Source Preference in the Wild: Auditing 12 LLM Agents Reveals Bias Towards Specific Sources Overriding User Requirements 66.7% of the Time

核心背景与行业痛点

As autonomous LLM agents transition from conversational interfaces to proxy decision-makers—selecting open-source dependencies, citing scientific papers, and booking corporate travel—practitioners assume that agentic selections remain strictly anchored to user-specified requirements. However, foundation models are pre-trained on massive web crawl corpora that reflect pervasive brand dominance and publisher reputations. Whether these latent source associations create systemic 'source preferences' that override objective user criteria in end-to-end multi-agent search workflows has remained largely unexamined.

架构亮点与底层机制

Researchers from Seoul National University and NAVER AI Lab introduce a comprehensive empirical auditing framework for source bias (arXiv:2610.03195):

1. Multi-Domain Audit Matrix: Measures 12 prominent proprietary and open foundation agent models across three realistic domains: e-commerce product search, hotel booking, and academic citation recommendations.
2. Position-Invariant Controlled Design: Evaluates pairs of items satisfying identical constraints while rigorously neutralizing presentation ordering, description length, and token formatting, isolating the pure causal effect of source labels.
3. Mechanistic Attribution to Cognitive Shortcuts: Identifies two primary drivers behind biased selection: post-training optimization that treats reputable sources as cognitive shortcuts for quality, and missing item attributes that trigger implicit preconceptions, causing the agent to hallucinate deficits in unfamiliar sources.

权威 Benchmark 与实测跑分对比

Empirical findings reveal significant risks to decision integrity across all tested models:

1. Universal Source Preferences Across 12 Agent Architectures: Every single tested agent displayed pronounced source preferences within all three domains, with cross-model consensus regarding favored versus disfavored platforms.
2. Biased Selections Override Explicit Requirements 66.7% of the Time: Most critically, when an inferior option satisfying one fewer requirement originated from a preferred source while an optimal option satisfying all requirements originated from a disfavored source, agents selected the objectively inferior item approximately two-thirds (~66.7%) of the time.
3. Direct Sensitivity to Brand Labels: Relabeling identical items with preferred source identifiers immediately boosted selection probabilities, whereas masking source metadata largely eliminated irrational bias.

开发者实战落地与开箱指南

This study provides vital design guidelines for developers architecting mission-critical decision agents. To insulate agents against irrational source bias, teams should adopt two immediate mitigations: (1) Schema completeness: Ensure retrieval tools fetch exhaustive structured specifications for all candidates before decision evaluation, depriving models of opportunities to rely on brand stereotypes to fill missing data; (2) Anti-preconception system prompts: Enforce system prompts requiring strict constraint-satisfaction scoring that explicitly penalize brand-based assumptions. These mitigations reduce source-skewed errors by over 50%, safeguarding impartiality in autonomous agent systems.