# What Structural Inductive Bias Helps Transformers Reason Over Knowledge Graphs? A Study with Tabula RASA

arXiv: https://arxiv.org/abs/2602.02834 · HTML: https://mlsanity.com/paper/2602.02834/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 26 of its 32 references that are in the map.
- First posted on arXiv 2 Feb 2026.

## Builds on (most cited first, 26 of 26)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Graph Attention Networks](https://mlsanity.com/paper/1710.10903/) (2017)
- [Modeling Relational Data with Graph Convolutional Networks](https://mlsanity.com/paper/1703.06103/) (2017)
- [Measuring and Narrowing the Compositionality Gap in Language Models](https://mlsanity.com/paper/2210.03350/) (2022)
- [How Attentive are Graph Attention Networks?](https://mlsanity.com/paper/2105.14491/) (2021)
- [Recipe for a General, Powerful, Scalable Graph Transformer](https://mlsanity.com/paper/2205.12454/) (2022)
- [A Generalization of Transformer Networks to Graphs](https://mlsanity.com/paper/2012.09699/) (2020)
- [Faith and Fate: Limits of Transformers on Compositionality](https://mlsanity.com/paper/2305.18654/) (2023)
- [Rethinking Graph Transformers with Spectral Attention](https://mlsanity.com/paper/2106.03893/) (2021)
- [The Web as a Knowledge-base for Answering Complex Questions](https://mlsanity.com/paper/1803.06643/) (2018)
- [QA-GNN: Reasoning with Language Models and Knowledge Graphs for Question Answering](https://mlsanity.com/paper/2104.06378/) (2021)
- [Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning](https://mlsanity.com/paper/2310.01061/) (2023)
- [Do Transformers Really Perform Bad for Graph Representation?](https://mlsanity.com/paper/2106.05234/) (2021)
- [Structure-Aware Transformer for Graph Representation Learning](https://mlsanity.com/paper/2202.03036/) (2022)
- [The Parallelism Tradeoff: Limitations of Log-Precision Transformers](https://mlsanity.com/paper/2207.00729/) (2022)
- [Query2box: Reasoning over Knowledge Graphs in Vector Space using Box Embeddings](https://mlsanity.com/paper/2002.05969/) (2020)
- [Variational Reasoning for Question Answering with Knowledge Graph](https://mlsanity.com/paper/1709.04071/) (2017)
- [Exphormer: Sparse Transformers for Graphs](https://mlsanity.com/paper/2303.06147/) (2023)
- [Graph Inductive Biases in Transformers without Message Passing](https://mlsanity.com/paper/2305.17589/) (2023)
- [Saturated Transformers are Constant-Depth Threshold Circuits](https://mlsanity.com/paper/2106.16213/) (2021)
- [Representational Strengths and Limitations of Transformers](https://mlsanity.com/paper/2306.02896/) (2023)
- [GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning](https://mlsanity.com/paper/2405.20139/) (2024)
- [Improving Multi-hop Knowledge Base Question Answering by Learning Intermediate Supervision Signals](https://mlsanity.com/paper/2101.03737/) (2021)
- [Where Did the Gap Go? Reassessing the Long-Range Graph Benchmark](https://mlsanity.com/paper/2309.00367/) (2023)
- [Simple Is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation](https://mlsanity.com/paper/2410.20724/) (2024)
- [ReaRev: Adaptive Reasoning for Question Answering over Knowledge Graphs](https://mlsanity.com/paper/2210.13650/) (2022)

## Abstract (arXiv)

What structural inductive bias helps transformers reason over knowledge graphs? Through controlled ablations of a minimal transformer modification with four independently removable components (sparse adjacency masking, edge-type biases, query scaling, value gating), we isolate which structural signals drive multi-hop reasoning. Our finding is sharp: sparse adjacency masking alone accounts for the dominant share of improvement over unmasked transformers (+72.5pp on 3-hop MetaQA, +45.5pp on WebQSP, +53.9pp on CWQ), while learned relation parameters add only modest refinement and can actively hurt without structural guidance. A zero-shot experiment provides architecturally independent corroboration: masking-based attention degrades 4.0x less than relation-specific weights when edge types are held out. The useful inductive bias for multi-hop KGQA is predominantly topological, not relational.