When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction

2026-05-15 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

Large language models (LLMs) reliably follow single-turn instructions but degrade in long multi-turn interactions, losing track of initial instructions, persona, and rules. This study provides a mechanistic explanation, proposing a "channel-transition" account where goal-defining tokens become less accessible via attention, while goal-related information may persist in residual representations. Researchers introduce the Goal Accessibility Ratio (GAR) to measure attention to task-defining goal tokens and combine it with sliding-window ablations and residual-stream probes. They found that GAR monotonically declines across all tested architectures (Mistral, Qwen, LLaMA, Mixtral), and forcing attention channel closure via a sliding-window intervention causes predictable behavioral failures. Despite attention loss, some models preserve substantial goal-conditioned behavior, with linear probes recovering per-episode recall outcomes with AUC up to 0.99 from residual representations, indicating architecture-specific encoding depths.

Key takeaway

For AI engineers developing long-horizon conversational agents, your focus should extend beyond merely extending context windows. Prioritize architectural designs that maintain robust goal representations in the residual stream, even after direct attention to initial instructions fades. This ensures consistent instruction following and persona compliance in prolonged interactions, mitigating degradation and improving reliability.

Key insights

LLM multi-turn instruction-following degradation stems from attention channel closure, with residual stream encoding determining post-closure behavior.

Principles

• Attention to goal tokens decays over multi-turn interactions.
• Residual streams can retain goal information even without direct attention.
• Architectural design influences goal information retention capacity.

Method

The Goal Accessibility Ratio (GAR) quantifies attention to goal tokens. Sliding-window ablations causally close the attention channel. Linear probes on residual activations measure goal information encoding capacity.

In practice

• Use GAR as a diagnostic for attention channel openness.
• Implement sliding-window interventions to test channel closure effects.
• Employ linear probes to assess residual stream encoding capacity.

Topics

• Multi-Turn LLM Degradation
• Attention Channel
• Residual Stream
• Goal Accessibility Ratio
• Sliding-Window Intervention

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

Related on AIssential

• Understanding Generalization and Forgetting in In-Context Continual Learning — Standard attention mechanisms in LLMs inherently cause intertask interference and systematic bias in sequential in-context learning.
• Talking to a Know-It-All GPT or a Second-Guesser Claude? How Repair reveals unreliable Multi-Turn Behavior in LLMs — LLM multi-turn conversational repair behavior is unreliable and highly model-specific, challenging consistent user interaction.
• When the Loop Closes: Architectural Limits of In-Context Isolation, Metacognitive Co-option, and the Two-Target Design Problem in Human-LLM Systems — Prompt-level isolation is architecturally insufficient for multi-modal LLM systems due to context contamination within the attention window.
• A Guide to Context Engineering for LLMs — More context can degrade LLM performance due to architectural limitations and uneven attention distribution.
• Relational Intervention During Functional Collapse in Large Language Models: A Lexical-Statistical Ablation and a Structure x Register Factorial — Relational interventions combining structure and first-person register significantly improve LLM behavior during functional collapse.
• Your AI Agent Isn’t Dumb. It Has ADHD — LLMs forget instructions mid-conversation due to context window attention bias, causing significant performance drops.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.