Year: 2026
Type: conference-paper
Abstract: As large language models (LLMs) increasingly serve as the backbone of agentic systems with tool-use capabilities, predicting the correctness, latency, and cost for a given tool-calling query is essential for efficient model evaluation, selection, and routing.Existing approaches predominantly adopt a semantics-first perspective, without capturing the complex interaction patterns inherent in tool-calling queries and without jointly accounting for cost and latency.We observe that query-model interaction matrices in tool-calling scenarios exhibit a low-rank structure, indicating that behavioral outcomes are governed by latent factors beyond surface semantics.We leverage this observation to propose a dual-pathway framework that decomposes interaction matrices into lowdimensional latent spaces for multi-objective prediction.The Capability-First (CF) pathway discovers latent structure via matrix factorization and then projects natural language queries into latent space.The Joint Capability-Semantic (JCS) pathway concurrently optimizes the text encoder and model embeddings, implicitly shaping query embeddings through capability-semantic interactions.Experiments on 75 LLMs demonstrate strong performance in predicting correctness, latency, and cost.Further experiments show that our method enables few-shot unseen-model generalization, parameter-efficient training, and reveals structured latent behavioral patterns.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Award: G5784019134
OpenAlex ID: W7224254027