Parth Asawa is a computer science PhD researcher at UC Berkeley and the project lead behind Continual Learning Bench, which measures whether AI systems become more effective through experience. His research challenges a basic limitation of deployed models: they can perform sophisticated work yet repeatedly rediscover the same codebase, database, or user preferences.
Advised by Matei Zaharia and Joey Gonzalez, Asawa is a Laude Open Research Resident. He previously earned bachelor’s degrees in electrical engineering and computer sciences and business administration through Berkeley’s Management, Entrepreneurship, & Technology program, working with researchers Shreya Shankar and Aditya Parameswaran. His engineering experience includes Rox, Databricks, Robust Intelligence, and Amazon.
Asawa coauthored SPADE, which derives data-quality checks for language-model pipelines from developers’ prompt revisions, and BARE, which combines base-model generation with instruction-tuned refinement to produce synthetic training data. His subsequent research focuses on making adaptation measurable, efficient, and durable.
Making AI systems learn after deployment
- Continual Learning Bench: Asawa leads a collaboration involving Berkeley, Snorkel AI, and the University of Wisconsin–Madison that evaluates AI systems across connected tasks in software engineering, database exploration, signal processing, epidemiology, strategic games, and sales forecasting. Its stateful-versus-stateless gain compares performance when a system retains experience against its performance when reset between tasks, separating actual learning from a model’s starting strength. Early benchmark results found that straightforward in-context learning outperformed several specialized memory-management systems.
- Adaptation under concept drift: A database agent should need fewer SQL queries as it learns unfamiliar schemas and table relationships, then revise its assumptions when a migration changes those schemas. Asawa frames this as the stability-plasticity trade-off: retaining useful experience while updating obsolete beliefs. His conference presentation also argues that meaningful evaluation must consider capability, learning gains, and computational cost separately.
- Advisor Models: In research on adapting black-box models, Asawa and collaborators train a smaller model to provide situation-specific guidance to a more powerful system. Reinforcement learning improves the advisor using feedback from the larger model’s results, enabling specialization without access to the frontier model’s weights.
- SIEVE: His SIEVE research turns natural-language instructions, domain knowledge, and feedback into durable model updates. The method decomposes contextual guidance, generates relevant training questions, and distills selected information into model parameters, with experiments beginning from as few as three example queries.
Asawa’s longer-term interest is sample-efficient continual learning built into model architectures, training algorithms, and deployment itself. He also advocates stronger independent AI research and rejects the assumption that model safety and concentrated institutional power must be treated as opposing concerns, a position he has articulated publicly with Joey Gonzalez.
