Research map / Reinforcement learning / RL from feedback
On Effectiveness and Efficiency of Agentic Tool-calling and RL Training
Tong Liu, Cheng Qian, Matej Cief, Yuan He, Daniele Dan, Nikolaos Aletras, Gabriella Kazai
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add online pre-rollout filtering and max-variance rollout down-sampling to GRPO | GRPO, Qwen2.5-3B-Instruct single-turn · BFCL single-turn tool-calling | training speedup at comparable performance: 1.7x speedup | Improved |
| add online pre-rollout filtering and max-variance rollout down-sampling to GRPO | GRPO, Qwen3-4B multi-turn · BFCL multi-turn tool-calling | training speedup at comparable performance: 2.6x speedup | Improved |
| use native multi-turn template instead of context concatenation (ablation) | Qwen3-8B, Qwen3-4B, Qwen2.5-7B-Instruct · BFCL multi-turn | accuracy: about 6-8% gain | Improved |
| retain thinking history across turns (ablation) | Qwen3 models · BFCL multi-turn | accuracy: about 3-5% for Qwen3-8B and Qwen3-4B | Improved |
| add multi-turn instructions to the default BFCL system prompt (ablation) | Qwen3-4B · BFCL multi-turn | accuracy: 22.7 -> 37.2 | Improved |
| train on multi-turn data instead of single-turn data (ablation) | Qwen3-4B · BFCL multi-turn | accuracy: 22.7 -> 15.9 | Hurt |
| train on single-turn data (ablation) | Qwen3-4B · BFCL multi-turn | accuracy: 22.7 -> 20.2 | No clear effect |
| evaluate with multiple random seeds instead of one run (ablation) | BFCL evaluation of five models · BFCL multi-turn | score variance: up to ~3% deviation in multi-turn | Mixed |
| apply RL tool-calling training | Qwen3-4B · ACEBench English overall | overall accuracy: 65.4 -> 77.5 | Improved |
| apply RL tool-calling training on top of stronger system prompt | Qwen3-4B · BFCL multi-turn | accuracy: 37.2 -> 39.4 | Improved |
| apply tool-calling RL training | Qwen3-4B · MMLU, HellaSwag, TruthfulQA, Winogrande | accuracy: MMLU 68.3 -> 68.3 | No clear effect |
Closest papers by meaning
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.91)
- ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning (NeurIPS 2026 · similarity 0.91)
- Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.90)
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.90)
- Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning (ICLR 2026 · similarity 0.90)
- LLM Agents Already Know When to Call Tools - Even Without Reasoning (NeurIPS 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
