Skip to content

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

🕒 Published (v1): 2025-11-21 06:01 UTC · Source: Arxiv · Venue: AAAI 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Tree-of-Claims (ToC) reframes patent claim optimization as a structured search problem by coupling Monte Carlo Tree Search (MCTS) with a two-agent LLM system (EditorAgent + ExaminerAgent). The framework produces traceable, legally grounded claim revisions that consistently outperform zero-shot and few-shot LLM baselines across novelty, scope, and consistency metrics.

Problem

Existing LLM-based patent claim drafting relies on single-shot or few-shot generation, which lacks iterative, structured reasoning. These approaches cannot reliably balance novelty enhancement, legal scope preservation, and technical coherence simultaneously, and produce opaque revision histories that are unacceptable in high-stakes legal contexts.

Method

ToC models claim editing as a sequential decision-making problem over a ten-operation atomic action space (AddNovelFeature, ReplaceSynonym, ReframeViaFigure, DropElement, MergeElements, SplitElement, AddLimitation, ModifyRelationship, ChangeOrder, AddDependency). An MCTS tree is built where each node stores a claim state plus uncertainty estimates (epistemic σ^epi and aleatoric σ^ale). The four MCTS phases are:

  1. Selection via UCT score with σ-gating: paths where σ^epi > 0.2 are pruned or flagged for human review.
  2. Expansion by the ExaminerAgent (identifies disclosed claim elements) followed by the EditorAgent (proposes atomic edits); branching is controlled by progressive widening K(n) = α·N(n)^δ with α=2.0, δ=0.5.
  3. Simulation via a rule-guided rollout policy (entropy-based, confidence-based, or hybrid 0.6/0.4).
  4. Backpropagation of a five-component scalar reward: R = w1·R_coverage − w2·R_scope + w3·R_novelty + w4·R_consistency − w5·R_uncertainty, with weights (1.0, 0.5, 1.5, 0.8, 0.3).

The ExaminerAgent outputs strict JSON with disclosure status, evidence quotes, confidence, and epistemic uncertainty; the EditorAgent outputs structured edit operations with rationale, conditioned on the examiner's chain-of-thought feedback.

Key Contributions

  • Formulation of patent claim optimization as a structured MCTS search over a discrete atomic edit action space.
  • Uncertainty-aware MCTS with σ-gating (threshold 0.2) that separates epistemic from aleatoric uncertainty and prunes speculative edits.
  • Collaborative two-agent architecture (ExaminerAgent + EditorAgent) with strict JSON schemas ensuring auditable, step-by-step revision history.
  • Multi-objective reward function jointly optimizing novelty coverage, scope retention, consistency, and uncertainty penalty.
  • Benchmark construction from 1,145 USPTO Office Action patents (28,261 claims, 8,418 prior art references) with multimodal (text + figures) inputs.

Results

  • Primary metric (composite score, N=500 hold-out): ToC with GPT-4o achieves 0.701 vs. GPT-4o zero-shot 0.647 and few-shot 0.678 — an ~8% average improvement, up to 9% in certain cases.
  • Best open-source: ToC + Qwen2.5-VL-72B scores 0.658 vs. zero-shot 0.611 / few-shot 0.636, within 5% of GPT-4o+ToC.
  • Novelty (R_novelty): GPT-4o+ToC 0.732 vs. zero-shot 0.659; Claude-3.5 Sonnet+ToC 0.703 vs. zero-shot 0.633.
  • Perplexity (PPL): GPT-4o+ToC 8.72 vs. zero-shot 8.93; lower is better.
  • ROUGE-L / BLEU: GPT-4o+ToC 0.624 / 0.554 vs. zero-shot 0.590 / 0.522.
  • Human expert preference: ToC revisions preferred in ~2/3 of evaluations by five patent specialists (≥5 years experience).
  • Ablation: Removing uncertainty gating, progressive widening, or multi-agent collaboration each causes measurable degradation in novelty and coverage.
  • Sensitivity: Accuracy stable at 0.72–0.81 across hyperparameter grid; ~70% of eventual reward gain achieved within the first 6 MCTS iterations.

Limitations

  • Search budget is large: up to 800 iterations or 3,600 seconds per claim, making throughput a bottleneck at scale.
  • Dominant error category is "System Control Failures" (12.2% of errors): excessive branching under uncertainty, zero-confidence decisions, and missed pruning.
  • "Unsupported Novelty" errors (9.9%): the ExaminerAgent certifies novelty that is not genuinely supported by prior art evidence.
  • Dataset is domain-narrow (wireless communications patents, 2016 USPTO office actions) — generalization to other technology fields is untested.
  • Reward weights are empirically tuned on a development set; no principled derivation or cross-domain validation is provided.
  • Progressive widening and σ-gating hyperparameters require careful tuning; extreme settings (α=0.8, Tmax≤5) reduce accuracy by 3–5 points.

Relevance to Agentic AI / LLM Agents

ToC is a concrete instance of the search-augmented multi-agent paradigm: it replaces open-loop LLM generation with a closed-loop MCTS planner whose nodes are LLM agent calls, directly relevant to researchers studying how structured search (Tree-of-Thoughts, MCTS) can ground agentic reasoning in verifiable intermediate states. The ExaminerAgent/EditorAgent role decomposition — with one agent critiquing and one acting — is a clean realization of the critic-actor pattern increasingly used in agentic pipelines. The σ-gating mechanism for human-in-the-loop escalation addresses a key open problem in agentic AI: when should an agent defer to a human rather than proceed autonomously? The framework's focus on auditability (complete, justified edit history) also speaks directly to the transparency and controllability requirements for deploying agents in high-stakes domains.