ToolSense: A Diagnostic Framework for Auditing Parametric Tool Knowledge in LLMs

2026-05-23 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

ToolSense, an open-source diagnostic framework, evaluates Large Language Models' (LLMs) parametric tool knowledge, addressing the limitation of existing benchmarks that may not reveal true tool understanding. It automatically generates three benchmarks: a Realistic Retrieval Benchmark (RRB) with varied ambiguity, an MCQ probing benchmark, and a QA probing benchmark. Applied to ToolBench's ~47k tools and five parametric model configurations, ToolSense uncovered a "knowledge-retrieval dissociation." Models, despite strong performance on verbose, fully-specified ToolBench benchmarks, experienced a significant performance collapse of approximately 50-64 percentage points on RRB queries, often falling below embedding-model baselines. Furthermore, these models frequently scored near-random on factual probes, indicating a lack of genuine semantic understanding despite effective retrieval.

Key takeaway

For Machine Learning Engineers deploying LLM agents over extensive tool catalogs, relying solely on standard retrieval benchmarks is insufficient. Your models may exhibit a knowledge-retrieval dissociation, performing poorly on realistic queries despite high in-distribution scores. You should integrate diagnostic frameworks like ToolSense to uncover true semantic understanding gaps. Consider using flat token formats and LoRA fine-tuning to mitigate knowledge erosion during retrieval training, ensuring your agents genuinely comprehend their tools for robust real-world performance.

Key insights

Parametric tool retrieval in LLMs often shows a knowledge-retrieval dissociation, failing on realistic queries despite high benchmark scores.

Principles

• Parametric tool retrieval often learns surface-level query patterns.
• Constrained decoding can mask genuine tool knowledge deficits.
• Stage 2 retrieval fine-tuning can degrade semantic tool understanding.

Method

ToolSense generates three benchmarks: Realistic Retrieval Benchmark (RRB) with tiered queries, MCQ probing for discriminative factual knowledge, and QA probing for inferential factual knowledge. It also defines an Internalization Score (IS@k) for trie-dependency.

In practice

• Audit LLM tool knowledge using ToolSense's diagnostic benchmarks.
• Prefer flat token formats for parametric tool encoding.
• Apply LoRA during Stage 2 fine-tuning to preserve semantic knowledge.

Topics

• LLM Agents
• Parametric Tool Retrieval
• Diagnostic Benchmarking
• Knowledge-Retrieval Dissociation
• LoRA Fine-tuning
• Virtual Tokens

Code references

• SAP/toolsense

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

Related on AIssential

• Diagnosing Knowledge Gaps in LLM Tool Use: An Agentic Benchmark for Novel API Acquisition — LLM tool use with novel APIs requires specific knowledge components, with usage examples being critical, and retrieval/tuning playing complementary roles.
• How LoRA Remembers? A Parametric Memory Law for LLM Finetuning — The Parametric Memory Law quantifies LoRA's exact memory capacity in LLMs, linking loss reduction to effective parameters and sequence length.
• AI Doesn’t Search. It Remembers. — LLMs remember knowledge through trained weights, not by searching, akin to human semantic memory.
• LLMSurgeon: Diagnosing Data Mixture of Large Language Models — LLMSurgeon diagnoses LLM pretraining data mixtures from generated text, enabling post-hoc auditing without direct data access.
• ParaTool: Shifting Tool Representations from Context to Parameters — ParaTool shifts LLM tool representations from context to parameters, reducing overhead and hallucination risks.
• LLM-Powered Deep Parsing for Industrial Inventory Search — LLM-powered deep parsing extracts structured data from messy industrial text for improved search and deduplication.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.