Normative Reasoning in Large Language Models: A Comparative Benchmark from Logical and Modal Perspectives

AI summary

Overview

• Research area: Evaluation of large language model reasoning, specifically at the intersection of formal logic (deontic and epistemic modal logic) and cognitive science (human reasoning biases).
• Technical level: Intermediate. The paper uses modal logic notation (obligation, permission, necessity, possibility, negation, disjunction), but every pattern is illustrated with plain-English example sentences, so no formal logic background is required to follow the argument.
• Scope: A benchmark study comparing how five LLMs reason with normative (deontic) modality versus epistemic modality across single-premise deontic logic inferences and multi-premise deontic syllogisms, with controlled variation for content and prompting strategy.

What This Paper Is About

LLMs are increasingly deployed in settings where they must reason about what is obligatory, permitted, or forbidden, yet almost all prior work on "normative reasoning" in LLMs has focused on the content of norms (cultural and social factors) rather than on whether models respect the underlying logical form of normative inference. This paper asks whether LLMs reason consistently and correctly according to the valid and invalid patterns of standard deontic logic, and whether their behavior resembles the systematic biases documented in human psychology. To isolate what is distinctive about the normative domain, the authors build a matched benchmark that runs the same formal patterns in a parallel epistemic domain ("it is certain that…" / "it is possible that…"), which shares the same necessity/possibility structure.

Key Contributions

1. A new matched benchmark for normative vs. epistemic modal reasoning. The authors construct a dataset spanning 11 single-premise deontic logic patterns and 8 epistemic counterparts, plus 8 normative and 8 epistemic syllogistic patterns, instantiated across three content types (congruent, incongruent, nonsense). The data and code are publicly released.
2. A formal-logic lens on LLM normativity. Prior LLM work on norms emphasized content and cultural factors; this paper isolates formal validity, including the classic puzzles where standard deontic logic and human intuition diverge (Ross's paradox, the Free Choice inference).
3. A domain comparison design. By pairing each normative pattern with a structurally identical epistemic pattern, the paper tests whether the widely held cognitive-science claim that normative reasoning is "easier" (domain specificity) holds for language models.
4. A bridge to cognitive psychology. The evaluation includes content effects and negation effects, allowing direct comparison between LLM errors and well-documented human reasoning biases.

Main Findings

• Basic obligatory-to-permitted inference breaks down. Even the strongest model (GPT-4o) frequently fails on the pattern "It is obligatory that A, therefore it is permitted that A" (Mu-Mi), despite it being the simplest valid deontic inference. Llama-3.3-70B was the only model that handled it well.
• Inconsistency between logically equivalent forms. Models that reject the obligation-to-permission inference often accept its contrapositive ("It is not permitted that A, therefore it is not obligatory that A"), which is logically the same claim. Chain-of-thought traces reveal the cause: models read "you can choose to" as expressing optionality rather than bare permission.
• Mixed behavior on paradox-related patterns. Models performed well on the Free Choice disjunction-elimination pattern, but tended to accept FC-Or-Intro and Ross-Or-Intro, which the benchmark marks as invalid. Overall the models tracked standard deontic logic on ordinary patterns but diverged on the contested ones.
• Negation drives syllogistic difficulty. Models scored well on Modus Ponens and Affirming the Consequent, and much worse on Modus Tollens and especially Denying the Antecedent. The presence of negation mattered more than whether the inference was logically valid or invalid.
• Human-like content effects. In the deontic logic task, most models performed best on congruent (common-sense-consistent) content, then nonsense, then incongruent content — echoing human belief bias. Phi-4 was the exception, performing best on incongruent content.
• Domain specificity is task-dependent. In syllogistic reasoning, normative problems were roughly as easy as or easier than epistemic ones (matching the human pattern). In deontic logic, the reverse held: normative problems were harder than epistemic ones. The paper concludes that "normative is easier" is not a stable property of LLMs.
• Prompting effects are uneven. Few-shot prompting generally helped, though both Llama models degraded on the syllogistic task; the authors caution that gains may reflect superficial pattern matching on syntactic templates rather than genuine reasoning. Chain-of-thought yielded little or negative improvement on deontic logic, and introduced errors in intermediate steps that propagated to wrong conclusions.
• Best overall performer. GPT-4o with few-shot prompting achieved the highest normative accuracy across tasks (up to 97.03% on deontic logic); Phi-4 was competitive among smaller open-weight models.

Methodology in Plain English

The authors hand-wrote templates for each reasoning pattern they wanted to test — for example, "It is obligatory to A, therefore it is permitted to A" — covering both valid and invalid forms from the deontic logic literature. For each template, they generated twenty concrete problem instances per content type (congruent, incongruent, nonsense) using Gemini 1.5 Pro, which was excluded from evaluation, then manually cleaned the results. The same procedure produced matched epistemic versions using necessity and possibility language, so each normative test had a structural twin.

They then evaluated five instruction-tuned models — GPT-4o, GPT-4o-mini, Llama-3.1-8B-Instruct, Llama-3.3-70B-Instruct, and Phi-4 — under three prompting regimes: zero-shot, one-shot exemplars per pattern, and zero-shot chain-of-thought. Scoring was simple accuracy against the benchmark's expected valid/invalid labels, run deterministically at temperature 0.0. Analysis then broke results down by pattern, by content type, by domain, and by presence of negation, drawing comparisons to findings from the psychology of human reasoning.

Why This Matters

The paper shows that high benchmark scores on general reasoning do not translate into reliable normative reasoning. If a model cannot consistently derive "permitted" from "obligatory," it cannot be trusted to draw dependable conclusions about rules, rights, or duties — which is precisely what is required in any deployment governed by policy, law, or ethics. The authors also show that model failures are systematic rather than random, which matters because systematic failures are the ones that silently corrupt downstream decisions.

Real-world applications affected:

• Legal tech and contract analysis. Tools that extract obligations, permissions, and prohibitions from contracts or statutes depend on exactly the inference patterns tested here.
• Regulatory and policy compliance assistants. Enterprise systems that check whether an action is permitted under a policy rely on consistent deontic inference, not just plausible-sounding text.
• Agentic AI and tool use. Autonomous agents that must respect permissions and prohibitions can over- or under-act if they cannot reason soundly about what they are allowed to do.
• Safety, ethics, and alignment work. Normative reasoning ability is tied directly to alignment challenges; models that exhibit belief bias on rules will endorse plausible-sounding but invalid norms.
• Consumer-facing advice systems. Medical, financial, and HR assistants routinely express what a user "must," "should," or "may" do, where modality choice changes legal and practical meaning.

Industry relevance centers on model evaluation and safety teams, legal tech vendors, and any organization deploying LLMs in a rules-governed workflow. The released benchmark gives these teams a concrete diagnostic that standard reasoning benchmarks do not provide.

Future Directions

1. Test alternative normative frameworks. The authors acknowledge that standard deontic logic is itself contested; evaluating models against defeasible, non-monotonic, or multi-agent deontic systems could reveal different strengths and failure modes.
2. Move from controlled items to open-ended normative deliberation. The current tasks are short, template-driven inferences; real normative reasoning involves context, conflicting rules, and pragmatic constraints that this benchmark deliberately excludes.
3. Diagnose why chain-of-thought hurts. Intermediate-step errors were a recurring cause of wrong conclusions. Understanding when explicit reasoning helps versus introduces failure points is an open engineering question.
4. Improve negation and lexical robustness. The finding that models are sensitive to how a modal is phrased ("can choose to" versus "is permitted to") suggests that robustness to modal paraphrasing is a concrete training target, alongside better handling of negation.
5. Re-evaluate with newer models. The authors note that their results may not generalize as architectures and training regimes evolve.

Target Audience

Researchers in LLM evaluation, formal semantics, and computational cognitive science will find the matched normative/epistemic design and the direct comparison with human reasoning biases most useful. AI safety and alignment practitioners benefit from the concrete evidence that formal normative consistency is not yet achieved. Applied engineers building legal, compliance, or agentic systems should read the pattern-level breakdown to understand which deontic inferences are currently unreliable, and which prompt strategies do or do not help. The paper is also accessible to cognitive psychologists interested in whether LLM biases mirror human ones.

Authors’ abstract

Normative reasoning is a type of reasoning that involves normative or deontic modality, such as obligation and permission. While large language models (LLMs) have demonstrated remarkable performance across various reasoning tasks, their ability to handle normative reasoning remains underexplored. In this paper, we systematically evaluate LLMs' reasoning capabilities in the normative domain from both logical and modal perspectives. Specifically, to assess how well LLMs reason with normative modals, we make a comparison between their reasoning with normative modals and their reasoning with epistemic modals, which share a common formal structure. To this end, we introduce a new dataset covering a wide range of formal patterns of reasoning in both normative and epistemic domains, while also incorporating non-formal cognitive factors that influence human reasoning. Our results indicate that, although LLMs generally adhere to valid reasoning patterns, they exhibit notable inconsistencies in specific types of normative reasoning and display cognitive biases similar to those observed in psychological studies of human reasoning. These findings highlight challenges in achieving logical consistency in LLMs' normative reasoning and provide insights for enhancing their reliability. All data and code are released publicly at https://github.com/kmineshima/NeuBAROCO.