IC-177GPT-4o mini, GPT-4o, and Llama-3-8B all over-rely on incorrect external context, producing wrong answers at high rates when the context conflicts with their internal knowledge

Yukun Huang, Sanxing Chen, Hongyi Cai, Bhuwan Dhingra

SourceTo Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts

When each model is given a question paired with an incorrect context that contradicts its internal knowledge, the model's direct input augmentation (DIA) baseline follows the wrong context in the vast majority of cases. Across six QA datasets, accuracy given false contexts (acc_f) under DIA ranges from 2.3% to 24.7% for GPT-4o, 2.3% to 12.5% for GPT-4o mini, and 2.7% to 14.3% for Llama-3-8B, meaning the models are misled by incorrect context in over 75% of cases. This over-reliance is consistent across all three models and all six datasets, establishing it as a reproducible failure condition rather than an isolated error.

Evidence
correlational
Key metric
DIA acc_f: GPT-4o mini 12.5/2.3/12.0/12.3/11.0/10.3; GPT-4o 12.5/5.7/24.7/23.3/12.3/12.7; Llama-3-8B 9.7/2.7/10.7/12.0/14.3/12.0 (RedditQA/FreshQA/ClashEval/TriviaQA/PopQA/NaturalQA); overall situated faithfulness (ov) under DIA: 51.7/53.9/51.0 respectively
Caveat
Evaluation uses 300 sampled examples per dataset (176 for RedditQA); incorrect contexts are a mix of LLM-generated and human-written (RedditQA) errors, so the failure rate may vary with context type.
Model
GPT-4o mini, Llama 3 8B
Concepts
Failure mode
Datasets
RedditQA [eval], FreshQA [eval], ClashEval [eval], TriviaQA [eval], PopQA [eval], Natural Questions / NaturalQA [eval]
Related work
ClashEval [compared-to], TACS [compared-to]
Related findings
IC-178, IC-179, IC-180
Extraction
automatic-extraction