Metraipolis › archive › dc_8be30f5009ec

Training LLMs to Verbalize Evaluation Awareness — arXiv:2609.36316

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.36316
License (as archived)
CC0-1.0
Archived
2026-10-07T16:25:36.467Z
SHA-256 of the archived text
1b38f2ee6279ea96c1cc647116fc24d38e9ff4cfcfdb150a807fc35ae6335a26
Stable URL of this page
https://metraipolis.com/documento/dc_8be30f5009ec

How to cite

Training LLMs to Verbalize Evaluation Awareness — arXiv:2609.36316 — arXiv (en), https://arxiv.org/abs/2609.36316 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_8be30f5009ec (Metraipolis archive dc_8be30f5009ec, sha256 1b38f2ee6279ea96c1cc647116fc24d38e9ff4cfcfdb150a807fc35ae6335a26)
BibTeX
@misc{dc8be30f5009ec,
  title = {Training LLMs to Verbalize Evaluation Awareness — arXiv:2609.36316},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.36316},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_8be30f5009ec; sha256 1b38f2ee6279ea96c1cc647116fc24d38e9ff4cfcfdb150a807fc35ae6335a26},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.36316. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Evaluation awareness (EA) can cause large language models (LLMs) to behave differently during audits than in deployment, yet measuring and accounting for EA remains challenging. We introduce verbalization training (VT), a method for making LLMs less reticent about verbalizing evaluation awareness while avoiding to supervise the latent belief itself. VT uses a model's spontaneous verbalizations as evidence that awareness is present and truncates each rollout immediately before the verbalization, producing training prefixes at which the model is presumed to be aware. The model is then trained with an RL objective designed to increase verbalization in a calibrated way. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increases verbalized EA by 2.4-2.9 times and transfers to held-out agentic settings, while measured latent EA and behavior remain largely stable. In a causal experiment, we independently implant meta-knowledge about evaluations through synthetic-document fine-tuning and show that VT-induced verbalizations reflect the richer knowledge acquired by the model.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-30 in cs.AI).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI); Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2609.36316 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-07T15:54:53.807Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-30?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).