ERGO: Entropy-guided Resetting for Generation Optimization in Multi-turn Language Models

2025-10-15

AI Summary

In multi-turn dialogues, the incremental presentation of information causes significant performance degradation in large language models (LLMs), hindering their practical deployment. To address this, we propose a model uncertainty-driven dynamic context regulation method: for the first time, we employ Shannon entropy—computed over the next-token probability distribution—as a semantic misalignment detection signal; entropy surges trigger adaptive prompt integration and context reset, enabling proactive uncertainty mitigation. Our approach comprises three core components: dynamic entropy-threshold detection, context pruning, and prompt fusion. Evaluated on progressive instruction-following tasks, it achieves an average accuracy gain of 56.6%, a 24.7% improvement in peak capability, and a 35.3% reduction in performance volatility. These results demonstrate substantial enhancements in dialogue system stability, reliability, and consistency.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Large Language Models (LLMs) suffer significant performance degradation in multi-turn conversations when information is presented incrementally. Given that multi-turn conversations characterize everyday interactions with LLMs, this degradation poses a severe challenge to real world usability. We hypothesize that abrupt increases in model uncertainty signal misalignment in multi-turn LLM interactions, and we exploit this insight to dynamically realign conversational context. We introduce ERGO (Entropy-guided Resetting for Generation Optimization), which continuously quantifies internal uncertainty via Shannon entropy over next token distributions and triggers adaptive prompt consolidation when a sharp spike in entropy is detected. By treating uncertainty as a first class signal rather than a nuisance to eliminate, ERGO embraces variability in language and modeling, representing and responding to uncertainty. In multi-turn tasks with incrementally revealed instructions, ERGO yields a 56.6% average performance gain over standard baselines, increases aptitude (peak performance capability) by 24.7%, and decreases unreliability (variability in performance) by 35.3%, demonstrating that uncertainty aware interventions can improve both accuracy and reliability in conversational AI.

Problem

• Addresses performance degradation in multi-turn LLM conversations
• Mitigates model uncertainty spikes via entropy-guided resetting
• Improves accuracy and reliability with adaptive prompt consolidation

Innovation

• Uses entropy to detect model uncertainty spikes
• Triggers adaptive prompt consolidation dynamically
• Improves multi-turn conversation accuracy and reliability