| --- |
| license: apache-2.0 |
| language: |
| - en |
| base_model: |
| - Qwen/Qwen3-4B-Thinking-2507 |
| library_name: transformers |
| --- |
| |
|
|
|
|
|
|
| <p align="center"> |
| <img src="https://raw.githubusercontent.com/PKU-ML/GRASP/main/logo-new.png" width="15%"/> |
| <p> |
| |
| # PKU-ML/GRASP-4B |
|
|
| ## 📊 Overview |
|
|
| Integrating graph knowledge into Large Language Models (LLMs) via passive representation faces critical bottlenecks: limited context windows, unreliable numerical computation, and structural hallucinations. |
| To solve this, we propose **GRASP** (Graph Reasoning via Agentic Solving and Probing), shifting the paradigm from passive ingestion to proactive agentic exploration. |
| By 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. |
| We 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. |
| Evaluated 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. |
|
|
|
|
|
|
| ## 📌 Key Takeaways |
|
|
| 1️⃣ **Agentic Probing over Passive Ingestion**. |
| We 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. |
|
|
| 2️⃣ **Structure-Blind RL Training**. |
| We 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. |
|
|
| 3️⃣ **From Million-Node Graphs to Hard LeetCode**. |
| Evaluated 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. |
|
|
|
|
|
|
|
|
| ## 🌊 Evaluation on Graph Reasoning Benchmarks |
|
|
|
|
| | Model | Arxiv |PubMed |Products | WikiCS | fb15k237 |wn18rr |TSG-Bench |ExplaGraphs |Erdős |RealErdős |Average | |
| |------------------|-----------|-----------|-----------|-----------|------------|-----------|------------|------------|------------|------------|------------| |
| | 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 | |
| | GPT-4o |52.00 |43.00 |72.00 |24.00 |52.00 |24.00 |72.00 |77.00 |40.60 |18.07 |47.46 | |
| | 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 | |
| | GRASP-4B |**73.00** |**90.00** |**77.00** |**88.00** |**82.00** |**67.00** |85.00 |97.00 |**91.00** |**88.57** |**83.85** | |
|
|
|
|
|
|
|
|
|
|
|
|
| ## Quickstart |
|
|
| The code of Qwen3 has been in the latest Hugging Face `transformers` and we advise you to use the latest version of `transformers`. |
|
|
| With `transformers<4.51.0`, you will encounter the following error: |
| ``` |
| KeyError: 'qwen3' |
| ``` |
|
|
| The following contains a code snippet illustrating how to use the model generate content based on given inputs. |
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model_name = "PKU-ML/GRASP-4B" |
| |
| # load the tokenizer and the model |
| tokenizer = AutoTokenizer.from_pretrained(model_name) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_name, |
| torch_dtype="auto", |
| device_map="auto" |
| ) |
| |
| # prepare the model input |
| prompt = "Give me a short introduction to large language model." |
| messages = [ |
| {"role": "user", "content": prompt} |
| ] |
| text = tokenizer.apply_chat_template( |
| messages, |
| tokenize=False, |
| add_generation_prompt=True, |
| ) |
| model_inputs = tokenizer([text], return_tensors="pt").to(model.device) |
| |
| # conduct text completion |
| generated_ids = model.generate( |
| **model_inputs, |
| max_new_tokens=8192 |
| ) |
| output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() |
| |
| # parsing thinking content |
| try: |
| # rindex finding 151668 (</think>) |
| index = len(output_ids) - output_ids[::-1].index(151668) |
| except ValueError: |
| index = 0 |
| |
| thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n") |
| content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n") |
| |
| print("thinking content:", thinking_content) # no opening <think> tag |
| print("content:", content) |
| |
| ``` |
|
|
| ## Agentic Use |
|
|
| For 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. |
|
|
|
|
|
|
| ## Citation |
|
|
| If you find our work helpful, feel free to give us a cite. |
|
|
| ``` |
| |
| ``` |
