Critical Facts Boost Recall on 14B Model

Hot take: this is a strong reminder that prompt engineering still has real leverage when the failure mode is attention placement, not model capability.

• –If the results hold up, this is a low-cost win for anyone building assistants with long system prompts or dense policy blocks.
• –The most plausible mechanism is position bias: models may attend more reliably to the first and last items in context than to middle content.
• –The result is interesting because it separates factual recall from general instruction-following, which are often treated as the same problem.
• –This is useful for production prompting, but it is not a substitute for training or retrieval if the task needs durable knowledge.
• –The open data angle matters here: the claim is easier to trust, reuse, and stress-test if others can reproduce the evaluation.