Research map / Reinforcement learning / RL from feedback
STAR: Similarity-guided Teacher-Assisted Refinement for Super-Tiny Function Calling Models
Jiliang Ni, Jiachen Pu, Zhongyi Yang, Jingfeng Luo, Conggang Hu
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 Sim-RL refinement after CKD distillation (ablation) | CKD student (Qwen3-0.6B) · ACEBench Normal | accuracy: 39.00 -> 53.00 | Improved |
| add Sim-RL refinement after CKD distillation (ablation) | CKD student (Qwen3-0.6B) · BFCLv3 Overall | accuracy: 49.56 -> 51.70 | Improved |
| replace SFT-think initialization with CKD before Sim-RL (ablation) | STAR Qwen3-0.6B · BFCLv3 Overall | accuracy: 50.41 -> 51.70 | Improved |
| use Sim-RL similarity reward instead of binary reward (ablation) | CKD student · ACEBench Normal | accuracy: 35.70 -> 53.00 | Improved |
| use Sim-RL similarity reward instead of ToolRL reward (ablation) | CKD student · BFCLv3 Overall | accuracy: 48.59 -> 51.70 | Improved |
| use Sim-RL similarity reward instead of SWiRL PRM reward (ablation) | CKD student · ACEBench Normal | accuracy: 40.30 -> 53.00 | Improved |
| use CKD instead of stabilized RKL as distillation before RL (ablation) | Qwen3-0.6B · ACEBench Normal (w/ RL) | accuracy: 41.30 -> 53.00 | Improved |
| use CKD instead of stabilized AKL as distillation before RL (ablation) | Qwen3-0.6B · ACEBench Normal (w/ RL) | accuracy: 49.00 -> 53.00 | Improved |
| use CKD instead of top-k FKL as distillation before RL (ablation) | Qwen3-0.6B · ACEBench Normal (w/ RL) | accuracy: 50.00 -> 53.00 | Improved |
| use SFT instead of no fine-tuning | Qwen3-0.6B · ACEBench Normal Summary | accuracy: 27.20 -> 2.10 | Hurt |
| increase top-k from 10 to 100 in CKD (ablation) | CKD+Sim-RL · ACEBench Normal (w/ RL) | accuracy: 49.20 -> 53.00 | Improved |
| increase top-k from 100 to 1000 in CKD (ablation) | CKD+Sim-RL · ACEBench Normal (w/ RL) | accuracy: 53.00 -> 52.20 | No clear effect |
Closest papers by meaning
- Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning (NeurIPS 2024 · similarity 0.87)
- AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · similarity 0.87)
- Offline Training of Language Model Agents with Functions as Learnable Weights (ICML 2024 · similarity 0.87)
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.87)
- Tina: Tiny Reasoning Models via LoRA (ICLR 2026 · similarity 0.87)
- Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.86)
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.
