FutureBench tests how trust cues in prompts shape LLM answers
A new arXiv framework measures knowledge-context conflict in six models, finding that prompts telling models to prioritize retrieved passages can displace correct answers.
In plain words
Large language models can answer from their own training knowledge or from passages you give them. FutureBench is a reusable test set that checks what happens when the two conflict. Its authors report that a prompt asking a model to verify passages and prioritize them often made models adopt false answers, especially for events the models already knew about.
What was announced
FutureBench, introduced on arXiv by its authors as a renewable framework for evaluating knowledge-context conflict, combines timestamped questions, controlled passage triplets, and no-context probes. Across 3,629 questions and six models, the authors report that Verify, a prompt combining checking with passage priority, raised false-answer adoption in all five models tested on pre-cutoff events, mainly by displacing answers that were correct under a Baseline prompt that allowed both passage evidence and model knowledge. On post-cutoff events, only o4-mini increased adoption, by 13.7 percentage points.
A four-prompt comparison on 902 post-cutoff questions found that passage priority without checking raised adoption, while moving to Verify lowered it in all five models. A prompt called Adversarial-Aware, which warns of misleading passages and prioritizes model knowledge, restored 82.5 to 98.9 percent of displaced pre-cutoff answers with under one point of change in true-passage accuracy, according to the paper. The framework is designed to support repeated evaluation on fixed cohorts and extension to new events as models change.
What it does not show
- The evidence is the paper's abstract only; full methodology, model list details, and experimental settings are not available in the supplied passages.
- All reported figures are the authors' own claims and have not been independently confirmed.
- The abstract does not name all six models or explain why only five were tested on each event type.
For security and IT teams
- Scope
- The findings cover six models evaluated by the paper's authors on timestamped question cohorts, not a general audit of deployed systems.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.