Cornell's Malleable Prompting Turns Vague AI Style Requests Into Sliders
A new interactive technique turns vague prompt phrases like funny or concise into sliders and toggles that steer LLM generation at decoding time.
ProTakeaways
- Malleable Prompting turns preference words like funny or concise into GUI sliders, toggles, dropdowns, and steppers.
- A training-free decoding algorithm modulates token log-probabilities based on widget values, working on any open model.
- The method roughly doubled steering strength versus prompt-based intensity instructions (β=0.61 vs 0.32) in tests.
- Attribution highlights predicted the direction of token probability changes 81.4% of the time, versus 67.2% baseline.
- In a N=12 user study, target-matching quality scores rose from 3.31 to 3.61 out of 4.
- Sliders accounted for 52% of created widgets and over 85% of user adjustments during tasks.
Anyone who has spent time nudging a chatbot toward a specific tone knows the frustration. Ask for something funnier, and it goes overboard. Ask for less sarcasm, and now it reads like a press release. A team from Cornell, UIUC, UT Austin, and Loyola Maryland has proposed a way out that treats subjective preferences as GUI controls rather than words to be argued with. The technique is called Malleable Prompting, and it will be presented at UIST 2026.
The core move is to reify preference expressions in natural language prompts into GUI widgets (sliders, dropdowns, toggles) that users can directly configure to steer generation, while visualizing each control's influence on the output. Behind the interface sits a new decoding algorithm that changes token probabilities during generation based on each widget's value.

Two gulfs that natural language cannot cross
The paper frames the problem as two classic HCI gaps. There is a gulf of execution, where users may not have clear preferences at the outset, and articulating nuanced ones in language is notoriously difficult, forcing people to lean on vague modifiers like "more concise" or "less formal." Then there is a gulf of evaluation, where iterations pile up in a chat log and it becomes hard to tell which prompt tweak caused which text change.
Prior hybrid systems have tried widgets before, but usually with predefined dimensions drawn from a database, or with pretrained LoRA modules or steering vectors. Those approaches need training or curated data, which limits how flexible the controls can be.