💬 NLP

RCBSF: A Multi-Agent Framework for Automated Contract Revision via Stackelberg Game

The paper proposes the Risk-Constrained Bilevel Stackelberg Framework (RCBSF), a multi-agent system that models automated contract revision as a non-cooperative Stackelberg game to enforce safety constraints and achieve state-of-the-art performance with a 84.21% Risk Resolution Rate.

Original authors: Shijia Xu, Yu Wang, Xiaolong Jia, Zhou Wu, Kai Liu, April Xiaowen Dong

Published 2026-04-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shijia Xu, Yu Wang, Xiaolong Jia, Zhou Wu, Kai Liu, April Xiaowen Dong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to fix a very complicated, high-stakes legal contract. You ask a smart AI (a Large Language Model) to do it.

The Problem:
If you just ask a standard AI to "fix this," it's like hiring a very confident but slightly reckless intern. The intern might:

Hallucinate: Make up rules that don't exist (e.g., "The company can seize all your personal assets if you sneeze").
Miss the Point: Change the wrong things or leave dangerous loopholes open.
Be Wasteful: Write a 50-page revision when a 5-page one would do.

The Solution: RCBSF (The "Strict Boss and the Team" Framework)
The authors of this paper created a new system called RCBSF. Instead of asking one AI to do everything, they set up a game with three specific roles, inspired by a concept called a "Stackelberg Game" (think of it as a chess match where one player moves first and forces the other to react).

Here is how it works, using a simple analogy:

1. The Characters

The Global Prescriptive Agent (GPA) = The Strict Boss.
- This AI doesn't write the contract. Its only job is to audit and set the rules.
- It looks at the contract and says, "Okay, here are the 5 specific things that are dangerous: The time limit is too long, the location is too vague, and this clause is illegal. Here is exactly how to fix them."
- It acts like a strict project manager who hands you a checklist and says, "Do not leave this office until every item on this list is checked off."
The Constrained Revision Agent (CRA) = The Hard-Working Editor.
- This is the AI that actually rewrites the text.
- But unlike a normal AI, it is chained to the Boss's checklist. It cannot just "guess." It must follow the Boss's specific instructions to fix the exact problems identified.
The Local Verification Agent (LVA) = The Quality Control Inspector.
- After the Editor makes a change, the Inspector checks it immediately.
- "Did you actually fix the time limit? Or did you just move the words around?"
- If the fix isn't perfect, the Inspector sends it back to the Editor to try again.

2. The Process: A "Game" of Iteration

Instead of one-and-done, this system plays a multi-round game:

Round 1: The Boss (GPA) finds the risks and gives a strict list of instructions. The Editor (CRA) tries to fix them.
Round 2: The Inspector (LVA) checks the work. "You missed the part about the 50-mile radius!"
Round 3: The Editor goes back, fixes the specific error, and the Boss checks again.
The End: They keep looping until the contract is perfect, safe, and follows all the rules.

3. Why is this better? (The "Magic" Ingredients)

No More "Fake" Fixes:
- Old Way: The AI might say, "I fixed the risk!" but actually leave the dangerous text unchanged.
- RCBSF Way: The Boss demands evidence. "Show me the specific sentence where you added the 48-hour notice." If the AI can't point to it, the system rejects the change.
The "Budget" Constraint:
- Imagine the Editor has a limited amount of "ink" (tokens) to use. The Boss tells them, "You only have enough ink to fix the most dangerous risks. Don't waste it on fixing typos." This stops the AI from writing long, useless paragraphs and forces it to focus on what actually matters.
The "Stackelberg" Advantage:
- In a normal game, everyone tries to win at the same time (which leads to chaos).
- In this game, the Boss moves first. By setting the rules before the Editor starts, the system guarantees a better outcome. It's like a coach drawing up a play before the players run onto the field. The players (the Editor) are free to run, but they must run the play the coach designed.

The Result

The paper tested this system on thousands of real legal contracts.

Risk Resolution: It fixed 84% of the dangerous legal risks (compared to about 70-79% for other methods).
Efficiency: It did this using fewer "words" (tokens), meaning it was faster and cheaper to run.
Quality: The final contracts were clearer, more professional, and less likely to get someone sued.

In Summary:
RCBSF turns AI contract revision from a "magic black box" that guesses answers into a structured, disciplined team. It uses a strict Boss to define the problem, a focused Editor to solve it, and a tough Inspector to ensure the job is done right, all while playing a strategic game to guarantee the best possible result.

1. Problem Statement

The paper addresses the critical limitations of using Large Language Models (LLMs) for automated legal contract revision. While LLMs show promise in legal AI, their direct deployment in high-stakes scenarios faces two primary hurdles:

Hallucinations and Safety: Models often generate factually incorrect clauses, false citations, or "hallucinated" fixes that introduce new legal risks.
Lack of Rigorous Constraints: Standard prompting or simple iterative refinement (e.g., Chain-of-Thought) lacks a formal mechanism to enforce strict risk budgets or hierarchical auditing. This leads to outputs that may be linguistically fluent but legally unenforceable or biased.

Existing Multi-Agent Systems (MAS) often rely on cooperative information aggregation, failing to capture the adversarial nature of real-world legal practice, where a senior partner (auditor) must impose binding constraints on an associate's (revisor) draft.

2. Methodology: The RCBSF Framework

The authors propose the Risk-Constrained Bilevel Stackelberg Framework (RCBSF), which models contract revision as a non-cooperative Stackelberg game (a hierarchical game where one player moves first, and the other responds).

Core Architecture

The system consists of three specialized agents organized into two strategic levels:

The Leader (Global Prescriptive Agent - GPA):
- Acts as the "Outer Auditor."
- Analyzes the contract and projects it onto a 5-Dimensional Risk Hyperplane: Category, Location, Evidence, Issue, and Suggestion.
- Optimizes a Risk-Budget Objective, generating a specific instruction vector ( $h$ ) that defines the constraints for the revision.
- Uses a Q-Score mechanism to weight risks based on severity (Validity, Liability Scope, Control Allocation, Remediability).
The Follower System (Revision System):
- Constrained Revision Agent (CRA): The "Drafter" that generates the revised contract text. It optimizes its output to maximize likelihood under the LLM distribution while strictly adhering to the Leader's constraints.
- Local Verification Agent (LVA): The "Inner Auditor" that performs iterative checks on the CRA's output, calculating residual risk and providing feedback gradients.

Mathematical Formulation

Bilevel Optimization: The problem is formulated as a Bilevel Optimization Problem (BOP).
- Upper-Level (Leader): Maximizes risk mitigation ( $J_L$ ) subject to a computational budget (token limits).
- Lower-Level (Follower): Minimizes residual risk and maximizes semantic quality ( $J_F$ ) given the Leader's constraints.
Iterative Gradient Approximation: The system uses a discrete approximation of gradient ascent. The LVA computes a fusion gradient to guide the CRA's text generation, updating the contract state $x^{(t+1)}$ until convergence.
Equilibrium: The framework enforces a Stackelberg Equilibrium, theoretically proven to yield strictly superior utility compared to unguided Nash Equilibria (standard prompting).

3. Key Contributions

Novel Framework (RCBSF): Introduction of a game-theoretic, hierarchical multi-agent framework that decouples strategic auditing (GPA) from execution (CRA/LVA), mimicking real legal team dynamics.
Theoretical Guarantees: The authors provide mathematical proofs demonstrating that the Stackelberg formulation converges to a stationary point and offers strictly superior utility over standard unguided generation, particularly in non-convex risk manifolds.
Unified Benchmark: Construction of a comprehensive legal revision benchmark derived from four authoritative datasets (MAUD, CUAD, ContractNLI, PrivacyQA), covering 711 contracts across 41 categories.
5-Dimensional Risk Taxonomy: A structured approach to risk extraction that grounds suggestions in specific evidence and location, significantly reducing hallucinations.

4. Experimental Results

The framework was evaluated against four baselines (Standard Zero-Shot, CoT-Refinement, RAG-Based, and Iterative Refinement) using various LLM backbones (Qwen2.5, Mistral, LawLLM, LexiLaw).

Risk Resolution Rate (RRR): RCBSF achieved a state-of-the-art average RRR of 84.21%, significantly outperforming the best baseline (Iterative Refinement at 79.56%) and the Standard method (70.81%).
Contract Quality (CQ): RCBSF dominated in clarity, rigor, balance, and professionalism (scoring ~86.87 on a 0-100 scale with Qwen2.5-7B-Chat), avoiding the readability degradation often seen in rigid rule-based methods.
Token Efficiency (TES): The framework achieved the highest Token Efficiency Score (87.29%), proving that the Leader's discrete prompt updates guide the Follower more efficiently than continuous CoT reasoning, minimizing token waste.
Ablation Studies:
- Removing the 5-Dimensional Constraints caused a massive drop in RRR (-11.06%), confirming the necessity of specific evidence and suggestions.
- Removing Budget Constraints improved RRR slightly but drastically reduced token efficiency (-12.14%), highlighting the trade-off management.
- Iterative refinement (K=3) was found to be the Pareto-optimal stopping point, balancing performance gains with computational cost.

5. Significance and Impact

Paradigm Shift: RCBSF moves legal AI from "generative" models to "constrained optimization" models, treating contract revision as a strategic game rather than a simple text completion task.
Reliability: By enforcing a hierarchical audit loop and grounding revisions in specific evidence, the framework significantly mitigates the risk of hallucinations, a critical requirement for legal adoption.
Scalability: The framework is model-agnostic (tested on various LLM sizes) and demonstrates that smaller models can achieve expert-level performance when guided by the Stackelberg structure.
Practical Application: The system provides a blueprint for building reliable, constraint-aware AI assistants that can be deployed in high-stakes commercial environments to reduce legal risk and improve operational efficiency.

In conclusion, RCBSF establishes a new standard for automated legal text generation by rigorously combining game theory, multi-agent collaboration, and risk-constrained optimization to produce legally compliant, high-quality contract revisions.