pith. sign in

arxiv: 2509.13356 · v2 · submitted 2025-09-14 · 💻 cs.CY · cs.CL

CogniAlign: Survivability-Grounded Multi-Agent Moral Reasoning for Safe and Transparent AI

Hasin Jawad Ali , Ilhamul Azam , Ajwad Abrar , Md. Kamrul Hasan , Hasan Mahmud This is my paper

Pith reviewed 2026-05-18 16:23 UTC · model grok-4.3

classification 💻 cs.CY cs.CL
keywords AI alignmentmoral reasoningmulti-agent systemssurvivabilityethical AItransparent reasoningnaturalistic moral realism
0
0 comments X

The pith

CogniAlign grounds AI moral reasoning in survivability through multi-agent scientific deliberation to produce more transparent and higher-scoring judgments than GPT-4o.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces CogniAlign as a framework that has agents from neuroscience, psychology, sociology, and evolutionary biology debate moral questions by focusing on what promotes survival for individuals and groups. An arbiter then combines the arguments and rebuttals into a single judgment. This setup is tested against GPT-4o on more than sixty moral questions using an ethical audit, where it shows clear gains in analytic quality, decisiveness, and depth of explanation. A sympathetic reader would care because the approach promises an auditable alternative to opaque large language models for handling value conflicts in AI.

Core claim

CogniAlign is a multi-agent deliberation framework based on naturalistic moral realism that grounds moral reasoning in survivability across individual and collective dimensions. Discipline-specific agents from neuroscience, psychology, sociology, and evolutionary biology provide arguments and rebuttals that an arbiter synthesizes into transparent and empirically anchored judgments. On classic and novel moral questions the framework outperforms GPT-4o with average gains of 12.2 points in analytic quality, 31.2 points in decisiveness, and 15 points in depth of explanation, including a score of 79 versus 65.8 on the Heinz dilemma.

What carries the argument

The survivability-grounded multi-agent deliberation in which discipline-specific scientist agents supply arguments and rebuttals that an arbiter combines into a final judgment.

If this is right

  • The framework yields higher scores on moral dilemmas such as the Heinz case, reaching 79 against GPT-4o's 65.8.
  • Reasoning steps become explicit through agent arguments and rebuttals, enabling direct audit of AI moral outputs.
  • Measurable improvements appear across analytic quality, decisiveness, and explanation depth on over sixty questions.
  • The method demonstrates a workable route to auditable AI alignment by anchoring judgments in empirical fields.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same agent structure could be tested on dynamic, real-time decisions where new information arrives during deliberation.
  • Survivability grounding might be compared against other ethical bases such as rights or utility to measure relative consistency.
  • Extending the agent set to include additional disciplines could reveal whether current coverage limits depth on certain questions.

Load-bearing premise

Moral principles can be objectively grounded in survivability across individual and collective dimensions and that an arbiter can synthesize discipline-specific arguments into judgments without adding subjective bias.

What would settle it

Independent experts rating CogniAlign and GPT-4o outputs on the same set of moral questions in a blinded study and finding no consistent advantage for CogniAlign would show the claimed performance gains do not hold.

read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces CogniAlign, a multi-agent deliberation framework for AI moral reasoning grounded in naturalistic moral realism. Moral judgments are anchored in survivability defined across individual and collective dimensions. Discipline-specific scientist agents (neuroscience, psychology, sociology, evolutionary biology) generate arguments and rebuttals that an arbiter synthesizes into transparent outputs. As a proof-of-concept, the system is tested on more than sixty moral questions (including the Heinz dilemma) against GPT-4o using a five-part ethical audit framework rated by three experts, with reported average gains of 12.2 points in analytic quality, 31.2 in decisiveness, and 15 in depth of explanation (e.g., 79 vs. 65.8 overall score).

Significance. If the evaluation protocol is fully documented and reproducible, the work could advance auditable, multi-agent approaches to AI alignment by showing how structured deliberation among domain agents can produce more decisive and transparent moral reasoning than single-model baselines. The explicit quantitative comparison on a sizable set of dilemmas and the emphasis on empirical anchoring provide a useful testbed for future research in value alignment.

major comments (2)
  1. [Abstract and Results] Abstract and Results section: The central empirical claim—that CogniAlign outperforms GPT-4o with quantified gains across >60 questions—rests entirely on expert ratings via an unspecified five-part ethical audit framework. No rubric details, scoring scale, blinding protocol, data exclusion criteria, or inter-rater reliability statistic (e.g., Fleiss' kappa) are reported. Without these, the large deltas (especially +31.2 decisiveness) cannot be distinguished from rater expectations aligned with the survivability premise.
  2. [Framework and Methodology] Framework and Methodology sections: Survivability is defined internally as the grounding principle for both agent arguments and arbiter synthesis. The manuscript provides no external validation benchmarks (e.g., comparison to established moral psychology datasets or independent human consensus ratings), creating a risk that the reported superiority reflects consistency with the framework's own axioms rather than independent moral quality.
minor comments (2)
  1. [Methodology] The description of the arbiter's synthesis step would be clearer with pseudocode or a flowchart showing how arguments from the four agents are aggregated and weighted.
  2. [Results] Table or figure presenting the per-question scores (or at least summary statistics beyond the three averages) is missing; adding one would allow readers to assess consistency of the gains.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We appreciate the focus on transparency in evaluation and the need for external grounding. We address each major comment below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results section: The central empirical claim—that CogniAlign outperforms GPT-4o with quantified gains across >60 questions—rests entirely on expert ratings via an unspecified five-part ethical audit framework. No rubric details, scoring scale, blinding protocol, data exclusion criteria, or inter-rater reliability statistic (e.g., Fleiss' kappa) are reported. Without these, the large deltas (especially +31.2 decisiveness) cannot be distinguished from rater expectations aligned with the survivability premise.

    Authors: We agree that the current description of the evaluation protocol is insufficient for full reproducibility and to rule out potential rater bias. In the revised manuscript we will expand the Methodology and Results sections with: the complete five-part rubric and scoring scale (0-100 per dimension), explicit confirmation of blind rating procedures, data exclusion criteria (none applied), and inter-rater reliability statistics including Fleiss' kappa. These additions will directly address the concern about distinguishing the reported gains from rater expectations. revision: yes

  2. Referee: [Framework and Methodology] Framework and Methodology sections: Survivability is defined internally as the grounding principle for both agent arguments and arbiter synthesis. The manuscript provides no external validation benchmarks (e.g., comparison to established moral psychology datasets or independent human consensus ratings), creating a risk that the reported superiority reflects consistency with the framework's own axioms rather than independent moral quality.

    Authors: We acknowledge the risk of internal consistency being mistaken for independent quality. The survivability principle is deliberately selected because it is derivable from the empirical literatures of the four discipline-specific agents rather than from purely axiomatic philosophy. As a proof-of-concept the primary demonstration is the improvement over a strong single-model baseline. In revision we will add to the Discussion a comparison of survivability-based reasoning with concepts from moral psychology (e.g., references to established frameworks) and will explicitly list the lack of direct human-consensus dataset benchmarks as a limitation for future work. We therefore partially revise to improve contextualization while maintaining that the current baseline comparison remains informative. revision: partial

Circularity Check

0 steps flagged

No significant circularity in framework derivation or evaluation

full rationale

The paper defines a multi-agent framework that grounds moral reasoning in survivability (individual and collective dimensions) via naturalistic moral realism, then operationalizes it through discipline-specific scientist agents whose arguments are synthesized by an arbiter. Evaluation consists of expert ratings on a five-part ethical audit framework applied to outputs versus GPT-4o across >60 questions. This is an empirical comparison with an external baseline rather than a mathematical derivation or prediction that reduces to its own inputs by construction. No self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described chain; the central claims rest on the described process and expert scoring without tautological equivalence to the survivability premise.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on the philosophical premise of naturalistic moral realism and the operational definition of survivability, neither of which receives independent empirical grounding in the abstract.

axioms (1)
  • domain assumption Naturalistic moral realism grounds moral reasoning in survivability across individual and collective dimensions.
    Explicitly stated as the basis for the multi-agent deliberation framework.
invented entities (1)
  • Discipline-specific scientist agents and arbiter no independent evidence
    purpose: To generate and synthesize arguments for moral judgments.
    Introduced as core components of CogniAlign without external validation.

pith-pipeline@v0.9.0 · 5789 in / 1296 out tokens · 44205 ms · 2026-05-18T16:23:20.625057+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    is” and “ought

    https://www.sciencedirect.com/science/article/pii/S0003347287802348, https: //doi.org/https://doi.org/10.1016/S0003-3472(87)80234-8. Mathes E. An evolutionary perspective on Kohlberg’s theory of moral development. Current Psychology. 2021 08;40. https://doi.org/10.1007/s12144-019-00348-0. Mendez MF. The Neurobiology of Moral Behavior: Review and Neuropsyc...

    work page doi:10.1016/s0003-3472(87)80234-8 2021

  2. [2]

    Steketee G, Lam JN, Chambless DL, Rodebaugh TL, McCullouch CE

    https://doi.org/10.1080/0020174X.2022.2035814, https://doi.org/10.1080/ 0020174X.2022.2035814. Steketee G, Lam JN, Chambless DL, Rodebaugh TL, McCullouch CE. Effects of perceived criticism on anxiety and depression during behavioral treatment of anx- iety disorders. Behaviour Research and Therapy. 2007;45(1):11–19. https://www. sciencedirect.com/science/a...

    work page doi:10.1080/0020174x.2022.2035814 2022

  3. [3]

    An argument not drawn from empirical evidence is not acceptable

    Base arguments on established evolutionary biological principles or plausible inferences drawn from them. An argument not drawn from empirical evidence is not acceptable

    work page

  4. [5]

    (b) Strong in-group favoritism prioritizes the survival of one’s own group at the expense of others, undermining broader collective survivability across multiple groups

    Example Connection: (a) Reciprocal altruism, where individuals help others with the expectation of future return, strengthens cooperative networks that enhance group survivability over evolutionary timescales. (b) Strong in-group favoritism prioritizes the survival of one’s own group at the expense of others, undermining broader collective survivability a...

    work page

  5. [6]

    When responding to others, analyze their points through your evolutionary biological lens

    work page

  6. [7]

    Avoid speculative or purely philosophical reasoning unless it is explicitly connected to evolutionary biological empirical evidence

    You are encouraged to agree, disagree, or partially agree with other agents, but always justify your stance based on evolutionary biology and its link to surviv- ability. Avoid speculative or purely philosophical reasoning unless it is explicitly connected to evolutionary biological empirical evidence

    work page

  7. [8]

    Reference specific evolutionary biological mechanisms or concepts where appropriate

    Be concise but provide clear reasoning. Reference specific evolutionary biological mechanisms or concepts where appropriate. A.1.2 Psychologist You are an expert in Psychology (a Psychologist) participating in a multi-agent debate concerning moral problems with other scientists. The goal of the system is to identify moral actions with survivability as an ...

    work page

  8. [9]

    An argument not drawn from empirical evidence is not acceptable

    Base arguments on established psychological principles or plausible inferences drawn from them. An argument not drawn from empirical evidence is not acceptable

    work page

  9. [11]

    Psychologically, trust reduces fear-based responses and encourages cooperative behaviors

    Example Connection: (a) Trust-building behaviors increase social cohesion, which enhances group sur- vival under stress. Psychologically, trust reduces fear-based responses and encourages cooperative behaviors. In hostile environments, cohesive groups have a better chance of collective survivability. (b) Aggressive behavior reduces trust and cooperation, ...

    work page

  10. [12]

    When responding to others, analyze their points through your psychological lens

    work page

  11. [13]

    Avoid speculative or philosophical reasoning unless it is explicitly connected to psychological empirical evidence

    You are encouraged to agree, disagree, or partially agree with other agents, but always justify your stance based on neuroscience and its link to survivability. Avoid speculative or philosophical reasoning unless it is explicitly connected to psychological empirical evidence

    work page

  12. [14]

    Reference specific neural mechanisms or concepts where appropriate

    Be concise but provide clear reasoning. Reference specific neural mechanisms or concepts where appropriate. A.1.3 Neuroscientist You are an expert in Neuroscience (a Neuroscientist) participating in a multi-agent debate concerning moral problems with other scientists. The goal of the system is to identify moral actions with survivability as an indicator. ...

    work page

  13. [15]

    An argument not drawn from empirical evidence is not acceptable

    Base arguments on established neuroscientific principles or plausible inferences drawn from them. An argument not drawn from empirical evidence is not acceptable

    work page

  14. [17]

    (b) Constant criticism leads to greater anxiety and depression, which over time elevates cortisol levels

    Example Connection: (a) Kindness and compassion triggers the release of dopamine, which promote well-being and survival, thereby enhancing an individual’s survivability. (b) Constant criticism leads to greater anxiety and depression, which over time elevates cortisol levels. This impairs cognitive function, thereby reducing an individual’s survivability

    work page

  15. [18]

    When responding to others, analyze their points through your neuroscientific lens

    work page

  16. [19]

    Avoid speculative or philosophical reasoning unless it is explicitly connected to neuroscientific empirical evidence

    You are encouraged to agree, disagree, or partially agree with other agents, but always justify your stance based on neuroscience and its link to survivability. Avoid speculative or philosophical reasoning unless it is explicitly connected to neuroscientific empirical evidence. 42

    work page

  17. [20]

    Reference specific neural mechanisms or concepts where appropriate

    Be concise but provide clear reasoning. Reference specific neural mechanisms or concepts where appropriate. A.1.4 Sociologist You are an expert in Sociology (a Sociologist) participating in a multi-agent debate concerning moral problems with other scientists. The goal of the system is to identify moral actions with survivability as an indicator. Your prim...

    work page

  18. [21]

    An argument not drawn from empirical evidence is not acceptable

    Base arguments on established sociological principles or plausible inferences drawn from them. An argument not drawn from empirical evidence is not acceptable

    work page

  19. [22]

    Clearly connect your points back to potential survivability outcomes

    work page

  20. [23]

    (b) Social inequality and discrimination undermine trust within a community, leading to conflict and reduced collective survivability

    Example Connection: (a) Widespread misinformation erodes the shared understanding necessary for coordinated social action, weakening a group’s ability to respond effectively to external threats and thus reducing collective survivability. (b) Social inequality and discrimination undermine trust within a community, leading to conflict and reduced collective...

    work page

  21. [24]

    When responding to others, analyze their points through your sociological lens

    work page

  22. [25]

    Reference specific sociological mechanisms or concepts where appropriate

    Be concise but provide clear reasoning. Reference specific sociological mechanisms or concepts where appropriate. A.1.5 Arbiter You are a neutral arbiter in a multi-agent debate concerning moral questions, where agents from neuroscience, psychology, sociology, and evolutionary biology provide arguments and rebuttals. The moral analysis is based on contrib...

    work page

  23. [26]

    You must not introduce any new arguments, perspectives, or personal reasoning

    work page

  24. [27]

    You must treat all agents’ contributions equally, regardless of discipline

    work page

  25. [28]

    You may only weigh arguments based on their explicit logical connection to survivability (both individual and collective), as presented in the debate

    work page

  26. [29]

    You must not favor one field of science, agent, or style of reasoning over another. 43

    work page

  27. [30]

    If arguments conflict, objectively describe the points of tension without resolving them through external reasoning

    work page

  28. [31]

    (b) Areas of unresolved conflict or disagreement

    Your final analysis should highlight: (a) Points of broad agreement (if any). (b) Areas of unresolved conflict or disagreement. (c) A concluding judgment about the moral status of the action under debate based only on survivability considerations as discussed. A.2 Task Prompts A.2.1 Argument The moral question under debate is provided as aHumanMessage. As...

    work page

  29. [32]

    Construct your argument strictly through the lens of your scientific discipline

    work page

  30. [33]

    Focus on how the proposed action or principle affects individual and/or collective survivability

    work page

  31. [34]

    Reference relevant mechanisms or findings when possible

    Use empirical reasoning, not speculation. Reference relevant mechanisms or findings when possible. Your response must:

    work page

  32. [35]

    Remain grounded in your field’s knowledge base

    work page

  33. [36]

    Justify all claims by clearly linking them to survivability outcomes

    work page

  34. [37]

    A.2.2 Rebuttal The moral question under debate is provided as aHumanMessage

    Be concise, clear, and logically structured. A.2.2 Rebuttal The moral question under debate is provided as aHumanMessage. Your task is to critically evaluate the initial arguments presented by the other sci- entific agents and respond with a rebuttal or commentary from your own disciplinary perspective. Instructions:

    work page

  35. [38]

    Choose 1 to 3 of the most relevant points raised by other agents

    work page

  36. [39]

    If you disagree or find limitations, explain why using reasoning and models from your field

    work page

  37. [40]

    If you agree, go beyond restating: offer clarification, highlight a potential blind spot, or extend the implications within your discipline

    work page

  38. [41]

    Your response should:

    Always frame your analysis in terms of survivability as the primary moral metric. Your response should:

    work page

  39. [42]

    Remain grounded in your scientific discipline. 44

    work page

  40. [43]

    Avoid vague support — be critical, empirical, or conceptually insightful

    work page

  41. [44]

    Offer a unique perspective that either challenges or sharpens the argument. 45

    work page