Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Haziq Mohammad Khalid, Athikash Jeyaganthan, Timothy Do, Yicheng Fu, Sean O'Brien, Vasu Sharma, Kevin Zhu

TL;DR

ERGO tackles multi-turn degradation in large language models by monitoring token-level uncertainty via Shannon entropy over next-token distributions and triggering adaptive prompt consolidation when entropy spikes are detected. The method uses a lightweight, model-agnostic intervention that rewrites the user prompt to a single-turn optimized form, followed by isolated generation and a branch continuation, thereby realigning context without altering internal model computations. Across five generation tasks and multiple models, ERGO achieves a 56.6% average performance gain, a 24.7% improvement in aptitude, and a 35.3% reduction in unreliability relative to standard multi-turn baselines, outperforming existing intervention strategies. These results demonstrate that uncertainty-aware input restructuring can significantly enhance both accuracy and reliability in real-world, long-horizon conversational AI applications.

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.

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

TL;DR

ERGO tackles multi-turn degradation in large language models by monitoring token-level uncertainty via Shannon entropy over next-token distributions and triggering adaptive prompt consolidation when entropy spikes are detected. The method uses a lightweight, model-agnostic intervention that rewrites the user prompt to a single-turn optimized form, followed by isolated generation and a branch continuation, thereby realigning context without altering internal model computations. Across five generation tasks and multiple models, ERGO achieves a 56.6% average performance gain, a 24.7% improvement in aptitude, and a 35.3% reduction in unreliability relative to standard multi-turn baselines, outperforming existing intervention strategies. These results demonstrate that uncertainty-aware input restructuring can significantly enhance both accuracy and reliability in real-world, long-horizon conversational AI applications.

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.
Paper Structure (31 sections, 6 equations, 5 figures, 5 tables)

This paper contains 31 sections, 6 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background and Related Works
  3. Entropy Based Uncertainty Estimation
  4. Inference-Time Interventions
  5. Backtracking and Prompt Restructuring
  6. Entropy-Guided Context Resetting
  7. Rise in Average Token Level Entropy
  8. Threshold-Based Trigger for Context Reset
  9. Context Reset Protocol
  10. Experimentation Background
  11. Simulation Scale & Parameters
  12. Tasks
  13. Metric Selection
  14. Per-Run Scoring
  15. Aggregate Metrics
  16. ...and 16 more sections

Figures (5)

  • Figure 1: Illustrative comparison of a standard multi-turn conversational AI and the ERGO system
  • Figure 2: Example Llama3.1-8B run on a GSM8K question with FULL, SHARDED and ERGO settings. Each row represents a separate prompt given to the model while each table represents a context window.
  • Figure 3: Effect of SHARDED and ERGO on Aptitude and Unreliability. Icons represent models FULL performance. Green dots represent performance with ERGO while red dots represent SHARDED performance
  • Figure 4: Comparison of performance point gains (percentage-point increase in accuracy relative to SHARDED) and number of resets across entropy-guided, random, and quintet reset methods on Database, Actions, and Math tasks. Icons represent their respective task with their color determining method used.
  • Figure 5: Comparison of maximum performance points gains (highest increase in accuracy when compared to ) and number of resets between different thresholds across Database, Actions, and Math tasks.