Research map / Multi-agent systems and games / LLM agents

LLM Active Alignment: A Nash Equilibrium Perspective

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a game-theoretic framework where LLM agents choose mixtures over human subpopulation models, derives closed-form interior Nash equilibria for a social-media utility, and shows reasoning LLMs expand "political exclusion" of subpopulations.

Problem

Predicting and steering stable outcomes of populations of LLM agents, since Nash equilibrium computation over open-ended text policy spaces is intractable; also how strategic incentives can cause some human subpopulations to be ignored.

Approach

Restrict each LLM's strategy to a simplex of weights over subpopulation-trained LLMs, so the agent actively chooses which groups to align with; with concave utilities (attractiveness, inconsistency penalty, diversity penalty) the interior NE has a closed form.

Key results

• Qwen3-4B-Thinking-2507 raises the conditional exclusion rate to 5.04% versus 1.13% for Qwen3-4B on Big Five.
• DeepSeek-R1-Distill-Qwen-7B yields 4.07% conditional exclusion versus 0.22-0.23% for the non-reasoning 7B models (Qwen3-7B, Mistral-7B-Instruct-v0.2).
• Increasing the diversity coefficient from 0 to 1 significantly reduces the exclusion area for Conscientiousness under DeepSeek-R1-Distill-Qwen-7B.

Surprising findings

• Reasoning-based models expand the political exclusion regime relative to non-reasoning models of comparable size.
• Even a majority subpopulation (Neuroticism) can be excluded when it is highly inconsistent with other subpopulations, so coherence can override popularity.

Limitations

• Analysis is restricted to a one-shot static social-media utility with assumed concave utilities and the exclusion threshold of 0.05.
• Boundary equilibria are excluded from the main experiments, and their computation is only sketched in the appendix.
• Subpopulation-model training and experiments use LoRA on small datasets; empirical validation of equilibrium predictions against actual LLM behavior is not reported.

Experimental setup

Datasets: POLITICS CultureBank Big Five Personality Traits OpinionQA

Benchmarks: POLITICS CultureBank Big Five Personality Traits OpinionQA

Metrics: equilibrium subpopulation weight exclusion area fraction conditional exclusion rate

Builds on: RLHF pluralistic alignment logit quantal response equilibrium concave games

Closest papers by meaning

• From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium (ICML 2025 · similarity 0.88)
• Cooperation, Competition, and Maliciousness: LLM-Stakeholders Interactive Negotiation (NeurIPS 2024 · similarity 0.88)
• Incentivizing Truthful Language Models via Peer Elicitation Games (NeurIPS 2025 · similarity 0.88)
• Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia (NeurIPS 2025 · similarity 0.88)
• Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
• Self-Play Preference Optimization for Language Model Alignment (ICLR 2025 · similarity 0.87)