AI Paper Atlas

Research map / Large language models / Agents, code and math

Large Language Models as Topological Thinkers: A Benchmark on Graph Persistent Homology

Hao Li, Hao Wan, Yixue Huang, Yuzhou Chen, Yulia Gel, Hao Jiang

ICML 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add supervised fine-tuning on small graphs (ablation)LLaMA-3.1-8B-Instruct · 1D Simplex Counting (medium graphs)accuracy: 0.02 -> 0.20Improved
add supervised fine-tuning on small graphs (ablation)LLaMA-3.1-8B-Instruct · Component Merge Timeaccuracy: 0.09 -> 0.43Improved
add supervised fine-tuning on small graphs (ablation)LLaMA-3.1-8B-Instruct · Component Count Under Filtrationaccuracy: 0.16 -> 0.66Improved
transfer model tuned on small graphs to large graphs (ablation)LLaMA-3.1-8B-Instruct · 1D Simplex Counting (large graphs)accuracy: 0.005 -> 0.235Improved
use text-style graph input instead of matrix-based input (ablation)LLMs · LLM4PH prompt ablationaccuracyImproved
use graph-theoretic phrasing instead of topological phrasing (ablation)LLMs · LLM4PH prompt ablationaccuracyMixed
use PH terminology instead of graph-theoretic terms (ablation)GPT-4o · Component Count Under FiltrationaccuracyHurt
add few-shot CoT or rule-emphasized prompting (ablation)LLMs · Non-Uniform Filtration GenerationrankNo clear effect
increase number of sampled ego-Facebook subgraphs from 10 to 150 (ablation)GPT-4.1-mini · 1D Simplex Countingaccuracy: 14% -> 24%Improved
let model generate persistence diagrams instead of code (ablation)LLM-PH hybrid pipeline · BZR classificationaccuracyHurt
let model compare diagram distances in hybrid pipeline (ablation)Claude and DS-V3 · BZR classificationaccuracy: 0.840Improved
use direct classification instead of PH-guided filtration designLLMs · real-world graph classificationaccuracy: 0.98 to 1.0Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy