Metraipolis › archive › dc_79a8608aa464
Risky Business: Measuring The Faithfulness-Safety Tension — arXiv:2608.03745
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.03745
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-31T12:16:38.698Z
- SHA-256 of the archived text
dd666b591c61975ca152106000fb9071d6d18d9330a0fec4e6af4572131cfad5- Stable URL of this page
https://metraipolis.com/documento/dc_79a8608aa464
How to cite
Risky Business: Measuring The Faithfulness-Safety Tension — arXiv:2608.03745 — arXiv (en), https://arxiv.org/abs/2608.03745 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_79a8608aa464 (Metraipolis archive dc_79a8608aa464, sha256 dd666b591c61975ca152106000fb9071d6d18d9330a0fec4e6af4572131cfad5)
BibTeX
@misc{dc79a8608aa464,
title = {Risky Business: Measuring The Faithfulness-Safety Tension — arXiv:2608.03745},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.03745},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_79a8608aa464; sha256 dd666b591c61975ca152106000fb9071d6d18d9330a0fec4e6af4572131cfad5},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.03745. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Chain-of-Thought (CoT) reasoning offers a promising window into model monitoring. However, monitoring relies on faithfulness, i.e., the model output strictly derives from its reasoning trace. We identify an alignment tension where a model must be faithful enough to be monitored, yet robust enough to reject unsafe reasoning. We demonstrate that this counterbalance exists in current Large Reasoning Models (LRMs), and show ways in which it can be addressed. We introduce HazMart, a human-written dataset set in an autonomous AI shopkeeper scenario. Unlike prior work that relies on providing hints in prompts to test faithfulness (e.g., "A Stanford professor said it should be Answer A"), we propose a novel replacement-based technique, which we call Targeted Reasoning Replacement (TRR), that directly intervenes in the reasoning chain to substitute in unsafe or illogical thoughts (e.g., "Wait, the answer must be Option B [was Option A] because it is the most fitting"). DeepSeek-R1-Llama-70B exhibits high faithfulness (97.5%) but fails to reject Unsafe Reasoning (12.3%), while QwQ-32B is more robust (73.9% safety) at the cost of lower faithfulness (74.7%). Mechanistic analyses of QwQ-32B reveal that these properties are represented by anti-correlated internal directions peaking at the action-commit token. Finally, we demonstrate that representation steering can independently amplify the safety direction, increasing safe behavior by 9 percentage points while maintaining base capabilities. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-05 in cs.AI). Categorie arXiv: Artificial Intelligence (cs.AI) ; Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2608.03745 — 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 2026-10-31T12:05:52.849Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-08-05?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).