RASP-Tuner: Retrieval-Augmented Soft Prompts for Context-Aware Black-Box Optimization in Non-Stationary Environments

2026-04-20 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

RASP-Tuner is a novel black-box optimization system designed for non-stationary environments where optimal configurations shift with external context. It addresses the high adaptation costs of traditional optimizers like Gaussian Processes (GP) by employing a retrieval-augmented soft prompt approach. The system decomposes online tuning into identifying regime proxies, predicting short-horizon loss with a mixture-of-experts surrogate, and adapting primarily in a low-dimensional prompt subspace. RASP-Tuner utilizes a RealErrorComposer to map heterogeneous streaming metrics to a single differentiable training target. Evaluated on nine synthetic non-stationary benchmarks, an adversarial-context check, and three real-world tabular streams, RASP-Tuner matched or improved cumulative regret compared to GP-UCB and CMA-ES on seven of nine synthetic tasks at horizon T=100. It also demonstrated 8-12 times lower wall-clock time per step than sliding-window GP-UCB on identical hardware.

Key takeaway

For research scientists developing online optimization systems in dynamic, non-stationary environments, RASP-Tuner offers a method to significantly reduce computational overhead while maintaining or improving performance. You should investigate integrating retrieval-augmented soft prompts and mixture-of-experts surrogates to manage context-dependent shifts, especially when full model refits are computationally prohibitive. This approach can lead to more efficient and scalable black-box optimization.

Key insights

RASP-Tuner optimizes black-box systems in non-stationary environments using retrieval-augmented soft prompts for efficient context adaptation.

Principles

• Decompose online tuning into regime identification and loss prediction.
• Adapt primarily in low-dimensional prompt subspaces.
• Map heterogeneous metrics to a single differentiable target.

Method

RASP-Tuner identifies regime proxies via past context retrieval, predicts loss with a mixture-of-experts surrogate using parameters, context, and a soft prompt, and adapts in a low-dimensional prompt subspace, with full surrogate updates only on error spikes.

In practice

• Apply retrieval-augmented prompts for dynamic black-box optimization.
• Use EMA-stabilized logistic scores for metric aggregation.
• Consider prompt-based adaptation for non-stationary tuning.

Topics

• RASP-Tuner
• Black-Box Optimization
• Non-Stationary Environments
• Retrieval-Augmented Learning
• Soft Prompts

Code references

• Jian-Lang/RAGPT
• ritaranx/Collab-RAG
• SEU-VIPGroup/RAAP

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Large-scale empirical tuning and comparison of default optimizers for variational inference — A selection of 5 adaptive stochastic optimizers reliably approximates optimal performance for black-box variational inference without extensive tuning.
• RASFT: Rollout-Adaptive Supervised Fine-Tuning for Reasoning — RASFT calibrates expert SFT guidance with on-policy rollouts to prevent overfitting and activate LLM's intrinsic reasoning.
• Softmax as Linear Attention in the Large-Prompt Regime: a Measure-based Perspective — Softmax attention behaves linearly in the large-prompt, infinite-token limit, simplifying theoretical analysis.
• Optimizers in Deep Learning: From Gradient Descent to Adam — Optimizers guide neural network training by efficiently updating parameters to minimize loss, evolving from basic GD to adaptive Adam.
• Continual Fine-Tuning with Provably Accurate and Parameter-Free Task Retrieval — Proteus offers provably accurate, parameter-free task retrieval for continual fine-tuning by leveraging distinct representation signatures.
• Parameter Tuning with Generalization Guarantees for GPU-Accelerated Linear Programming — Data-driven algorithm design provides generalization guarantees for hyperparameter tuning in complex optimization solvers.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.