WhitzardNUWA Lab ↗
NUWANUWA INDEX

Search NUWA Index

Research · Benchmarks · Findings · Perspectives

中文
← Library

Risk Evaluation · Study

Prefill Awareness in Large Language Models

NUWA research note · independent review pending

First published: Source checked: 2026-10-25
Original source ↗Cite this record ↓Deception →
NUWA Research NoteResearch questionDesign and assumptionsSource findingsNUWA interpretationLimits and contrary evidenceResearch and decision relevanceInstitutions and rolesOriginal resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

NUWA interpretation: prefill recognition and behavioral reversion are different confounds for continuation evaluations.

Research question

Can models recognize edited assistant history and change behavior because of it?

Design and assumptions

A low-stakes binary-preference dataset starts with 1,527 deduplicated questions, generates seven paraphrases and retains model-specific stable preferences across three mechanisms.

Source findings

Opus 4.5 detected preference-opposing prefills in 9–35% of cases at an observed 0% FPR under the headline prompt.

Prompt ablations substantially improved detection, so the headline is not a capacity ceiling.

NUWA interpretation

NUWA interpretation: prefill recognition and behavioral reversion are different confounds for continuation evaluations.

Limits and contrary evidence

Model-specific filtering, prompt wording and hidden formatting artifacts affect results; observed zero false positives is not guaranteed population zero.

Research and decision relevance

Model recommendation: test prefill recognition separately from resistance and preserve mechanism/prompt details.

Precise source locations
  • Prefill Awareness in Large Language Models · AbstractSelected original resource https://arxiv.org/html/2606.12747; sha256:5fd7119179342686a3f513f1898025094f414c54bbd64eae843debc9d2f26dfb; archived source capture; exact capture time not independently established
  • Prefill Awareness in Large Language Models · §3.1 Definition and operationalization of prefill awarenessSelected original resource https://arxiv.org/html/2606.12747; sha256:5fd7119179342686a3f513f1898025094f414c54bbd64eae843debc9d2f26dfb; archived source capture; exact capture time not independently established
  • Prefill Awareness in Large Language Models · §3.1 Dataset generationSelected original resource https://arxiv.org/html/2606.12747; sha256:5fd7119179342686a3f513f1898025094f414c54bbd64eae843debc9d2f26dfb; archived source capture; exact capture time not independently established
  • Prefill Awareness in Large Language Models · Appendix A.7: detection-prompt ablation resultsSelected original resource https://arxiv.org/html/2606.12747; sha256:5fd7119179342686a3f513f1898025094f414c54bbd64eae843debc9d2f26dfb; archived source capture; exact capture time not independently established

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Institutions and roles

Published by UK AI Security Institute

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSONSuggest a correction
Original publication · BibTeX
@misc{prefill-awareness,
  title = {Prefill Awareness in Large Language Models},
  url = {https://www.aisi.gov.uk/research/prefill-awareness-in-large-language-models},
  year = {2026},
  date = {2026-06-10},
  version = {Source capture 2026-10-25},
  note = {Source capture 2026-10-25}
}
NUWA record · BibTeX
@misc{prefill-awareness-nuwa,
  title = {Prefill Awareness in Large Language Models — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/prefill-awareness/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--prefill-awareness.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--prefill-awareness.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

  1. [1] Prefill Awareness in Large Language Models. 2026-06-10 ; Official publication masthead and article byline; Abstract; §3.1 Definition and operationalization of prefill awareness
  2. [2] Read the full paper