Research
Giving language models a richer interface for customization
A position paper studies how vector prompts absorb supervision and influence model computation, then proposes exposing them as an additional interface for model customization.
Model customization begins at an interface. That interface defines the inputs an application can supply, the feedback it can use, and the artifacts it can preserve between revisions. Choosing the shape of those inputs is a systems decision with consequences for how customization develops over time.
A position paper co-authored by Haolin Chen explores vector prompts as an additional interface. A vector prompt is a set of continuous inputs that conditions the model’s computation. The paper combines a proposal for model APIs with diagnostic experiments on how these inputs learn and influence attention.
Follow the learning curve
The study holds a LLaMA3-8B Instruct backbone and the SST-5 sentiment task fixed while varying the amount of labeled supervision available for prompt optimization. Vector prompts continue to gain from additional examples across the range studied.
The experiments use gradient-based prompt tuning to explore the capacity of the interface. This setup provides a controlled view of what the learned inputs can encode. A learning curve then becomes an interface measurement: it shows how additional task information translates into behavior under the chosen experimental conditions.
Inspect where the control enters
A second analysis examines attention patterns with 32 learned vector tokens prepended to the task input. The paper visualizes representative middle and later layers. Task tokens repeatedly attend to the vector prompts across these layers, giving the researchers a view of how the learned control signal participates in computation.
This analysis connects a familiar customization artifact to an internal mechanism. The prompt has a defined location in the sequence, an optimization procedure, and an observable pattern of use inside the network.
Design an interface teams can evaluate
The proposal asks model providers to expose vector-prompt inputs alongside existing customization options. It also discusses inference-only optimization, deployment constraints, and the responsibilities of an API that accepts continuous control signals.
For application teams, this suggests useful evaluation questions. How much labeled feedback can the interface absorb? How many control tokens does a customization require? How stable is the behavior across task variants and repeated revisions? Recording these properties would make customization artifacts easier to compare and maintain as a model-backed application evolves.
Research by Liangwei Yang, Shiyu Wang, Haolin Chen, and collaborators at Salesforce AI Research. This article covers the paper’s interface proposal and diagnostic experiments.