LLM Flow Processes for Text-Conditioned Regression

2026-01-05

arXiv.org

AI Summary

This work proposes a general framework that addresses the limitations of existing meta-learning regression methods, which struggle to effectively integrate expert priors with textual metadata, and the constrained regression performance of large language models (LLMs) despite their rich semantic capabilities. The approach uniquely combines the text-conditional probability outputs of LLMs with neural diffusion or flow-matching processes through a product-of-experts mechanism, enabling joint sampling between a binning-based probabilistic density “expert” model and a diffusion generative model. By doing so, the method preserves the semantic understanding of LLMs while significantly enhancing regression accuracy. Empirical evaluations across multiple benchmark tasks demonstrate consistent superiority over both standalone LLMs and neural process–based approaches, validating the effective contribution of text-conditioned knowledge to regression performance.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Natural Language Processing: Generation

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Meta-learning methods for regression like Neural (Diffusion) Processes achieve impressive results, but with these models it can be difficult to incorporate expert prior knowledge and information contained in metadata. Large Language Models (LLMs) are trained on giant corpora including varied real-world regression datasets alongside their descriptions and metadata, leading to impressive performance on a range of downstream tasks. Recent work has extended this to regression tasks and is able to leverage such prior knowledge and metadata, achieving surprisingly good performance, but this still rarely matches dedicated meta-learning methods. Here we introduce a general method for sampling from a product-of-experts of a diffusion or flow matching model and an `expert'with binned probability density; we apply this to combine neural diffusion processes with LLM token probabilities for regression (which may incorporate textual knowledge), exceeding the empirical performance of either alone.

Problem

• text-conditioned regression
• large language models
• meta-learning
• prior knowledge
• neural processes

Innovation

• product-of-experts
• diffusion processes
• flow matching
• large language models
• text-conditioned regression