tao-hpu/cog-canvas-benchmark

CogCanvas Benchmark

Paper | GitHub

CogCanvas is a training-free framework for maintaining long-term memory in LLM conversations. This repository contains the Controlled Benchmark used to evaluate CogCanvas's ability to extract verbatim-grounded cognitive artifacts and perform temporal and multi-hop reasoning.

Dataset Summary

The CogCanvas benchmark consists of two primary evaluation sets:

1. Standard Benchmark: Focuses on verbatim extraction and recall with 5000 dialogue turns and 400 planted facts.
2. Multi-hop Reasoning Benchmark: Focuses on complex reasoning over dialogue history with 2500 turns and 176 specific questions.

Available configurations: standard_turns, standard_facts, multihop_turns, multihop_questions.

Sample Usage

from datasets import load_dataset

# Standard benchmark (5000 turns, 400 planted facts)
standard = load_dataset("tao-hpu/cog-canvas-benchmark", "standard_turns")

# Multi-hop reasoning benchmark (2500 turns, 176 questions)
multihop = load_dataset("tao-hpu/cog-canvas-benchmark", "multihop_turns")

Citation

@article{an2026cogcanvas,
  title={CogCanvas: Verbatim-Grounded Artifact Extraction for Long LLM Conversations},
  author={An, Tao},
  journal={arXiv preprint arXiv:2601.00821},
  year={2026},
  url={https://arxiv.org/abs/2601.00821}
}