DARA PhD fellowship to support research on more reliable AI reasoning
Sayambhu Sen has been awarded a prestigious PhD fellowship from the Danish Advanced Research Academy (DARA) to pursue research into how large language models can reason more reliably over long and complex tasks.
Sayambhu will carry out his PhD project, Beyond Text Traces: Graph-based State Representations for Reliable LLM Reasoning, under the supervision of Akhil Arora and Davide Mottin. The project explores new ways of representing reasoning processes in AI systems, with the goal of making them more transparent, robust, and easier to correct when errors occur.
Large language models have made remarkable progress in recent years and can now solve increasingly complex tasks through multi-step reasoning. However, they remain unreliable in long horizon tasks especially for mistakes early in the reasoning process, which can silently propagate through later steps and lead to incorrect conclusions.
To address this challenge, the project proposes replacing traditional text-based reasoning traces with a structured graph representation that explicitly captures claims, evidence, and dependencies. This makes it possible to identify and repair only the affected parts of a reasoning process when new information emerges, rather than restarting the entire chain from scratch.
The research will investigate how such graph-based reasoning states can be constructed, maintained, and updated through minimal-change repair mechanisms. The project will also evaluate how well these systems cope with unexpected changes and interventions during reasoning tasks.
The fellowship is awarded by the Danish Advanced Research Academy (DARA), a national PhD academy funded by the Novo Nordisk Foundation. DARA supports outstanding early-career researchers through competitive fellowships, interdisciplinary training, international collaboration, and professional development opportunities.