Uploaded October 2025 | Updated September 2026, 21 hours ago
Language models are often used for complex reasoning tasks that require information synthesis over long inputs. In this talk, Adithya will present his work on augmenting language models with text compression tools to improve their long-context performance. First, he compares compression-based methods against a full-context setup for multi-document summarization. Based on this analysis, he will discuss our hybrid retrieval-
augmented method that estimates the optimal number of tokens to retrieve for a given configuration of dataset, retriever and summarizer. Finally, he will highlight future research directions on building efficient
systems and evals for complex information synthesis tasks.
Adithya Pratapa is a Ph.D. candidate at the Language Technologies Institute at Carnegie Mellon University, advised by Teruko Mitamura. His research focuses on developing methods to improve performance and efficiency of language models for large-scale multi-document tasks. His work received an outstanding paper award at EMNLP and a best paper award at a NAACL workshop.
Language models are often used for complex reasoning tasks that require information synthesis over long inputs. In this talk, Adithya will present his work on augmenting language models with text compression tools to improve their long-context performance. First, he compares compression-based methods against a full-context setup for multi-document summarization. Based on this analysis, he will discuss our hybrid retrieval-
augmented method that estimates the optimal number of tokens to retrieve for a given configuration of dataset, retriever and summarizer. Finally, he will highlight future research directions on building efficient
systems and evals for complex information synthesis tasks.
Adithya Pratapa is a Ph.D. candidate at the Language Technologies Institute at Carnegie Mellon University, advised by Teruko Mitamura. His research focuses on developing methods to improve performance and efficiency of language models for large-scale multi-document tasks. His work received an outstanding paper award at EMNLP and a best paper award at a NAACL workshop.
