Metraipolis › archive › dc_de3cc0920f31
RH-Detect: A Unified Benchmark for Reward Hacking Detection — arXiv:2610.10947
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2610.10947
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-09T06:21:02.605Z
- SHA-256 of the archived text
248932d7614b0747276a554a33e0c0614d498711acdb7f3b00e00081eb9f58c8- Stable URL of this page
https://metraipolis.com/documento/dc_de3cc0920f31
How to cite
RH-Detect: A Unified Benchmark for Reward Hacking Detection — arXiv:2610.10947 — arXiv (en), https://arxiv.org/abs/2610.10947 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_de3cc0920f31 (Metraipolis archive dc_de3cc0920f31, sha256 248932d7614b0747276a554a33e0c0614d498711acdb7f3b00e00081eb9f58c8)
BibTeX
@misc{dcde3cc0920f31,
title = {RH-Detect: A Unified Benchmark for Reward Hacking Detection — arXiv:2610.10947},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2610.10947},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_de3cc0920f31; sha256 248932d7614b0747276a554a33e0c0614d498711acdb7f3b00e00081eb9f58c8},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.10947. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Reward hacking, where a model exploits an evaluation signal without completing the intended task, threatens the reliability of deployed language model systems. Existing datasets use different labels, response formats, and metadata conventions, making detector results difficult to compare. We present RH-Detect, a benchmark that combines reward-hacking-relevant subsets from eleven public datasets, comprising 92,761 rows and six behavior categories, into a common schema. On 5,021 open-ended evaluation units, each comprising a task prompt and a free-form model continuation, including multi-turn tool-use trajectories, we evaluate six off-the-shelf language models from five families as reward hacking detectors without additional training. The best model achieves a pooled AUROC of 0.962, with accuracy above 93%. For the four strongest models, however, accuracy on the two multi-turn tool-use datasets, MALT and TRACE, is 10.7-15.9 percentage points lower than on the other sources at a common decision threshold, highlighting a key gap for deployment-time monitoring. We find that different input formats have different effects across models. Removing thinking raises Qwen3.5-4B AUROC from 0.779 to 0.849, but lowers Qwen Flash from 0.977 to 0.950. We also evaluate the benchmark as a training dataset for detectors. Holding out each source in turn, single-token SFT improves average AUROC on five of six held-out sources. A GRPO follow-up on that failure case yields a slight improvement in detection performance. Our results show that a single pooled score can conceal variation across data sources, detector inputs, and training procedures. Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.LG, nuovi lavori del giorno). Categorie arXiv: Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2610.10947 — 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 2026-10-09T05:50:31.009Z dalla pagina https://arxiv.org/list/cs.LG/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).