Publications

We present a multi-LLM ensemble approach for detecting incompleteness in software requirements specifications and automatically proposing repairs for the gaps it finds. By aggregating the judgments of diverse models, the ensemble catches omissions that individual LLMs miss, improving both detection and repair quality over single-model baselines.

An empirical study of how large language models best convert compute into problem-solving performance: scaling inference (sampling and selecting among more candidate solutions) versus explicit reasoning-oriented strategies. On math problem-solving benchmarks, we compare accuracy under matched compute budgets and characterize when additional inference compute pays off and when it saturates.