Research map / Multi-agent systems and games / LLM agents

AgentODRL: A Large Language Model-based Multi-agent System for ODRL Generation

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AgentODRL, an Orchestrator-Workers multi-agent system that translates natural-language data-rights rules into ODRL policies, combining SHACL-validated correction loops and a LoRA-tuned semantic reflection model, evaluated on a new 770-case dataset.

Problem

Translating complex natural-language usage rules (parallel and recursive logical structures) into syntactically valid and semantically faithful ODRL policies is hard for single LLMs, and no NL-to-ODRL benchmark existed.

Approach

Decompose NL-to-ODRL into specialized worker agents (Rewriter, Splitter, Generator) coordinated by an Orchestrator that routes each use case by complexity, with a SHACL validator loop for grammar and a LoRA-finetuned expert validator providing semantic checkpoint reflection.

Key results

• Full pipeline (AOFP) improves average grammar score by 5.39% and semantic score by 14.52% over SCR-Enhanced across the three GPT-4.1 models on all use cases.
• On GPT-4.1-nano recursive cases, semantic score rises from 34.87 (OGS) to 61.53 with AOFP.
• Orchestrator-Workers on GPT-4.1-nano reaches average semantic score 80.22 vs 72.35 for single ODRL Generator, using 46.2M tokens vs 47.9M for the fixed Splitter path.

Surprising findings

• The fixed Splitter→Generator path outperformed the full Rewriter→Splitter→Generator path on average semantic score (84.02 vs 88.07 is actually higher for Rewriter; the Orchestrator underperformed the best fixed path), so the Orchestrator does not reach the best fixed-path ceiling.

Limitations

• Semantic score depends on LLM jurors and an identifier LLM rather than human judgment.
• Dataset is synthetic, augmented by Gemini 2.5 Pro from 70 seed cases, and limited to data-space scenarios.
• Orchestrator average semantic score falls below the best fixed workflow path.

Experimental setup

Datasets: 2,380 synthetic samples for LoRA training 70 seed use cases from literature, CESI standard drafts, Creative Commons 4.0, GDPR and CCPA templates

Benchmarks: 770-case ODRL NL-to-policy benchmark (own)

Metrics: grammar score (SHACL error-based) semantic score (checkpoint-based LLM jury) average number of reflections token consumption

Baselines: Ontology-Guided Strategy (OGS) SCR-Enhanced ODRL Generator (single-agent)

Builds on: SCR (self-correction rules) ontology-guided ODRL generation SHACL

Closest papers by meaning

• SecureClaw: Clawing Back Control of LLM Agents (NeurIPS 2026 · similarity 0.84)
• OpenCUA: Open Foundations for Computer-Use Agents (NeurIPS 2025 · similarity 0.84)
• DRIFT: Dynamic Rule-Based Defense with Injection Isolation for Securing LLM Agents (NeurIPS 2025 · similarity 0.83)
• AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.83)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.83)
• \$OneMillion-Bench: How Far are Language Agents from Human Experts? (NeurIPS 2026 · similarity 0.83)