AdviSD trains a small consultant to guide giant models, raising function-call accuracy without extra inferences

A new research paper by Rishab Agrawal, Heiji Kui, Shasha Li, Shanshan Wu, and Sirkan Arik introduces a technical framework called “AdviSD” (AdviSD), which relies on training a small language model to act as a consultant that guides a large, frozen execution model through natural-language advice and instructions. The core problem the paper addresses is how to train the consultant: typically it learns from task rewards and subsequent evaluation feedback, yet the suggested corrections after task completion may appear persuasive but do not necessarily alter the practical execution trajectory, while training the model on all those corrections can negatively affect its later decisions in other contexts because the parameters are shared.

The research team mathematically demonstrated that in models with shared parameters, corrections that do not sufficiently support useful guidance constrain the learning trajectory and weaken its effectiveness.The solution offered by the AdviSD framework combines result-based reinforcement learning with selective self-distillation from a consultant version conditioned on feedback.The process begins with a reflective mechanism that proposes specific corrections, after which the consultant assigns evaluation scores to the executive model’s responses with and without the guidance, using the numerical difference to select decisions worthy of training and supervision. The most notable operational advantage is that this method does not require access to the executive model’s probabilities nor does it require running additional inferences.

Experimental results showed remarkable efficiency for a consultant model built on the Qwen3-8B architecture (an 8 billion-parameter model) when guiding large models such as Gemini (Google’s multimodal model) and Claude. The new framework outperformed the GRPO-Consultant algorithm by 4.2-6.4 percentage points on the advanced function-calling benchmark BFCL-v3, and by 3.9-5.1 points on the EnvScaler agent-environment evaluation. It also surpassed a matched-size random-selection baseline, confirming the validity of the filtering and selection rule, and the trained consultant successfully generalized its capabilities to tasks beyond the training set and smoothly transitioned across different families and versions of executive models.

This development is directly important for AI engineering and agent-building teams in the Gulf, Egypt, and the Levant, where companies incur high compute and invocation costs by relying exclusively on major commercial models for every inference and planning step.Separating the independent guidance layer and training it locally at eight billion parameters changes the cost and technical sovereignty equation.This approach allows an engineer in Riyadh, Dubai, or Cairo to train an open-source consultant model on local infrastructure to guide closed-source executive models with greater precision, without needing access to their weights or internal probabilities, reducing paid API usage and giving projects the flexibility to switch the execution model when needed without losing accumulated guidance logic.

This trajectory reveals a fundamental shift in the architecture of independent agents, as performance improvements no longer depend on model size or waiting for provider updates. Today you can build a lightweight intelligence layer dedicated to guidance and self-distillation, achieving higher effectiveness with minimal resource consumption.