Peripheral Attention Engineering: Structured Peripheral Context Improves LLM Decision Quality

Abstract

Large language models answer queries using directly relevant material, but they lack the lateral, situational context a competent human consults silently before responding. This paper introduces Peripheral Attention Engineering (PAE): the deliberate structuring and provision of unrequested-but-decision-relevant context to shape LLM responses. A controlled within-subjects experiment across four domains (customer service, medical triage, legal advisory, financial advisory; N=500 total scenarios) and six conditions tests whether structured peripheral context improves response quality when controlling for token count, domain relevance, and prompt structure. Results confirm all preregistered hypotheses: peripheral context produces large gains in response quality (baseline mean=1.63–2.06 vs anomaly-flagged mean=4.01–4.90 on a 5-point scale, Cohen's d=2.02–3.10, all p<0.0001); the improvement is attributable to informational content, not prompt length (noise d=0.04, n.s.) or domain relevance (expanded direct d=0.06, n.s.); slot-structured context outperforms equivalent unstructured context (d=0.13, p<0.0001); and the effect generalizes across three model families and all four domains. A Phase 2 fine-tuning experiment demonstrates that contextual reach — the unprompted instinct to retrieve peripheral context before answering — is trainable, with a fine-tuned Gemma 4 12B model achieving 100% context-request and context-use rates, and outperforming all Phase 1 generators on the full benchmark. A knowledge-graph substrate (Phase 3) and a practical Python library (pae.context.wrap()) are released as open-source artifacts.

Keywords