All updates
Progress on papers, projects, and notes, from newest to oldest.
2026
PapersMultiLCR retrieves multi-defendant legal cases by reasoning about each defendant’s role and judgment outcome, beyond case-level textual similarity. The paper is published in the EMNLP 2026 main conference proceedings.
PapersSelSkill combines episode-level and step-level preference learning to teach agents when to invoke reusable skills and when to skip them. The paper is published in Findings of EMNLP 2026.
PapersOCSD distills supervision from future environment observations, using observation-calibrated self-distillation to refine token-level updates in agentic reinforcement learning. The preprint is available on arXiv.
PapersFutureBridge-OPD introduces teacher guidance at high-disagreement states and validates it through subsequent student trajectories, retaining bridges that lead to more teacher-preferred continuations. The preprint is available on arXiv, and the code is open source.
PapersQR-LJP brings quantitative reasoning about monetary values into legal judgment prediction, using legally grounded crime amounts to support judgment and prison-term predictions. The paper is published in ACM TKDD.