Instance Aware Dynamic Prompting

This direction is split into other directions: dynamic prompt generation (dynamic soft prompting). Its last version stays readable and citable here.

Version 1 of 1205 papers279 conceptsfrozen 2026-10-04

In the standard paradigm of parameter-efficient fine-tuning, "prompt tuning" involves learning a small set of continuous vectors—soft prompts—that are prepended to every input in a dataset. While this is efficient, it assumes a "one-size-fits-all" strategy: the same learned instructions are applied to every query, regardless of its specific nuances.

A growing body of research suggests that this static approach is suboptimal. Instead, researchers are moving toward Instance-Aware Dynamic Prompting, where a model generates or retrieves unique soft prompts tailored to each individual input on the fly. By making the prompt a function of the input, these methods allow frozen large-scale models to adapt to the specific semantic context of a query, significantly closing the performance gap with full-model fine-tuning.

The Shift from Static to Generative Prompts

Traditional prompt tuning learns a fixed "global" prompt for a task. However, as Xianjun Yang and colleagues argue in Dynamic Prompting: A Unified Framework for Prompt Tuning, the optimal prompt length, position, and content often depend on the specific input instance.

The core mechanism of this research direction is the introduction of a lightweight auxiliary network or prompt generator. Instead of storing a fixed vector, the system stores the weights of a small network that takes the input embedding as its own input and outputs a custom soft prompt. For example, Ali Shakeri and Lishan Yang developed FedDPG, which uses a tiny sub-network to generate context-aware prompts. This allows the model to "steer" its frozen backbone more precisely for each unique sample, which is particularly effective in diverse or non-identical data distributions.

Capturing Local Context and Structure

The intuition behind instance-aware prompting is that the "instructions" needed to solve a problem often reside within the local context of the data point itself. This is especially clear in structured data like graphs.

In Graph Neural Prompting with Large Language Models, Yijun Tian and Huan Song demonstrate that instead of a generic task prompt, a model can generate a unique prompt by encoding the local neighborhood of a knowledge graph relevant to a specific query. This "evidence-conditioned" approach ensures that the injected information is directly relevant to the individual problem. Similarly, Jiazheng Li and Jundong Li show in Instance-Aware Graph Prompt Learning that discrete, instance-specific prompts are essential for handling the high diversity found in molecular and citation datasets.

This concept extends to multimodal tasks as well. In In-context Prompt-augmented Micro-video Popularity Prediction, Jiao Li uses a cross-attention mechanism to transform features from retrieved similar items into learnable tokens. This allows the model to adapt its internal representations based on external semantic priors specific to that video.

Why It Works: Dynamic Computational Depth

Beyond just providing better instructions, dynamic prompting can be viewed as a way to provide "extra thinking time" or computational depth where it is most needed.

In Deliberation in Latent Space via Differentiable Cache Augmentation, researchers use input-dependent latent embeddings to inject information directly into the model's KV-cache. This provides a dynamic computational path that adapts to the complexity of the input. Similarly, Tianxiang Sun introduced Late Prompt Tuning (LPT), which conditions prompts on the intermediate hidden states of the model. By generating prompts based on how the model is already processing the input, LPT allows for a more surgical intervention in the model's reasoning process.

Efficiency and Robustness

One might assume that generating a new prompt for every input would be computationally expensive. However, these auxiliary generators are typically extremely lightweight. Muling Wu and colleagues' Token-wise Prompt Tuning (TPT) retrieves specific tokens from a bank, updating only 0.035% of parameters while outperforming full fine-tuning.

Furthermore, this instance-level adaptation provides a natural defense against "overrefusal" and distribution shift. Kartik Sharma and Yiqiao Jin developed Sysformer, which dynamically tailors system prompts to specific user queries in the embedding space. This allows the model to adjust its safety instructions on the fly, ensuring that a request is only refused if it is truly harmful in that specific context, rather than relying on a rigid, static safety prompt.

By moving from static task-level vectors to dynamic, input-dependent generation, this research direction enables frozen foundation models to act with the precision of a fine-tuned specialist while maintaining the efficiency and generality of a prompt-tuned generalist.

Papers in this version

Showing 60 of 205 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_instance_aware_dynamic_prompting_424fe42d,
  title = {Instance Aware Dynamic Prompting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/instance-aware-dynamic-prompting/rdr_424fe42dcd5a4a92a36384e39ab3b5a8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_instance_aware_dynamic_prompting_424fe42d_v1,
  title = {Instance Aware Dynamic Prompting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/instance-aware-dynamic-prompting/rdr_424fe42dcd5a4a92a36384e39ab3b5a8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}