Do LLMs Make More Mistakes If They Do Not Believe the Input Data?
Source
Peter Kochelka, Ale\v{s} Manuel Pap\'a\v{c}ek, Vojt\v{e}ch Dvo\v{r}\'ak, Ond\v{r}ej Du\v{s}ek
Author
Peter Kochelka, Ale\v{s} Manuel Pap\'a\v{c}ek, Vojt\v{e}ch Dvo\v{r}\'ak, Ond\v{r}ej Du\v{s}ek
Date
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
RAG — Retrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.
Why it matters
Finds only a weak context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →-memory conflict effect when LLMs are given counterfactual context, suggesting faithfulness issues in RAGRetrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.Full definition → systems may stem less from belief conflict than assumed.
Key quotes
“Contrary to our expectations, we observe only a weak context-memory conflict on the human-annotated sample.”
“We also find that a suboptimal choice of LLM judge would lead to overestimating the strength of the context-memory conflict.”
articleLarge Language Models and Language Server Protocol: a match made in contextAlessandro Schena, Ilgiz Mustafin, Julia Kotovich
articleSWORD: Wikidata-based Distortions Reveal Hidden Cross-Lingual Inconsistencies in LLM Factual Error RejectionSanghyeok Park, Minji Kang, Hosung Kwak, Jinhyuk Yun
articleA Removal Based Approach to Improve LLM Faithfulness at Test-TimeQinglan Luo, S M A Nahian, John Guttag, S. Mazdak Abulnaga, Katie Matton