publications
Knowledge Discovery & Management Lab (KDM Lab)
Knowledge Discovery & Management Lab (KDM Lab) started out at DA-IICT as a part of the research efforts by Prof. Sourish Dasgupta and his students and collaborators in 2015. The lab broadly works on natural language understanding (NLU) and the evaluation of state-of-the-art large language models (LLMs) w.r.t. their reasoning and personalization capabilities. We love designing theoretical frameworks and experimental setups for LLM evaluation. We also love dabbling with emerging methodologies, such as neuro-symbolic models, and understanding how that can be applied to specific NLU tasks. We highly emphasize lateral thinking (thereby creativity) and during problem-solving love connecting the dots across pretty much everything - linguistics. economics, political science, social sciences, and even food! Our motto is “Don’t bring the baggage and boredom of specific schools of thought!”my research
Evaluate whether large language models (LLMs) can detect the non-triviality of a proposed research question given a body of prior literature. Non-triviality is operationalized as how much background knowledge (and how influential that background is) reduces the novelty or difficulty of the target research question. Using a citation-aware dataset derived from IdeaBench, we will generate year-by-year background sets and measure how model judgments change as more prior work accumulates. This gives a quantifiable signal of whether a model can recognize incremental vs. genuinely novel research directions.
Evaluating Epistemic Non-Triviality in Scientific Hypothesis Generation
Under Review at ACL 2026 · Knowledge & Discovery Lab, DAIICT and Oak Ridge National Laboratory · Gandhinagar, India and Tennessee, USA
- Proposed DISCERN, a diagnostic framework to assess epistemic non-triviality in LLM-based hypothesis generators using year-indexed citation disclosure and model-agnostic signals: semantic similarity, citation graph, and temporal baselines.
- Benchmarked GPT-4o, O4-mini, Magistral, and GPT-OSS-20B on IdeaBench across three prompt conditions, measuring alignment via Spearman, Kendall, Pearson, and KL divergence with robustness validated on NLPeer V2 scores.