Framing
How it works
A model processes the prompt as a token sequence and conditions the distribution over next tokens on the whole context. Framing modifies that context along several axes: (1) role/persona โ giving the model an identity ("you are a legal expert") shifts style and scope; (2) perspective and stance โ asking it to take a given viewpoint; (3) context/scenario โ embedding the task in a situation (real, hypothetical or fictional); (4) question form โ positive vs negative wording, assumptions and the confidence stated in the user's claim. Because safety tuning (RLHF, filters) is also context-conditioned, changing the framing can shift behavior either toward better relevance or outside the learned guardrails (an unrestricted persona, a fictional scenario). Framing operates at inference time, without changing the model's weights.
Problem solved
Language model outputs are sensitive to the form of a query, not only its content. Prompt framing names and organizes this lever: it lets one deliberately steer answer quality, perspective and format, while also exposing the risk โ answer inconsistency across equivalent wordings and safety bypasses via careful framing.
Implementation
Semantically equivalent prompts with different framing can yield conflicting answers, hurting reproducibility.
Hypothetical, fictional or roleplay framing can be a jailbreak vector bypassing safety guardrails.
Confident/biased framing of a user's claim nudges the model to agree rather than judge the facts.
Evolution
Tversky and Kahneman show that equivalent options framed as a gain or a loss lead to different choices โ the cognitive precursor of framing.
Jailbreak work shows that persona/roleplay framing sharply raises the rate of harmful model completions.
Studies show the same task structure under different framings yields materially different model answers.
Hyperparameters (configurable axes)
The kind of frame applied to the prompt: role/persona, perspective, scenario (real/hypothetical/fictional).
Positive vs negative wording of a semantically equivalent question.
How much background/scenario is supplied alongside the actual question.