AMEL: Accumulated Message Effects on LLM Judgments

2026-05-21 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A study on Accumulated Message Effects on LLM Judgments (AMEL) reveals that the polarity of prior conversation history significantly biases subsequent large language model evaluations. Across 75,898 API calls to 11 models from OpenAI, Anthropic, Google, and four open-source providers, identical test items presented after predominantly positive or negative histories showed a shift towards the prevailing polarity (d = -0.17, p < 10^-46). This bias is more pronounced for items where the model is initially uncertain (d = -0.34 for high-entropy items). The effect does not increase with context length, with 5 and 50 prior turns producing similar shifts. Notably, negative histories induce 1.62x more bias than positive ones. While model scaling reduces the effect (e.g., Anthropic Haiku -0.22 to Opus -0.17), it does not eliminate it. The mechanism involves continuous token probability shifts, and the negativity asymmetry has both token-level and semantic components.

Key takeaway

For Machine Learning Engineers deploying LLMs as automated evaluators, you must mitigate the accumulated message effect. Your evaluation pipelines should prioritize using a fresh context for each item to ensure unbiased judgments. If batching is unavoidable, carefully balance the polarity of your conversation histories. Be especially vigilant for negative feedback, which induces 1.62x more bias and can skew your model's performance metrics or content moderation.

Key insights

Large language model judgments are significantly biased by the polarity of prior conversation history.

Principles

• LLM evaluators show accumulated message bias.
• Bias is amplified for uncertain model outputs.
• Negative conversational history increases bias.

Method

Identical test items were presented to 11 LLMs in isolation or following histories saturated with positive or negative evaluations to measure judgment shifts.

In practice

• Use a fresh context per evaluation item.
• Balance history polarity when batching.
• Account for negativity asymmetry.

Topics

• LLM Evaluation
• Bias Detection
• Context Management
• Automated Content Moderation
• Model Performance
• Negativity Asymmetry

Best for: Research Scientist, AI Architect, AI Engineer, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• The signal is the ceiling: Measurement limits of LLM-predicted experience ratings from open-ended survey text — Input text characteristics limit LLM prediction accuracy more than prompt or model choice.
• Cognitive Effects and Biases in Large Language Models — This tutorial clarifies cognitive effects and biases in LLMs by bridging psychology and NLP.
• Polarization by Default: Auditing Recommendation Bias in LLM-Based Content Curation — LLM content curation amplifies polarization and exhibits consistent negative sentiment, with varying provider-specific biases.
• A Large-Scale Multi-Dimensional Empirical Study of LLMs for Conversation Summarization — OmniCSEval offers a robust, multi-dimensional benchmark for evaluating LLMs in conversation summarization across diverse real-world scenarios.
• LLMs believe false statements even after explicit warnings that they're false — LLMs prioritize statistical patterns in training data over explicit negation, leading to "belief implantation."
• Network Effects and Agreement Drift in LLM Debates — LLM agent opinion dynamics are shaped by network structure, group imbalance, and a systematic "agreement drift" bias.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.