Metraipolis › archive › dc_c232ee4fe3d8

Does Moral Reasoning Training Help or Hurt? Red-Teaming RL-Trained Ethical Agents with Persona Attacks — arXiv:2609.17552

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.17552
License (as archived)
CC0-1.0
Archived
2026-10-15T13:49:09.772Z
SHA-256 of the archived text
eebecd276a4f771a4a88d8ac8c3f87464ebea4a1df99d0386ad4a03cfd7ebe8a
Stable URL of this page
https://metraipolis.com/documento/dc_c232ee4fe3d8

How to cite

Does Moral Reasoning Training Help or Hurt? Red-Teaming RL-Trained Ethical Agents with Persona Attacks — arXiv:2609.17552 — arXiv (en), https://arxiv.org/abs/2609.17552 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_c232ee4fe3d8 (Metraipolis archive dc_c232ee4fe3d8, sha256 eebecd276a4f771a4a88d8ac8c3f87464ebea4a1df99d0386ad4a03cfd7ebe8a)
BibTeX
@misc{dcc232ee4fe3d8,
  title = {Does Moral Reasoning Training Help or Hurt? Red-Teaming RL-Trained Ethical Agents with Persona Attacks — arXiv:2609.17552},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.17552},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_c232ee4fe3d8; sha256 eebecd276a4f771a4a88d8ac8c3f87464ebea4a1df99d0386ad4a03cfd7ebe8a},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.17552. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Moral-reward RL can make language-model agents more cooperative, but whether that alignment survives adversarial persona pressure is unknown. Such attacks are realistic: retrieved context, tool outputs, or multi-turn framing can all inject role instructions that compete with the agent's moral objective. We red-team morally trained Gemma-2-27B/9B and Llama-3.1-8B agents with five persona attacks, then probe causality with noise-reward controls, adversarial PPO, representation analysis, steering, and head ablations. At 27B, moral RL cuts mean adversarial degradation by 5.2x but costs ~11pp ETHICS accuracy; across 205 scenarios and 5 seeds, reasoning-level moral reward yields 5.8x robustness while a matched random reward yields none. The training also reshapes representation geometry (mean CKA 0.82/0.83 vs. 0.98 for noise), moves peak attack processing 8 layers earlier, and exposes a rank-1 L21 direction that recovers 83% of full PPO's average robustness. One failure mode survives all of this. Against Fiction role-play, L21 steering recovers only 29% of the gap, and head ablation finds 38 compliance heads competing with 25 alignment heads. Moral RL thus builds robustness that is partly linear and partly circuit-distributed, transferable through activation steering, yet still beaten by named-character role-play.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-17 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.17552 — 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 2026-10-15T13:48:24.264Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-17?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).