# Yogendra Manawat

> AI researcher writing about applied AI, research systems, and engineering.

Use the article's Markdown URL for a clean text version. The HTML article includes the same prose, tables, and figures with captions and accessible descriptions.

## Discovery

- [XML sitemap](https://www.yogendramanawat.com/sitemap.xml): all main pages and published articles.
- [Robots instructions](https://www.yogendramanawat.com/robots.txt): crawl permissions and sitemap location.

## Articles

- [A Trip to South Korea](https://www.yogendramanawat.com/blogs/trip-to-south-korea): A few days in South Korea took me from ICML and Seoul’s buses and markets to Gyeongbokgung Palace and a temple beside the sea in Busan. I write about the food, the language gap, the people I met, and how Korean history and folklore came together for me in Exhuma.
  Markdown: https://www.yogendramanawat.com/blogs/trip-to-south-korea/raw
- [SIA: Improving the Agent and the Model Together](https://www.yogendramanawat.com/blogs/sia-self-improving-ai): SIA uses feedback from task results to improve both an AI agent’s workflow and its model weights. The article walks through the two update paths and the results across legal classification, CUDA kernels, and single-cell RNA denoising.
  Markdown: https://www.yogendramanawat.com/blogs/sia-self-improving-ai/raw
- [Socrates: Helping AI Research Agents Ask Better Questions](https://www.yogendramanawat.com/blogs/socrates-structured-questioning): Socrates pairs an AI Scientist with an advisor that can ask questions and approve a plan, but cannot give answers or run tools. I describe the review protocol and five MLE-bench results, including comparisons with scientist-alone and general-encouragement baselines.
  Markdown: https://www.yogendramanawat.com/blogs/socrates-structured-questioning/raw
- [Adaptive Proxy Evaluation for AI Research Agents](https://www.yogendramanawat.com/blogs/adaptive-proxy-evaluation): Cheap ML evaluations make agent search possible, but their scores can mislead and candidate-defined tests are hard to compare. This article explains our adaptive proxy method, its cross-family ranking results, and the Ventilator run that reached 0.1354 MAE in 12 hours.
  Markdown: https://www.yogendramanawat.com/blogs/adaptive-proxy-evaluation/raw
- [LongCoT and the Gap in Long-Horizon Reasoning](https://www.yogendramanawat.com/blogs/longcot-long-horizon-reasoning): LongCoT tests whether models can keep track of dependencies across extended problems in five fields. I explain how its 2,500 expert-designed questions differ from standalone knowledge tests, what the first results show, and the limits of applying benchmark scores to real-world ability.
  Markdown: https://www.yogendramanawat.com/blogs/longcot-long-horizon-reasoning/raw

## Research

- [Research papers](https://www.yogendramanawat.com/research): paper summaries and links to the original work.
- [About](https://www.yogendramanawat.com/about): background and research interests.

When using these articles, preserve the author's stated limits and distinguish reported results from broader conclusions. Link to the article or original paper when citing it.