A graph serialization tweak doubled multi-hop accuracy for local models in a 10-format benchmark
Immediate-Cake6519 · reddit · 2026-07-27
A developer reports that the serialization format alone can dramatically change local-model performance on graph reasoning tasks.
Key findings
- Tested 10 graph serialization formats including JSON, GraphML, RDF variants, and edge lists
- Verbose formats waste about 70% of tokens on syntax alone
- On an 8K local model, that cuts the effective graph budget to roughly one-third
- Multi-hop accuracy moved from about 40% to 80% purely by changing the format
- Tabular or relational layouts consistently outperformed nested markup
What they built
- ISONGraph, a lightweight property-graph format optimized for LLM comprehension
- About 70% fewer tokens with 92% traversal accuracy
- Fully offline support for local models
- MIT licensed implementations in Python, JS/TS, Rust, Go, C++, and C#
More from Research
- CMU’s AI Science Foundry wins $20M NSF grant for a national lab network — jmuiuc · 2026-07-27
- Study of 775,323 scientists finds LLMs pushed research into more distant fields — i_dg23 · 2026-07-27
- IGGT4D proposes a streaming Transformer for real-time 4D scene understanding — liuziwei7 · 2026-07-27
- Study of 1.02M PRs Finds Agentic Code Review Cuts Time, but Not Quality — rohanpaul_ai · 2026-07-27
- Anthropic’s Opus 5 hits 30% on ARC-AGI-3, but loses its edge on a held-out puzzle suite — NathanpmYoung · 2026-07-27
- Paul David’s 106-page essay traces how open science emerged in the Scientific Revolution — michael_nielsen · 2026-07-27