Research

Publications

Detecting and Correcting Reasoning Failures: A Controlled Negative Result for a Residual-Stream Monitor

Arnav Kakani (lead author), Mrinal Agarwal. The 6th Workshop on Mathematical Reasoning and AI (MATH-AI), NeurIPS 2026. Poster.

I tested whether drift in a language model’s hidden state can flag a math error while the solution is still being written, and whether rolling back and asking the model to recheck fixes it. Under answer-based scoring and proper controls, it does neither.

OpenReview · PDF · Supplementary · Code and data · Blog post

@inproceedings{kakani2026detecting,
  title     = {Detecting and Correcting Reasoning Failures: A Controlled Negative Result for a Residual-Stream Monitor},
  author    = {Kakani, Arnav and Agarwal, Mrinal},
  booktitle = {The 6th Workshop on Mathematical Reasoning and AI (MATH-AI), NeurIPS 2026},
  year      = {2026},
  url       = {https://openreview.net/forum?id=AxOXuIaLWj}
}

Current interests

I am looking into world models, optimizers, and evaluating AI harnesses and how effective they are.

Profiles

OpenReview