Get Started
Home
Topics
Search
Library
Research questionHow can we train and evaluate LLM agents for tool use across single- and multi-turn workflows with serial or parallel calls?Tool-use studies often represent function-call interactions differently and cover uneven trajectory structures. This makes it difficult to align training data with benchmarks or compare agents when workflows vary in turn depth and call execution.
AI Agents
Evaluation & Benchmarks
LLM Pretraining & Post-training
Natural Language Processing
Latest papersRecent research connected to this question, newest first.UniToolCall: Unifying Tool-Use Representation, Data, and Evaluation for LLM AgentsThe source concerns structured function calling across single-hop and multi-hop, single-turn and multi-turn interactions, including serial and parallel execution with cross-turn dependencies. Its evidence uses a 22k-plus tool pool, more than 390k training instances, seven unified public benchmarks, and evaluation at function-call, turn, and conversation levels; reported model results come from fine-tuning Qwen3-8B, including a distractor-heavy Hybrid-20 setting.research paper · Sep 2, 2026
Related questions
How can LLM agents reuse execution traces without losing temporal and outcome-dependent behavior?How can LLM agents reconcile conflicting user instructions, stored knowledge, and live tool observations before acting?How can LLM agent configurations be jointly optimized when prompts, tools, and parameters interact?How can open-source LLM agents reliably chain multiple tool calls over live public APIs?