intervention / ITI / Neuron intervention (pinning activation) [primary], TruthX [primary]
Related work
DoLa [builds-on], Inference-Time Intervention / ITI / Neuron intervention (pinning activation) [builds-on], TruthX [builds-on]
Related findings
IC-338
Extraction
automatic-extraction