Beliz Gunel (Google DeepMind)

May 16, 2025
Cory 400

Title and Abstract

Long-Range Tasks Using Short-Context LLMs: Incremental Reasoning With Structured Memories

Long-range tasks require reasoning over long inputs. Existing solutions either need large compute budgets, training data, access to model weights, or use complex, task-specific approaches. We present PRISM, which alleviates these concerns by processing information as a stream of chunks, maintaining a structured in-context memory specified by a typed hierarchy schema. This approach demonstrates superior performance to baselines on diverse tasks while using at least 4x smaller contexts than long-context models. Moreover, PRISM is token-efficient. By producing short outputs and efficiently leveraging key-value (KV) caches, it achieves up to 54% cost reduction when compared to alternative short-context approaches. The method also scales down to tiny information chunks (e.g., 500 tokens) without increasing the number of tokens encoded or sacrificing quality. Furthermore, we show that it is possible to generate schemas to generalize our approach to new tasks with minimal effort.

Bio

Beliz Gunel is a Senior Research Scientist at Google DeepMind in the Gemini Post-Training team, where she currently focuses on tool understanding and coding capabilities within large language models. She earned her PhD from Stanford University in 2022, focusing on using prior knowledge and structure for data-efficient machine learning. During her PhD, Beliz interned at Microsoft Research, Meta AI, Google AI, and Google Brain.

Beliz regularly reviews for ML and NLP conferences such as NeurIPS, ICML, ICLR, ACL, EMNLP, and NAACL. She co-organized the Representation Learning on Graphs and Manifolds Workshop (ICLR 2019), Women in Machine Learning Workshop (NeurIPS 2022), and Knowledge and Logical Reasoning in the Era of Data-driven Learning Workshop (ICML 2023). Previously, she held a Visiting Scholar appointment at Stanford Electrical Engineering, where she co-advised PhD students. Beliz is also an active angel investor in the AI space and has advised approximately 10 students (BS, MS, PhD).