mlsanityData to 7 Oct 2026

MASCRDM: Multi-Agent System for Compliance Risk Detection and Mitigation in Training Process of Large Language Models

arXiv 2609.39107 ↗First posted 30 Sep 2026cs.AIcited by 0

On the map: LLM agents and tool use → Agent threats, guardrails and governance also fits: General LLM agent frameworks and benchmarks

Abstract (arXiv)

Large Language Models (LLMs) have been applied in various fields. However, ensuring compliance and safety of LLMs, such as avoiding discrimination and bias, still remains a challenge. Current efforts mainly focus on detecting and filtering inputs and outputs of the trained models, rather than studying the intrinsic architecture of the models in real-time. To tackle this challenge, we analyze the LLMs training process and discover two critical issues: 1) Most of the existing methods are predominantly static in their approach to detection and filtering, achieving only localized optimizations without systematically enhancing the compliance of LLMs. 2) Another issue with existing approaches is the lack of real-time risk detection and mitigation across the full training process, which leads to limited flexibility. Motivated by these, we propose MASCRDM (Multi-Agent System for Compliance Risk Detection and Mitigation) during the LLM training process. Firstly, we develop a set of compliance rules based on existing Artificial Intelligence (AI) laws and a compliance-specific LLM with the instruction of compliance law experts. Then, we deconstruct LLMs into several components and identify key nodes based on the compliance knowledge graph. During LLMs training, we implement our multiple agents in the whole process, giving compliance risk alerts and suggestions for LLM developers. Experiments on discrimination and bias benchmark demonstrate that our multi-agent system can effectively improve the compliance while maintaining reasonable semantic performance. The results indicate that our method provides an executable path for mitigating compliance risk from within the LLMs systematically.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Agent threats, guardrails and governance, one of 16 directions in LLM agents and tool use. Builds on 43 of its 52 references that are in the map. First posted on arXiv 30 Sep 2026.

Builds on

43 of 52 references.

References · 43

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  3. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
  4. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
  5. 2023Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversationscited by 1,124
  6. 2021A General Language Assistant as a Laboratory for Alignmentcited by 1,023
  7. 2023Large Language Models Cannot Self-Correct Reasoning Yetcited by 965
  8. 2020StereoSet: Measuring stereotypical bias in pretrained language modelscited by 911
  9. 2020CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Modelscited by 703
  10. 2021BBQ: A Hand-Built Bias Benchmark for Question Answeringcited by 683
  11. 2023Bias and Fairness in Large Language Models: A Surveycited by 626
  12. 2024WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMscited by 460
31 more
  1. 2021BOLD: Dataset and Metrics for Measuring Biases in Open-Ended Language Generationcited by 410
  2. 2023AI Transparency in the Age of LLMs: A Human-Centered Research Roadmapcited by 119
  3. 2024From LLMs to LLM-based Agents for Software Engineering: A Survey of Current, Challenges and Futurecited by 85
  4. 2024Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignmentcited by 50
  5. 2025Improving LLM Safety Alignment with Dual-Objective Optimizationcited by 40
  6. 2025Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMscited by 37
  7. 2025Safety Pretraining: Toward the Next Generation of Safe AIcited by 35
  8. 2025Defending Against Prompt Injection with DataFiltercited by 30
  9. 2025Toward universal steering and monitoring of AI modelscited by 29
  10. 2024Tradeoffs Between Alignment and Helpfulness in Language Models with Steering Methodscited by 26
  11. 2025Defending Against Prompt Injection With a Few DefensiveTokenscited by 26
  12. 2025NRC VAD Lexicon v2: Norms for Valence, Arousal, and Dominance for over 55k English Termscited by 25
  13. 2025When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasonerscited by 16
  14. 2024New Job, New Gender? Measuring the Social Bias in Image Generation Modelscited by 15
  15. 2025Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Modelscited by 15
  16. 2024Superficial Safety Alignment Hypothesiscited by 14
  17. 2025Revisiting Replay and Gradient Alignment for Continual Pre-Training of Large Language Modelscited by 13
  18. 2025Towards Safe Reasoning in Large Reasoning Models via Corrective Interventioncited by 11
  19. 2025Unforgotten Safety: Preserving Safety Alignment of Large Language Models with Continual Learningcited by 10
  20. 2025Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editingcited by 9
  21. 2025Fairness Mediator: Neutralize Stereotype Associations to Mitigate Bias in Large Language Modelscited by 8
  22. 2025InfoPO: On Mutual Information Maximization for Large Language Model Alignmentcited by 7
  23. 2025Context Reasoner: Incentivizing Reasoning Capability for Contextualized Privacy and Safety Compliance via Reinforcement Learningcited by 7
  24. 2025Mitigating Biases in Language Models via Bias Unlearningcited by 6
  25. 2026Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projectioncited by 6
  26. 2025LLM Bias Detection and Mitigation through the Lens of Desired Distributionscited by 4
  27. 2026Continual Safety Alignment via Gradient-Based Sample Selectioncited by 4
  28. 2025KLAAD: Refining Attention Mechanisms to Reduce Societal Bias in Generative Language Modelscited by 3
  29. 2026LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safetycited by 3
  30. 2025Inference-Time Intervention in Large Language Models for Reliable Requirement Verificationcited by 1
  31. 2026Alignment Drift in Long-Term Human-LLM Interaction: A Mechanism-Oriented Frameworkcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2024Large Language Model Safety: A Holistic Survey
  2. 2025A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deployment
  3. 2024Large Language Model Evaluation Via Multi AI Agents: Preliminary results
  4. 2025Responsible Federated LLMs via Safety Filtering and Constitutional AI
  5. 2025SafetyFlow: An Agent-Flow System for Automated LLM Safety Benchmarking
  6. 2025Safety Compliance: Rethinking LLM Safety Reasoning through the Lens of Compliance
  7. 2024Navigating the Risks: A Survey of Security, Privacy, and Ethics Threats in LLM-Based Agents
  8. 2026PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models