CogniAlign: survivability-grounded multi-agent moral reasoning for safe and transparent AI

Hasin Jawad Ali, Ilhamul Azam, Ajwad Abrar, Md. Kamrul Hasan, Hasan Mahmud

TL;DR

This paper introduces CogniAlign, a novel multi-agent AI framework that grounds moral reasoning in "survivability" – the capacity for individual and collective persistence. By simulating structured debates among discipline-specific AI agents, CogniAlign aims to provide more transparent, consistent, and empirically anchored moral judgments than monolithic LLMs, demonstrating significant improvements in key ethical evaluation metrics.

What Problem Does This Paper Address?

The core challenge this paper tackles is the "AI alignment problem," specifically how to ensure AI systems make decisions consistent with human values. This is difficult because human moral principles are often abstract, conflicting, and lack clear application in real-world scenarios. Existing approaches often rely on opaque "black-box" models or abstract guidelines that are hard to operationalize. The authors highlight the lack of consensus on abstract values, the difficulty in explaining AI decisions, and the potential for AI to misgeneralize. They propose a new approach to bridge this gap by grounding AI moral reasoning in a more concrete, empirically verifiable concept: survivability.

Methods

CogniAlign employs a multi-agent deliberation framework. The system comprises several "scientist agents," each representing a distinct discipline (neuroscience, psychology, sociology, evolutionary biology). These agents engage in a structured debate process:

  1. Argument Generation: Each agent independently constructs an argument about a moral question, grounded in its disciplinary principles and focusing on how actions impact individual and collective survivability.
  2. Rebuttal Generation: Agents then receive and critique each other's arguments, offering counterpoints or elaborations from their disciplinary lens.
  3. Arbiter Synthesis: A final arbiter agent synthesizes the entire deliberation transcript, identifying points of agreement and disagreement, and providing a final judgment based solely on survivability considerations.

This process is designed to be transparent and auditable, moving beyond the opacity of single-model outputs. The framework is built on naturalistic moral realism, positing that moral facts are discoverable through empirical means.

Datasets

The paper does not mention specific datasets used for training the underlying LLMs. However, for evaluation, they used:

  • Classic and Novel Moral Questions: A diverse set of moral dilemmas.
  • "Whispering Cure" Dilemma: A specific case from Neuman et al. (2025).
  • Heinz Dilemma: A classic dilemma from Kohlberg's research.
  • 60 Additional Cases: Sourced from Kabir et al. (2025) (DFAR dataset) and the "Battery III Dilemmas" (Neuman et al. 2025), excluding the Whispering Cure, and Thomson's (1985) "Trolley Dilemma."

Key Insights and Ideas

  • Survivability as a Moral Anchor: The central idea is that "survivability" (individual and collective) provides a concrete, empirically grounded, and cross-disciplinary anchor for moral reasoning, moving beyond abstract or subjective values.
  • Interdisciplinary Deliberation for Robustness: Simulating debates among agents from different scientific fields helps to identify blind spots, challenge assumptions, and produce more comprehensive and robust moral judgments than a single agent or model.
  • Transparency and Auditability: The structured deliberation process makes the reasoning behind AI moral judgments visible and auditable, addressing the "black-box" problem of current AI.
  • Naturalistic Moral Realism: The framework is grounded in the philosophical idea that moral truths are natural facts that can be studied empirically, allowing for objective evaluation.
  • Beyond Utilitarianism: CogniAlign aims to avoid the pitfalls of simple utilitarian calculus by considering short-term and long-term survivability across multiple dimensions and stakeholders, not just maximizing a single utility function.

Key Results

CogniAlign consistently outperformed GPT-4o across various evaluation metrics:

  • Overall Average Performance Gain: 12.2 points in analytic quality, 31.2 points in decisiveness, and 15 points in depth of explanation.
  • Whispering Cure Dilemma: CogniAlign achieved an overall score of 84, compared to GPT-4o's 74. CogniAlign recommended conventional medicine (Option B), emphasizing long-term survivability and trust, while GPT-4o recommended deceptive compromise (Option D), prioritizing short-term harm reduction.
  • Heinz Dilemma: CogniAlign scored 79, while GPT-4o scored 65.8. CogniAlign's judgment emphasized addressing systemic inequalities, whereas GPT-4o provided an indecisive response based on multiple ethical frameworks.
  • Secondary Cases: CogniAlign maintained an average score of 84, compared to GPT-4o's 72.5, across 60 additional cases.

Limitations and Future Work

The authors acknowledge several limitations:

  • Subjectivity in Evaluation: While using expert assessments, the metrics still involve subjective judgment.
  • Formalizing Trade-offs: The precise mechanisms for weighing individual vs. collective survivability trade-offs are not formally outlined.
  • Engineering Complexity: The prototype could benefit from more sophisticated engineering, including a larger agent roster, dynamic agent creation, and tool use for fact-checking.
  • Computational Intensity: The deliberation process can be computationally intensive, potentially making it unsuitable for simple queries.
  • Model Agnosticism Principle (MAP) Validation: The principle that moral truths can be discovered independent of model biases needs further empirical validation across different LLMs.
  • LLM Biases: The influence of specific LLM biases on survivability-based deliberation needs comprehensive identification.
  • Mechanistic Interpretability: While the process is transparent, the underlying LLM outputs might still be opaque or prone to hallucination.
  • Arbiter Vulnerabilities: The LLM arbiter could be susceptible to adversarial inputs or evaluation biases.
  • Deceptive Alignment: The system might still be susceptible to deceptive alignment, where agents learn to appear aligned with survivability without genuinely adopting the objective.

Future work includes distilling the process into a more efficient technique, integrating it into Constitutional AI models, developing game-theoretic models for agent interaction, and exploring human-in-the-loop arbitration.

Summary updated

Abstract

The challenge of aligning artificial intelligence (AI) with human values persists due to the abstract and often conflicting nature of moral principles and the opacity of existing approaches. This paper introduces CogniAlign, a multi-agent deliberation framework based on naturalistic moral realism, that grounds moral reasoning in survivability, defined across individual and collective dimensions, and operationalizes it through structured deliberations among discipline-specific “scientist agents.” Each agent, representing neuroscience, psychology, sociology, and evolutionary biology, provides arguments and rebuttals that are synthesized by an arbiter into transparent and empirically anchored judgments. As a proof-of-concept study, we evaluate CogniAlign on classic and novel moral questions and compare its outputs against GPT-4o using a five-part ethical audit framework with the help of three experts. Results show that CogniAlign consistently outperforms the baseline across more than sixty moral questions, with average performance gains of 12.2 points in analytic quality, 31.2 points in decisiveness, and 15 points in depth of explanation. In the Heinz dilemma, for example, CogniAlign achieved an overall score of 79 compared to GPT-4o’s 65.8, demonstrating a decisive advantage in handling moral reasoning. Through transparent and structured reasoning, CogniAlign demonstrates the feasibility of an auditable approach to AI alignment, though certain challenges still remain.
BibTeX
@inproceedings{ali2026cognialign,
  author = {Hasin Jawad Ali and Ilhamul Azam and Ajwad Abrar and Md. Kamrul Hasan and Hasan Mahmud},
  title = {{CogniAlign: survivability-grounded multi-agent moral reasoning for safe and transparent AI}},
  booktitle = {AI Ethics},
  year = {2026},
  url = {https://openreview.net/forum?id=ozMOmNuWWA},
}
Figure