# ArXiv paper proposes using frozen numerical foundation models to guide few-shot multitask optimization

> ArXiv CS.LG posted a preprint introducing ICG-MTO, a framework that uses a frozen numerical foundation model to improve inter-task coupling estimation in few-shot multitask optimization.

- **Topic**: Research
- **Published**: 2026-09-24T05:30:44.620Z
- **Canonical URL**: https://highsignal.sh/stories/arxiv-paper-proposes-using-frozen-numerical-foundation-models-to-guide-few-shot-34c95259

## Why It Matters

Multitask optimization often fails when the budget is too small to estimate relationships between tasks, causing negative transfer; the paper's bet is that a frozen model's in-context inference can fill that early data gap, though the evidence is the authors' own experiments on benchmarks and a robot-arm control task.

## Key Findings & Analysis

### What the paper claims

- ArXiv CS.LG published a preprint (arXiv:2609.25836v1) introducing In-Context Guidance Multitask Optimization (ICG-MTO), a framework for few-shot multitask optimization. The paper says conventional methods suffer from inaccurate inter-task relationship estimation under limited evaluation budgets, leading to negative transfer — ArXiv CS.LG (Machine Learning), the paper's authors, as stated in the abstract. Details are the authors' claims, not independent findings; this is an unverified preprint. The framework has three stages, per the abstract: build an algorithm-specific in-context query from evaluated solutions; use a frozen foundational model to infer a guidance signal describing predictive relationships among tasks; translate that signal into algorithm-specific guidance for maximum-a-posteriori coupling estimation. The authors say this regularizes the early, data-scarce stage and gradually relinquishes control as task-specific observations accumulate.

### Instantiations and evaluation

- The authors instantiate ICG-MTO in multitask Bayesian optimization as ICG-MTBO, using directional fitness-class queries to guide inter-task coupling estimation — ArXiv CS.LG, the paper's authors. They report a second instantiation in MFEA-II, using decision-space-overlap queries to guide random mating probability estimation — same source. They report experiments across synthetic benchmarks and a real-world robot arm control problem, plus evaluations under different acquisition functions and evolutionary multitasking, and describe the results as showing effectiveness and generality for few-shot multitask optimization — the authors' evaluation, on their own chosen tasks and baselines, not independently reproduced.

## Primary Sources & Citations

- [In-Context Guidance: Learning Inter-Task Synergies via Numerical Foundational Models for Few-Shot Multitask Optimization](https://arxiv.org/abs/2609.25836) — *ArXiv CS.LG (Machine Learning)* (Reporting)

---

[← Back to front page](https://highsignal.sh/) | [Daily Brief](https://highsignal.sh/brief) | [All stories](https://highsignal.sh/latest)