Skin-in-the-Game: Decision Making via Multi-Stakeholder Alignment in LLMs

TL;DR

The paper tackles the challenge of moral reasoning in LLMs by introducing Skin-in-the-Game (SKIG), a multi-turn, multi-perspective framework that simulates accountability by evaluating decisions from numerous stakeholder viewpoints. SKIG formalizes decision-making as an implicit mesa-optimization over aggregated stakeholder utilities, incorporating a scenario generator, an aggregation mechanism, and a scenario evaluator, with theoretical generalization guarantees that improve as the number of simulated scenarios increases and the LLM's modeling capacity improves. Empirically, SKIG outperforms standard prompting, chain-of-thought, and Thought Experiment baselines across multiple moral benchmarks (MMLU Moral Scenarios, Moral Stories, ETHICS Commonsense Morality, Social Chemistry 101) and models, with ablations showing empathy and risk assessment as the most impactful components. The work demonstrates that multi-turn, stakeholder-aware prompting yields more consistent, robust, and ethically aligned decisions, offering a path toward safer and more responsible AI-assisted decision making in ethically nuanced domains; limitations include domain scope and risk of generating harmful outputs that require mitigation strategies.

Abstract

Large Language Models (LLMs) have shown remarkable capabilities in tasks such as summarization, arithmetic reasoning, and question answering. However, they encounter significant challenges in the domain of moral reasoning and ethical decision-making, especially in complex scenarios with multiple stakeholders. This paper introduces the Skin-in-the-Game (SKIG) framework, aimed at enhancing moral reasoning in LLMs by exploring decisions' consequences from multiple stakeholder perspectives. Central to SKIG's mechanism is simulating accountability for actions, which, alongside empathy exercises and risk assessment, is pivotal to its effectiveness. We validate SKIG's performance across various moral reasoning benchmarks with proprietary and opensource LLMs, and investigate its crucial components through extensive ablation analyses.

Figures (5)

• Figure 1: Illustration outlining various strategies for tackling reasoning problems with LLMs. The red box contains existing methods that use single-turn methods Standard Prompting and zero-shot Chain-of-Thought. The blue box contains Thought Experiment, a multi-turn single-perspective framework. The green box contains SKIG, our proposed multi-turn multi-perspective reasoning framework.
• Figure 2: Ablation Analysis on MMLU Moral Scenarios, Moral Stories and ETHICS Commonsense Morality datasets comparing the improvement in accuracy resulting from each of the components in SKIG framework.
• Figure 3: Skin in the Game Workflow. Each box signifies a distinct thought, functioning as a unified string of words that forms an incremental pathway to reasoning.
• Figure 4: Illustration outlining various strategies for tackling reasoning problems with LLMs. The red box contains existing methods that use single-turn methods Standard Prompting and zero-shot Chain-of-Thought. The blue box contains Thought Experiment, a multi-turn single-perspective framework. The green box contains SKIG, our proposed multi-turn multi-perspective reasoning framework.
• Figure 5: Flowchart detailing SKIG reasoning stages in the context of an example. The stakeholder identification process, followed by motivation analysis, consequence exploration and risk assessment are shown as radiating semi-circles following each other respectively.

Theorems & Definitions (3)

• Theorem 3.1
• Theorem F.1
• proof