Gavel Paper on LLM Legal Summarization Accepted to EMNLP 2026

Yao Dou posted that the Gavel paper was accepted to EMNLP 2026 Main. The work examines frontier large language models on legal case summarization tasks where inputs exceed 500K tokens. According to the announcement, models encounter difficulties once inputs surpass 256K tokens. They prove much more likely to omit key information than to produce hallucinations. The authors also introduced Gavel-Agent, an agentic evaluation method that operates without references. This report originates from a researcher on the project and covers the acceptance along with the main findings described in the post.