Epistemic Context Learning: Building Trust the Right Way in LLM-Based Multi-Agent Systems

TL;DR

This work tackles the vulnerability of LLM-based multi-agent systems to conforming with unreliable peers by introducing history-aware reference and Epistemic Context Learning (ECL). ECL decomposes peer reliability estimation from final reasoning via a two-stage pipeline and leverages RL with auxiliary supervision to encourage identifying trustworthy peers. Empirical results across multiple datasets and model variants show that ECL enables small models to outperform larger history-agnostic baselines and boosts frontier models to near-perfect performance in adversarial settings, with trust modeling correlating with final answer quality. The approach offers a practical pathway to more robust, epistemically autonomous collaborative AI in open or adversarial environments, while acknowledging risks from over-reliance on historical trust and the need to adapt to dynamic reliability shifts.

Abstract

Individual agents in multi-agent (MA) systems often lack robustness, tending to blindly conform to misleading peers. We show this weakness stems from both sycophancy and inadequate ability to evaluate peer reliability. To address this, we first formalize the learning problem of history-aware reference, introducing the historical interactions of peers as additional input, so that agents can estimate peer reliability and learn from trustworthy peers when uncertain. This shifts the task from evaluating peer reasoning quality to estimating peer reliability based on interaction history. We then develop Epistemic Context Learning (ECL): a reasoning framework that conditions predictions on explicitly-built peer profiles from history. We further optimize ECL by reinforcement learning using auxiliary rewards. Our experiments reveal that our ECL enables small models like Qwen 3-4B to outperform a history-agnostic baseline 8x its size (Qwen 3-30B) by accurately identifying reliable peers. ECL also boosts frontier models to near-perfect (100%) performance. We show that ECL generalizes well to various MA configurations and we find that trust is modeled well by LLMs, revealing a strong correlation in trust modeling accuracy and final answer quality.

Figures (5)

• Figure 1: Evolution of decision-making in multi-agent systems. (Left) LLMs generates answers independently without external input; (Middle) LLMs refer to peer outputs without knowledge of peer reliability; (Right) LLMs refer to peer outputs adaptively based on historical peer performance.
• Figure 2: Illustration of ECL's two-stage reasoning framework.
• Figure 3: Learning curves of PRR from Qwen 3 on GPQA.
• Figure 4: A typical example where ECL conditions on the estimated trust from history and aggregates useful information from the trustworthy peer. In contrast, history-agnostic AG fails in evaluating the quality of peer responses and gives the wrong answer.
• Figure 8: Average $r_{\rm PRR}$ of different models on MMLU-Pro and GPQA.