ToC: Tree-of-Claims Search with Multi-Agent Language Models

AI summary

Overview

Research area: Automated patent claim drafting and revision, combining large language model (LLM) reasoning with Monte Carlo Tree Search (MCTS) and multi-agent collaboration in a legal-domain text-editing setting.

Technical level: Advanced. The paper assumes familiarity with MCTS, tree search mechanics (UCT, backpropagation, progressive widening), multi-agent LLM orchestration, reward shaping, and patent prosecution concepts such as novelty, prior art disclosure, and claim scope.

Scope: The paper proposes the Tree of Claims (ToC) framework, which recasts patent claim editing as a guided search problem driven by an LLM EditorAgent, an LLM ExaminerAgent, and a five-term multi-objective reward function, evaluated on a USPTO-derived benchmark of patent claims.

What This Paper Is About

Patent claims must simultaneously maximize novelty over prior art and preserve the legal scope of the original invention, and human drafting of these revisions is labor-intensive, costly, and inconsistent. Standard LLMs handle this poorly because single-shot or few-shot prompting lacks the structured, iterative reasoning needed to make and verify precise claim edits. ToC reframes claim revision as a sequential decision-making search over atomic edit operations, where specialized LLM agents propose and critique each edit under a multi-objective reward.

Key Contributions

1. A new formulation of claim optimization as structured search. The paper models claim editing as a sequential decision problem over a state space of claim configurations and an action space of ten atomic editing operations, solved with Monte Carlo Tree Search rather than single-pass generation.

2. A collaborative multi-agent architecture. An EditorAgent generates legally grounded edit operations and an ExaminerAgent simulates patent examiner scrutiny, producing structured chain-of-thought assessments of novelty and prior art disclosure, including a Status label (Disclosed, NotDisclosed, or PartiallyDisclosed), verbatim evidence quotes, confidence, and an uncertainty measure.

3. A multi-objective reward function with uncertainty gating. The reward linearly combines coverage, scope retention, novelty, consistency, and uncertainty terms with weights (w1,…,w5) = (1.0, 0.5, 1.5, 0.8, 0.3), and a σ-epistemic gating mechanism (threshold 0.2) prunes or flags high-uncertainty paths.

4. Extensive empirical evaluation. Experiments on a benchmark built from the USPTO Office Actions Research Dataset across five multimodal LLMs (three closed-source, two open-source) in zero-shot, few-shot, and ToC configurations, plus ablation, sensitivity, error, and expert-preference analyses.

Main Findings

• ToC improves over standard prompting. Across all five evaluated MLLMs, adding ToC raised performance relative to zero-shot and few-shot baselines, with an average composite score improvement of 8% and up to 9% in certain cases.

• GPT-4o with ToC scored highest overall. It posted coverage 0.582±0.050, novelty 0.732±0.050, consistency 0.956±0.017, the lowest uncertainty term 0.068±0.018, and the highest aggregate score of 0.701±0.03.

• Other closed-source results with ToC. OpenAI O1 reached coverage 0.560±0.048, novelty 0.712±0.050, consistency 0.942±0.012, and overall 0.680±0.03. Claude-3.5 Sonnet reached coverage 0.548±0.049, novelty 0.703±0.051, consistency 0.945±0.018, and overall 0.675±0.03.

• Open-source results with ToC. Qwen2.5-VL-32B reached coverage 0.507±0.047, novelty 0.665±0.050, and overall 0.639±0.03. Qwen2.5-VL-72B reached coverage 0.534±0.048, novelty 0.682±0.051, and overall 0.658±0.03, the best open-source result and within 5% of the GPT-4o benchmark according to the paper.

• Auxiliary generation metrics with ToC (GPT-4o). JSON parsing completeness 0.996±0.001, perplexity 8.72±1.00, ROUGE-L 0.624±0.049, and BLEU 0.554±0.052, the best values in Table 4. Qwen2.5-VL-72B reported JSON 0.993±0.004, perplexity 9.52±1.17, ROUGE-L 0.596±0.050, BLEU 0.525±0.051.

• Expert preference favors ToC. Five patent specialists with at least five years of industry experience evaluated revisions; ToC was preferred in approximately two-thirds of evaluations, particularly where prior art was available.

• Ablations show every module matters. Removing uncertainty gating, progressive widening, or multi-agent collaboration caused noticeable performance degradation, with the absence of uncertainty control and agent interaction significantly affecting novelty, coverage, and overall legal quality. The paper reports the direction of these effects rather than specific ablation numbers in the text.

• Hyperparameter sensitivity is modest. Accuracy stayed in the 0.72–0.81 range across the swept grid of progressive-widening coefficient α and maximum search depth T_max, peaking at α = 0.6 and T_max = 15. Very small trees (T_max ≤ 5) or aggressive widening (α = 0.8) reduced scores by 3–5 points.

• Reward convergence is fast. Roughly 70% of the eventual reward gain was achieved within the first six search iterations, with the curve plateauing after iteration 10.

• Error analysis identifies recurring failure modes. Across the categorized failures, System Control Failures were the largest category at 139 cases (12.2%), followed by Invalid Modifications at 120 (10.5%), Unsupported Novelty at 113 (9.9%), Input Misalignment at 78 (6.8%), and Legal or Style Issues at 31 (2.7%).

Methodology in Plain English

The system starts from an initial claim and a set of prior art documents and tries to produce a revised claim that is more novel while keeping the original technical scope intact. It treats every small edit as a move in a game tree.

Two LLM agents do the work. The ExaminerAgent looks at each element of the claim against the prior art and decides whether it is disclosed, partially disclosed, or not disclosed, quoting evidence and giving a confidence score and an uncertainty score. The EditorAgent takes that feedback and proposes concrete edits drawn from ten operation types, such as adding a novel feature, adding a limitation, replacing a synonym, merging or splitting elements, or changing element order. Edits follow precedence rules, for example a novel feature must be added before synonyms are swapped.

Monte Carlo Tree Search explores sequences of these edits. It selects promising paths by an upper confidence bound on tree score, expands them by having the EditorAgent propose candidates (with progressive widening limiting how many children a node can have), simulates the result, and backpropagates the reward. Uncertainty plays a gatekeeping role: the ExaminerAgent's variance-based uncertainty is split into epistemic and aleatoric parts, and only the epistemic part is penalized in the reward, so noise in prior art phrasing does not punish good edits. Paths whose epistemic uncertainty exceeds 0.2 are pruned or flagged for human review. Search stops at 800 iterations, when reward improvement stalls below 0.01, after 3,600 seconds, or after 20 consecutive failures.

Each candidate claim is scored by a weighted sum of five terms: coverage (turning disclosed points into not-disclosed), scope retention (penalizing unnecessary narrowing), novelty (counting only examiner-rated innovative changes), consistency (legal readability times technical coherence), and uncertainty.

The benchmark comes from the USPTO Office Actions Research Dataset, focused on wireless communications patents and 2016 office actions. It contains 1,145 unique patents (106 allowed, 1,039 rejected), 28,261 claims, and 8,418 prior art references, with 82.3% of claims having prior art references. The dataset is imbalanced at 9.3% allowed versus 90.7% rejected, averages 24.7 claims per patent, includes 27 overlapping applications (2.4%), and incorporates figures, making it multimodal. Evaluation used a 500-sample hold-out set, three seeds, and NVIDIA A100 GPUs (80GB).

Why This Matters

The work shows that pairing an LLM's generative ability with an explicit search process and an adversarial critic agent can produce more controllable, auditable outputs than prompting alone. Because every edit is an explicit, recorded operation with reasoning attached, the revision history is traceable, which is a prerequisite for adoption in high-stakes legal work. It also demonstrates a general pattern—specialized agents plus tree search over atomic edits—that transfers across model sizes, from open-source 32B and 72B Qwen2.5-VL models to closed-source frontier systems.

Real-world applications:

• Patent prosecution support: helping attorneys and agents respond to examiner rejections by generating candidate amendments with documented justification.
• Prior art analysis: automated element-by-element disclosure checks against cited references, with quotes and confidence scores.
• IP portfolio strategy: rapid exploration of alternative claim scope positions during drafting or continuation decisions.
• Human-in-the-loop review systems: using the uncertainty gate and human-review flags to route only genuinely ambiguous cases to a specialist.

Industry relevance centers on cost and consistency. Manual claim drafting is described as resource-intensive, costly, and inherently inconsistent, while existing LLM tools are characterized as black boxes with limited control over scope and legal consistency. ToC targets that gap directly by making each modification explicit and justified, which the paper argues is essential for real-world integration.

Future Directions

• Incorporating multimodal reasoning for figure-grounded edits. Although the dataset includes patent figures and a ReframeViaFigure operation exists, the paper lists stronger figure-grounded editing as future work.
• Improving search efficiency through distributed computing. Search is bounded by 800 iterations and 3,600 seconds, and the paper identifies efficiency gains as an open direction.
• Generalizing to other structured editing domains, named in the paper as legal contracts, medical protocols, and scientific methods.
• Improving uncertainty calibration and validation. The error analysis flags excessive branching due to uncertainty, zero-confidence decisions, missed pruning opportunities, and unsupported novelty claims as areas needing better calibration, validation mechanisms, and alignment with prior-art evidence.

Target Audience

This paper is most useful to researchers working on LLM agents, tree search for text generation, and reward design for constrained generation; to patent practitioners and legal-tech engineers interested in auditable AI-assisted claim drafting; and to applied machine learning engineers who want a concrete example of combining MCTS with multi-agent LLM critique. Readers seeking a beginner-level introduction to patent law or to MCTS will need supplementary background, since the paper assumes fluency in both.

Authors’ abstract

Optimizing patent claims is a critical yet challenging task, demanding careful balance between maximizing novelty and preserving legal scope. Manual claim drafting is labor-intensive, costly, and inherently inconsistent, while conventional Large Language Models (LLMs) often lack the structured, iterative reasoning essential for precise claim refinement. To address these challenges, we introduce Tree of Claims (ToC), an innovative framework that redefines claim editing as a guided search problem. ToC synergistically integrates Monte Carlo Tree Search (MCTS) with a collaborative multi-agent system, comprising an LLM-based EditorAgent that proposes contextually grounded edits, and an ExaminerAgent that mimics patent examiner critiques through structured, chain-of-thought analyses of novelty and prior art disclosure. Driven by a carefully designed multi-objective reward function, ToC jointly optimizes novelty, scope retention, and semantic coherence. Experimental evaluation on a benchmark of 1145 claims demonstrates that ToC significantly outperforms standard LLMs in zero-shot and few-shot scenarios, achieving an average composite score improvement of 8\%, and up to 9\% in certain cases. Extensive experiments, including detailed ablation studies, validate ToC's efficacy in generating superior, legally robust claim revisions. Overall, ToC establishes a transparent, controllable, and interpretable methodology that effectively bridges advanced LLM reasoning capabilities with strategic MCTS planning for structured patent claim optimization.The source code is available at https://github.com/ysy2003/ToC.