Metraipolis › archive › dc_ea3d89f4c90c
An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning — arXiv:2608.17804
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.17804
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-18T14:48:42.976Z
- SHA-256 of the archived text
a68f096dc697aa3302899fb577bc04fb83f33ba23dcaf80afd2fb0b3df49adec- Stable URL of this page
https://metraipolis.com/documento/dc_ea3d89f4c90c
How to cite
An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning — arXiv:2608.17804 — arXiv (en), https://arxiv.org/abs/2608.17804 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_ea3d89f4c90c (Metraipolis archive dc_ea3d89f4c90c, sha256 a68f096dc697aa3302899fb577bc04fb83f33ba23dcaf80afd2fb0b3df49adec)
BibTeX
@misc{dcea3d89f4c90c,
title = {An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning — arXiv:2608.17804},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.17804},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_ea3d89f4c90c; sha256 a68f096dc697aa3302899fb577bc04fb83f33ba23dcaf80afd2fb0b3df49adec},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.17804. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Practical LLM unlearning is usually evaluated through two objectives: suppress target-specific knowledge and preserve non-target utility. In generative QA, this leaves a third behavior underspecified: when a target-adjacent prompt admits a broader answer without target-specific leakage, the model should answer at that level rather than leak, evade, or refuse. We study this specification problem in a controlled LoRA-GRPO RWKU setting, comparing four reward designs that span lexical suppression, anti-refusal shaping, rubric-based broad answering, and an explicit refusal contrast, with and without SFT warm-up. The experiments show that optimization success is not equivalent to behavioral unlearning: RWKU forget scores, held-out completion audits, and training dynamics can point to different conclusions. We trace these disagreements to reward-hacking endpoints, policy-support limits in GRPO, benchmark probes that miss endpoint changes, and a rubric reward that selects broad-topic answering with low semantic leakage under held-out evaluation. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-19 in cs.LG). Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2608.17804 — 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 2026-10-18T14:47:59.414Z dalla pagina https://arxiv.org/catchup/cs.LG/2026-08-19?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).