This paper has been superseded
This paper has been consolidated into Context Management in Autoregressive Language Models (DOI: 10.5281/zenodo.19431948). The content below is retained for reference.
Executive Summary
Paper XIX showed that priming selection closes 48.9% of the coordination gap between neutral and expert processing, but did not identify which structural features of a priming sequence produce this effect. This paper systematically varies five features on Qwen 2.5-7B across 160 domain probes: sequence length, structural regularity, within-session repetition, vocabulary density, and cold-start reset.
The key discoveries overturn several intuitive assumptions about coordination. First, effective priming is remarkably compact: just 15 tokens capture 98.8% of the full benefit, meaning that coordination is about triggering a mode switch rather than transferring detailed information. Second, natural multi-turn conversation outperforms rigid call-response ritual by 0.48 nats; the rigid format strips the processing dynamics that natural conversation carries. Third, within-session repetition is actively harmful, degrading performance by +0.07 nats per repeat — within a single forward pass, repeated content is redundant rather than reinforcing.
The protocol that emerges is simple: reset then prime. A 15-token reset instruction ("disregard any prior context") followed by domain priming reduces cross-entropy from 1.20 to 0.733, a 39% improvement that surpasses the expert's own priming baseline. Artificially named patterns ("TROPONIN CASCADE") are counterproductive (+0.568 nats worse than natural vocabulary): imposed labels fail, whereas Paper XIX found that labels the receiver generates itself help.
Key Findings
- Extreme compactness: 15 tokens capture 98.8% of coordination benefit — coordination is a mode switch, not information transfer
- Natural beats rigid: Natural conversation outperforms rigid call-response ritual by 0.48 nats
- Repetition degrades: Each within-session repetition adds +0.07 nats
- Forced naming harmful: Artificially named patterns worse than natural vocabulary by +0.568 nats
- Reset then prime: Reset instruction before domain priming achieves CE 0.733, a 39% improvement over direct priming
Key References
- McEntire (2026) — Ensemble Gravity: priming selection closes 48.9% of the coordination gap (Paper XIX)
- McEntire (2026) — Sender Continuation Perplexity: KV-cache teacher-forcing protocol (Paper XVIII)