Parth Asawa

Asawa’s Berkeley research is advised by Matei Zaharia and Joey Gonzalez. Before the doctorate, he earned separate EECS and Business Administration bachelor’s degrees through Berkeley’s M.E.T. program. His earlier engineering work includes Rox, Databricks, Robust Intelligence and Amazon; he also describes enterprise AI consulting. His current research asks how learning can remain sample-efficient and stable over long horizons.[1]

Measure improvement against a reset baseline

In his Continual Learning Bench talk, Asawa compares a system that retains experience across connected tasks with the same system reset between tasks. The difference in reward measures whether accumulated experience helped, controlling for initial model capability and task difficulty. Database tasks, for example, can test whether an agent reuses knowledge of a schema to answer later questions with fewer queries. The setup evaluates useful adaptation across a sequence rather than recall of a single stored fact.[4]

Turn prompt revisions into checks on output

SPADE, which Asawa coauthored, uses a developer’s history of prompt revisions to propose data-quality assertions. It then selects a small set of checks that meet coverage and accuracy requirements. This turns issues noticed during prototyping into functions that can identify bad outputs after deployment, while avoiding an ever-growing collection of redundant or overly strict checks.[5]

Combine diversity with instruction-following refinement

BARE starts synthetic-data generation with a base language model, whose outputs can be more diverse than those of an instruction-tuned model given only a few seed examples. An instruction-tuned model then refines those candidates. The paper studies this two-stage method through downstream code, mathematics and retrieval-related tasks, connecting dataset diversity with the quality needed for training.[6]