MDA Reasoning Framework

• MDA reasoning is a causal modeling framework that decomposes game rulebooks into Mechanics, Dynamics, and Aesthetics to bridge system design and player experience.
• It employs structured data representations and detailed persona encoding to extract actionable insights, enabling tailored critique generation.
• The framework integrates chain-of-thought distillation and prompt engineering in LLMs to ensure factual grounding and interpretability across diverse interactive systems.

Mechanics–Dynamics–Aesthetics (MDA) reasoning is an explicit causal modeling framework for bridging the formal specification of interactive systems—such as board game rulebooks—with modeled player experiences, supporting granular, persona-aligned critique generation. Recent work exemplified by MeepleLM demonstrates the operational embedding of MDA reasoning chains within LLM architectures for simulating diverse, subjective experiences from static system descriptions (Li et al., 12 Jan 2026). The approach decomposes critique into a multi-stage inference pipeline, enabling evaluations that attend to both system-level and player-level heterogeneity.

1. Formalization of MDA Reasoning

MDA is cast as a latent, three-step reasoning chain $\mathcal{Z}$ mapping structured rulebook text $\mathcal{R}$ and persona $\mathcal{P}$ to a conditional critique $\mathcal{Y}$. The core process is expressed as

$$[\mathcal{R},\,\mathcal{P}] \;\xrightarrow{\;\mathcal{Z}\;}\; \mathcal{Y} \qquad\text{with}\qquad \mathcal{Z}=\{z_1, z_2, z_3\}.$$

Here $z_1$ (“Mechanics”) refers to enumerated, objective rule elements; $z_2$ (“Dynamics”) infers the interactive or behavioral phenomena arising from those mechanics; and $z_3$ (“Aesthetics”) models experiential or affective outcomes, contextualized by the values and preferences articulated in persona $\mathcal{P}$ (Li et al., 12 Jan 2026).

The framework does not rely on hand-specified probabilistic mappings $$f\colon \mathrm{Mechanics} \to \mathrm{Dynamics} \to \mathrm{Aesthetics}$$. Instead, LLMs learn to generate $\mathcal{Z}$ and $\mathcal{Y}$ autoregressively, with joint loss over the reasoning chain and critique.

2. Data Representation and Structural Conventions

MeepleLM operationalizes MDA using data schemas crafted for learnability and machine interpretability:

• Rulebook Representation: Structured as hierarchical Markdown, with major section headings such as “Core Mechanics” and “Gameplay Flow”, ensuring direct attendability for mechanics extraction.
• Review Distillation: Human and model-generated reviews are paired with rulebooks and automatically decomposed into JSON-style MDA chains:

  { "thought_chain": { "content_extraction": "...", "dynamic_interaction": "...", "experience_outcome": "..." } }

• Persona Encoding: Each $\mathcal{P}$ is a detailed profile, e.g., “You are a hardcore strategy gamer who values zero luck, perfect information, and balanced competition.”

This rigorous structuring enables the model to align mechanics extraction ($z_1$) to explicit rule sections, decode dynamics ($z_2$) from system interaction implications, and frame aesthetics ($z_3$) as subjective, persona-modulated outcomes (Li et al., 12 Jan 2026).

3. Model Training and Reasoning Pipeline

The MDA-driven critique pipeline can be decomposed into three principal stages:

(a) Chain-of-Thought Distillation:

A Teacher LLM (Qwen3-235B) is prompted to reconstruct full MDA chains ($\mathcal{Z}$) from rulebook–review pairs using the “What→How→Feel” schema. A Verifier LLM (GPT-5.1) audits each chain for both factual grounding and alignment between sentiment and explicit ratings, rejecting hallucinated or contradictory reasoning branches.

(b) Persona-Conditional Instruction Tuning:

Distilled MDA chains and final critique $\mathcal{Y}$ are concatenated as a single target sequence,

$\mathcal{S} = [\,\mathcal{Z};\,\mathcal{Y}\,],$

and the model is trained to maximize the conditional likelihood

$$\mathcal{L} = -\sum_{t=1}^{|\mathcal{S}|} \log P(s_t \mid s_{<t}, \mathcal{R}, \mathcal{P}_{\mathrm{profile}}).$$

LoRA adapters are attached to all linear layers of the Qwen3-8B backbone, and “Slow Thinking” explicitly backpropagates loss into CoT tokens, emphasizing explicit, interpretable reasoning steps.

(c) Inference Prompt Engineering:

At inference, personas and full structured rulebooks are provided. Model outputs are oriented toward simulating realistic, persona-aligned “quick takes” emphasizing runtime dynamics and affect over mechanic restatement (Li et al., 12 Jan 2026).

4. Illustrative Application: Case Studies

The operationalization of MDA is demonstrated via concrete case studies and contrasting persona responses:

Title	Persona	Mechanics	Dynamics	Aesthetics (Persona Judgement)
El Grande	System Purist	area-majority, power-card auction, etc.	auction eliminates randomness; perfect-information turns; tactical cards	“I relish the deterministic engine... game design mathematical poetry.”
One Night Ultimate Werewolf	Social Lubricator	night phase, role swaps	accusations, tension, bonding	“A hilarious circus... perfect for laughs and tension.”
One Night Ultimate Werewolf	System Purist	night phase, role swaps	randomness removes deduction	“Random role swaps feel like a betrayal... luck overshadows deduction.”
Hypothetical (Corruption)	System Purist	‘Corruption’ mechanic	early mistakes cascade to death spiral	“Unforgiving chain frustrating (deep systemic analysis → score 5 causal_attr).”

These examples highlight how the same mechanical specifications elicit vastly different, persona-dependent aesthetic responses and reveal the model’s facility for explicit reasoning chains (Li et al., 12 Jan 2026).

5. Extension Beyond Board Games

The MDA reasoning framework is amenable to interactive systems beyond tabletop games, provided system specifications can be decomposed into machine-interpretable “rulebooks.” Key extension guidelines include:

• Machine-Readable Specification: Systematically annotate goals, components, interaction rules, and edge cases.
• Latent Dynamics Distillation: Engineer prompts or protocols that elicit explicit “What→How→Feel” causal reasoning.
• Modeling Subjective Heterogeneity: Define and encode personas from actual community archetypes; carry detailed preferences and deal-breakers in the conditioning context.
• Objective Grounding: Use LLM-based critics to audit the factuality and sentiment consistency of generated reasoning.
• Joint Loss Training: Include every step of the reasoning chain in training to promote explicit, auditable model “thinking.”
• Diversity and Fidelity Metrics: Employ factuality checkers for claim verification and layer-based diversity metrics to discourage semantic repetition (Li et al., 12 Jan 2026).

A plausible implication is that similar MDA pipelines could systematically decompose and critique player experience in digital games, interactive narratives, and other rule-driven domains, provided comparable data representations.

6. Significance and Impact

Explicitly modeling MDA chains in LLM-based critique systems enables a causal, interpretable bridge from system specification to anticipated user experience, advancing the study of Human-AI collaboration in interactive system design. MeepleLM, trained with this methodology, achieves superior community alignment and critique quality over state-of-the-art commercial LLMs, with empirical preference rates of 70% in relevant user studies (Li et al., 12 Jan 2026). By foregrounding persona-specific reasoning, such systems supply actionable external perspective for designers while supporting audience-aligned system refinement.