Larger Language Models Compound Errors Faster, Study Finds

A new study indicates a critical paradox in the scaling of large language models (LLMs): as models grow in size and capability, their reliability paradoxically decreases. While they initially produce more accurate responses, they tend to degrade faster, snowballing mistakes. This phenomenon is attributed not just to knowledge gaps, but to an inherent auto-regressive risk that becomes more pronounced with scale. The researchers tracked this degradation using a metric called per-position disagreement, comparing a model's probability distribution to that of a stronger oracle. They found that while the "knowledge gap" (bias) decreases significantly with scaling, the "knowledge degradation" (risk) grows substantially. This leads to a "confident-but-precarious" risk regime where models commit to low-probability tokens and then condition subsequent outputs on these incorrect assumptions, leading to cascading errors or fabrications. Crucially, this auto-regressive risk regime is causal and structurally evades self-monitoring. Models exhibit high confidence even when in this precarious state, making it difficult for internal mechanisms (like semantic entropy) to detect impending errors. The study demonstrated that a targeted intervention to contract variance in this regime could cut web-verified hallucinations by 35-74%, highlighting the potential for mitigation despite the inherent challenge.