--- license: apache-2.0 base_model: Qwen/Qwen3-8B library_name: transformers pipeline_tag: text-generation language: - en tags: - tool-use - function-calling - agent - reinforcement-learning - qwen3 --- # StateAlign-8B StateAlign-8B is a multi-turn tool-use agent trained from [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) with **StateAlign**, introduced in *StateAlign: Internalizing State-Aware Decision-Making for Reliable Multi-Turn Tool Agents via Reinforcement Learning*. **Code:** [github.com/ZhexiLu/StateAlign](https://github.com/ZhexiLu/StateAlign) (training code, example data and the BFCL evaluation script). A tool agent should decide its next action from the current task state: what information is known or missing, whether the available tools can perform the request, and which parts of the task are already done. StateAlign constructs task states from tool-use trajectories (execute, ask the user, retrieve, decline, stop) and pairs states that differ in one condition. Training combines outcome rewards from interaction with a token-level state-contrastive signal that compares the same generated reasoning and actions under paired states. ## Results | Benchmark | Qwen3-8B | StateAlign-8B | |---|---|---| | BFCL v4 multi-turn (overall) ↑ | 45.5 | **55.2** | | ToolSandbox (similarity) ↑ | 73.4 | **79.3** | | When2Call (accuracy) ↑ | 66.4 | **72.4** | | SimpleToolHalluBench, no tool (NTA) ↓ | 59.8 | **56.1** | | SimpleToolHalluBench, distractor tool (DT) ↓ | 28.2 | **25.0** | BFCL is the mean of six evaluations (temperature 0.001, `Qwen/Qwen3-8B-FC` handler, YaRN factor 2 for a 65,536-token context). See the paper for all settings and baselines. ## Usage The model keeps the Qwen3 architecture and chat template, so it can be used like Qwen3-8B. **vLLM (OpenAI-compatible server with tool calling):** ```bash vllm serve ZhexiLu/StateAlign-8B \ --enable-auto-tool-choice --tool-call-parser hermes \ --reasoning-parser qwen3 ``` For contexts beyond 32,768 tokens, enable YaRN as for Qwen3: ```bash vllm serve ZhexiLu/StateAlign-8B ... \ --rope-scaling '{"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":32768}' \ --max-model-len 65536 ``` **Transformers:** ```python from transformers import AutoModelForCausalLM, AutoTokenizer name = "ZhexiLu/StateAlign-8B" tokenizer = AutoTokenizer.from_pretrained(name) model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="auto", device_map="auto") tools = [{"type": "function", "function": { "name": "cancel_order", "description": "Cancel an order by its ID.", "parameters": {"type": "object", "properties": {"order_id": {"type": "string"}}, "required": ["order_id"]}}}] messages = [{"role": "user", "content": "Please cancel my order."}] text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True) inputs = tokenizer(text, return_tensors="pt").to(model.device) out = model.generate(**inputs, max_new_tokens=1024) print(tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)) ``` ## Training - Base model: Qwen3-8B. - Method: StateAlign (GRPO with outcome rewards plus token-level state contrasts scored by the fixed base model). - Fine-tuning: LoRA (rank 64, alpha 128, all attention and MLP projections), 400 steps, merged into the base weights for this release. - Data: task states constructed from public tool-use trajectories; see the paper for details. ## License Released under Apache 2.0, following the Qwen3-8B license. ## Citation ```bibtex @misc{lu2026statealign, title = {StateAlign: Internalizing State-Aware Decision-Making for Reliable Multi-Turn Tool Agents via Reinforcement Learning}, author = {Lu, Zhexi and Chaudhury, Subhajit and Pedapati, Tejaswini and Murugesan, Keerthiram and Yu, Lei}, year = {2026} } ```