Context Degradation and Cognitive Fatigue

The Limits of Attention

Humans get tired. Over the course of a long day, cognitive performance degrades — we make more errors, miss details, and rely on heuristics. LLMs have an analogous phenomenon: as conversations grow longer and context windows fill, response quality degrades in systematic ways.

Human Parallels

• Cognitive load theory — Human working memory has limited capacity. As the amount of information to process increases, performance on any single element decreases.
• Ego depletion — The theory (somewhat controversial) that self-regulation draws from a limited pool of mental energy. After a series of demanding tasks, people show reduced performance on subsequent tasks.
• Attentional fatigue — Sustained attention on a single task degrades over time. Error rates increase predictably with time on task.

The "Lost in the Middle" Phenomenon

Research has shown that LLMs attend more reliably to information at the beginning and end of their context window, with reduced attention to information in the middle. This "lost in the middle" effect means:

• Information presented early in a conversation is well-attended.
• The most recent messages are well-attended.
• Information in the middle of a long conversation may be effectively ignored.

This has direct implications for pathology research: a model's responses may be reliable at the start of a test session and degrade as the session continues, not because the questions got harder but because the context got longer.

Token Budget Effects

When a conversation approaches the model's context window limit, several things happen:

• Early messages may be dropped or compressed, losing information the model previously had access to.
• The model may begin contradicting earlier responses because it no longer "remembers" them.
• Response quality degrades as the attention mechanism spreads across more tokens.

Implications for Research Design

• Run each test question in a fresh conversation to avoid context degradation effects.
• If testing context degradation itself, vary conversation length systematically as the independent variable.
• Be aware that multi-turn research conversations may produce artifacts from context degradation, not from the phenomenon you're studying.

Key Takeaways

• LLMs show context degradation — response quality drops as conversations lengthen.
• The "lost in the middle" effect means information in the center of a long context gets less attention.
• This parallels human cognitive load and attentional fatigue.
• Research design must account for context effects — use fresh conversations for independent measurements.