Cornell University
LEGAL REASONING WITH SEARCH-AUGMENTED REINFORCEMENT LEARNING: A MULTI-TASK FRAMEWORK
Abstract
dc:description.abstractLarge language models hallucinate in 69% to 88% of legal queries, yet current systems only access external tools at inference time, never learning legal research methodology during training. This thesis presents a search-augmented reinforcement learning framework that integrates 18 specialized tools directly into the training loop via a smart Model Context Protocol client. The framework introduces four contributions: multi-tool training with a 9-call episodic constraint, multi-task reward routing using LLM judge ensembles across four legal task types, US jurisdiction compliance gating across all 50 states plus the federal system, and an attention-based interpretability framework for mapping legal reasoning patterns. Implemented with Qwen 2.5 3B and GRPO optimization on 8 A100 GPUs, the system achieves a 10.5 percentage point improvement over baselines on LegalBench (59.6% vs. 49.7%), 97% jurisdictional compliance, and 0.72 Spearman correlation between model attention and expert fact importance rankings.
Degree
thesis:*- Name thesis:degree_name
- M.S., Computer Science
- Level thesis:degree_level
- Master of Science
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Cornell University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Akinboro, David
- Chair dc:contributor.chair
-
- Cardie, Claire
- Committee member dc:contributor.committeemember
-
- Manzoor, Emaad
Subjects
dc:subject × 6Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Other
-
Akinboro_cornell_0058O_12647
https://www.proquest.com/LegacyDocView/DISSNUM/32444577