Papers by Asish Joel Batha
PerDucer: Keyphrase-Driven Personalization Inducer for Summarization from User Histories (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior work reported that prepending long interaction histories to LLMs leads to unstable personalization, especially for multi-aspect documents. |
| Approach: | They propose a personalization inducer for frozen language models that maps latent preference signals to a small set of personalized keyphrases for the query document. |
| Outcome: | The proposed model outperforms the strongest history-prompting LLMs and SLMs in the PENS and OpenAI-Reddit benchmarks. |