← News

Agents & MCPSources: arXiv · 19 Nov 2026

CaDO framework targets faulty causal reasoning datasets with automated diagnosis and repair

A multi-agent system from a new arXiv paper finds and fixes annotation errors and structural imbalances in causal reasoning datasets, reporting gains over five cleaning baselines.

In plain words

Datasets used to teach models causal reasoning often contain wrong labels and unbalanced question types. CaDO is a proposed pipeline where software agents translate questions into formal logic, flag suspect samples, and generate corrected or new data. The authors report it beats existing cleaning methods in most tests.

What was announced

The framework, described in an arXiv abstract by its authors, works in three stages. A Formalizer Agent converts natural-language questions into formal causal queries, which the authors say makes retrieval and verification more explicit. Independent Critic Agents then perform a dual-level diagnosis: identifying errors in individual samples while also checking query-type coverage and label distributions for dataset-level imbalances. Finally, reusable repair operators correct faulty samples, while Explorer and Expander Agents generate new samples the authors describe as diverse and semantically consistent.

In reported experiments, CaDO outperformed the strongest of five data cleaning baselines in 16 of 18 settings, with downstream model accuracy improving by up to 9.1%. The authors also state that at least 95% of flagged samples were genuinely defective, and that on the CLadder dataset it modified 92.2% fewer correct labels than an LLM-based judge. Code and data are published on GitHub. These figures come from the paper's own evaluation and have not been independently verified.

What it does not show

  • Evidence is limited to the paper's abstract; full methodology, datasets, and evaluation details were not available for review.
  • All performance figures are self-reported by the authors and not independently confirmed.

For security and IT teams

Status
The authors state code and data are publicly available on GitHub.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.