Uncovering Graph Reasoning in Decoder-only Transformers with Circuit Tracing
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866912603028062208 |
|---|---|
| author | Dai, Xinnan Lo, Chung-Hsiang Guo, Kai Zeng, Shenglai Luo, Dongsheng Tang, Jiliang |
| author_facet | Dai, Xinnan Lo, Chung-Hsiang Guo, Kai Zeng, Shenglai Luo, Dongsheng Tang, Jiliang |
| contents | Transformer-based LLMs demonstrate strong performance on graph reasoning tasks, yet their internal mechanisms remain underexplored. To uncover these reasoning process mechanisms in a fundamental and unified view, we set the basic decoder-only transformers and explain them using the circuit-tracer framework. Through this lens, we visualize reasoning traces and identify two core mechanisms in graph reasoning: token merging and structural memorization, which underlie both path reasoning and substructure extraction tasks. We further quantify these behaviors and analyze how they are influenced by graph density and model size. Our study provides a unified interpretability framework for understanding structural reasoning in decoder-only Transformers. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_20336 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Uncovering Graph Reasoning in Decoder-only Transformers with Circuit Tracing Dai, Xinnan Lo, Chung-Hsiang Guo, Kai Zeng, Shenglai Luo, Dongsheng Tang, Jiliang Machine Learning Artificial Intelligence Transformer-based LLMs demonstrate strong performance on graph reasoning tasks, yet their internal mechanisms remain underexplored. To uncover these reasoning process mechanisms in a fundamental and unified view, we set the basic decoder-only transformers and explain them using the circuit-tracer framework. Through this lens, we visualize reasoning traces and identify two core mechanisms in graph reasoning: token merging and structural memorization, which underlie both path reasoning and substructure extraction tasks. We further quantify these behaviors and analyze how they are influenced by graph density and model size. Our study provides a unified interpretability framework for understanding structural reasoning in decoder-only Transformers. |
| title | Uncovering Graph Reasoning in Decoder-only Transformers with Circuit Tracing |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2509.20336 |