Introduction
Drafting patent claims is a high-stakes balancing act in intellectual property law. A claim defines the legal "fence" around an invention: if it is too broad, it risks rejection for overlapping with "prior art" (existing inventions); if it is too narrow, the resulting patent may be commercially worthless. While Large Language Models (LLMs) have shown promise in text generation, they often struggle with the iterative, strategic reasoning required for patent law. They frequently produce "one-shot" drafts that lack technical precision or inadvertently expand legal scope in ways that would not survive litigation.
To address these challenges, researchers have developed Tree-of-Claims (ToC), a framework that treats patent claim optimization as a structured search problem rather than a simple generation task. By combining Monte Carlo Tree Search (MCTS) with a collaborative multi-agent system, ToC provides a controllable, transparent, and strategically capable solution for legal drafting.
The Tree-of-Claims Framework
The core innovation of the ToC framework is the conceptualization of claim editing as a guided search through a tree of possible revisions. Instead of generating a final claim in a single pass, the system explores a tree where each node represents a specific atomic edit, such as "Add Novel Feature" or "Replace Synonym."
Multi-Agent Collaboration
The search process is driven by two specialized agents:
- ExaminerAgent: Acts as the "critic," mimicking the role of a patent office examiner. It analyzes claim elements against prior art and provides structured feedback on whether a claim is disclosed or novel.
- EditorAgent: Acts as the "creator." It utilizes the Examiner’s feedback to propose legally grounded, atomic edits to improve the claim's standing.
Strategic Search and Gating
ToC employs the standard MCTS phases—Selection, Expansion, Simulation, and Backpropagation—to identify the optimal path of revisions. To prevent the model from "hallucinating" legal certainties, the system incorporates Uncertainty-Aware Gating ($\\sigma$-gating). By estimating epistemic uncertainty, the system can prune branches where the model is unsure of a revision path or flag them for human review. Additionally, "progressive widening" is used to manage the search space, limiting the number of child nodes to ensure computational efficiency.
Experimental Results
The researchers evaluated ToC using a benchmark constructed from the USPTO Office Actions Research Dataset, focusing on over 1,100 unique wireless communications patents. The search was guided by a multi-objective reward function balancing novelty coverage, scope retention, technical novelty, logical consistency, and risk control.
Performance Gains
ToC consistently improved the performance of state-of-the-art LLMs. The framework boosted the average composite score of models by 8%, with gains of up to 9% in novelty-focused scenarios. Notably, GPT-4o integrated with ToC achieved the highest overall score (0.701), outperforming even the base OpenAI O1 model.
Expert Validation
In blind evaluations conducted by human patent specialists with over five years of experience, ToC-generated revisions were preferred in approximately two-thirds of cases compared to original drafts or standard LLM outputs. This suggests that the structured search approach aligns more closely with professional legal standards than traditional prompting methods.
Conclusion and Future Work
The ToC framework demonstrates that for complex, constraint-heavy tasks like patent drafting, structured search is superior to simple prompting. By separating the roles of "creator" and "critic" and employing uncertainty gating, ToC produces more defensible and robust legal text.
Future work will focus on incorporating multimodal reasoning to allow the system to understand supporting figures and diagrams. Additionally, the researchers plan to explore the application of ToC to other high-stakes domains, such as medical protocols and complex legal contracts, where strategic planning and precision are paramount.
Summary updated