PKU-ML/GRASP-4B

1---2license: apache-2.03language:4- en5base_model:6- Qwen/Qwen3-4B-Thinking-25077library_name: transformers8---9 10 11 12 13<p align="center">14    <img src="https://raw.githubusercontent.com/PKU-ML/GRASP/main/logo-new.png" width="15%"/>15<p>16 17# PKU-ML/GRASP-4B18 19## 📊 Overview20 21Integrating graph knowledge into Large Language Models (LLMs) via passive representation faces critical bottlenecks: limited context windows, unreliable numerical computation, and structural hallucinations. 22To solve this, we propose **GRASP** (Graph Reasoning via Agentic Solving and Probing), shifting the paradigm from passive ingestion to proactive agentic exploration. 23By interleaving **Neighbor Retrieval** for on-demand probing with **Code Interpreter** as a deterministic solver, GRASP enables LLMs to autonomously navigate and compute over complex topologies. 24We employ a staged reinforcement learning strategy (GRPO) that transitions from visible tuning to a structure-blind environment, forcing the agent to develop genuine topological awareness. 25Evaluated on multi-domain graph reasoning benchmarks, our 4B model achieves a 53.06% average performance boost, surpassing SOTA baselines like DeepSeek-V3.2 and successfully generalizing to unseen tasks, 26with high potential for tackling sampling on million-node graphs and solving Hard-level LeetCode graph problems.27 28 29 30## 📌 Key Takeaways31 321️⃣ **Agentic Probing over Passive Ingestion**.33We propose GRASP (Graph Reasoning via AgenticSolving and Probing), shifting the paradigm from passive ingestion to proactive agentic exploration. By interleaving Neighbor Retrieval (Eyes 👀) for on-demand probing with Code Interpreter (Hands 🙌) as a deterministic solver, GRASP enables LLMs to autonomously navigate and compute over complex topologies.34 352️⃣ **Structure-Blind RL Training**.36We employ a staged reinforcement learning strategy (GRPO) that transitions from visible tuning to a structure-blind environment, forcing the agent to develop genuine topological awareness.37 383️⃣ **From Million-Node Graphs to Hard LeetCode**.39Evaluated on multi-domain graph reasoning benchmarks, our 4B model achieves a 53.06% average performance boost, surpassing SOTA baselines like DeepSeek-V3.2 and successfully generalizing to unseen tasks, with high potential for tackling sampling on million-node graphs and solving Hard-level LeetCode graph problems.40 41 42 43 44## 🌊 Evaluation on Graph Reasoning Benchmarks45 46 47| Model            | Arxiv     |PubMed     |Products   | WikiCS    | fb15k237   |wn18rr     |TSG-Bench   |ExplaGraphs |Erdős       |RealErdős   |Average     |48|------------------|-----------|-----------|-----------|-----------|------------|-----------|------------|------------|------------|------------|------------|49| Qwen3-4B-Thinking|51.00      |25.00      |21.00      |29.00      |16.00       |13.00      |62.00       |45.00       |38.80       |7.11        |30.79       |50| GPT-4o           |52.00      |43.00      |72.00      |24.00      |52.00       |24.00      |72.00       |77.00       |40.60       |18.07       |47.46       |51| DeepsSeek-V3.2   |65.00      |47.00      |70.00      |79.00      |65.00       |26.00      |**88.00**   |**99.00**   |83.60       |66.44       |68.90       |52| GRASP-4B         |**73.00**  |**90.00**  |**77.00**  |**88.00**  |**82.00**   |**67.00**  |85.00       |97.00       |**91.00**   |**88.57**   |**83.85**   |53 54 55 56 57 58 59## Quickstart60 61The code of Qwen3 has been in the latest Hugging Face `transformers` and we advise you to use the latest version of `transformers`.62 63With `transformers<4.51.0`, you will encounter the following error:64```65KeyError: 'qwen3'66```67 68The following contains a code snippet illustrating how to use the model generate content based on given inputs. 69```python70from transformers import AutoModelForCausalLM, AutoTokenizer71 72model_name = "PKU-ML/GRASP-4B"73 74# load the tokenizer and the model75tokenizer = AutoTokenizer.from_pretrained(model_name)76model = AutoModelForCausalLM.from_pretrained(77    model_name,78    torch_dtype="auto",79    device_map="auto"80)81 82# prepare the model input83prompt = "Give me a short introduction to large language model."84messages = [85    {"role": "user", "content": prompt}86]87text = tokenizer.apply_chat_template(88    messages,89    tokenize=False,90    add_generation_prompt=True,91)92model_inputs = tokenizer([text], return_tensors="pt").to(model.device)93 94# conduct text completion95generated_ids = model.generate(96    **model_inputs,97    max_new_tokens=819298)99output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 100 101# parsing thinking content102try:103    # rindex finding 151668 (</think>)104    index = len(output_ids) - output_ids[::-1].index(151668)105except ValueError:106    index = 0107 108thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")109content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")110 111print("thinking content:", thinking_content) # no opening <think> tag112print("content:", content)113 114```115 116## Agentic Use117 118For the specific tool configuration and agentic usages of GRASP, please refer to our [example](https://github.com/PKU-ML/GRASP/blob/main/evaluation/example.py) on Github.119 120 121 122## Citation123 124If you find our work helpful, feel free to give us a cite.125 126```127 128```