Metraipolis › archive › dc_ee15cfa0d685
Direction-Aware Token Credit Assignment with Reward-Model Gradients — arXiv:2610.32772
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2610.32772
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-26T20:12:34.811Z
- SHA-256 of the archived text
49b9f7627a11fa0ed43cce14868cffc6fcd304269a4bf29ce65865f825ef08ce- Stable URL of this page
https://metraipolis.com/documento/dc_ee15cfa0d685
How to cite
Direction-Aware Token Credit Assignment with Reward-Model Gradients — arXiv:2610.32772 — arXiv (en), https://arxiv.org/abs/2610.32772 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_ee15cfa0d685 (Metraipolis archive dc_ee15cfa0d685, sha256 49b9f7627a11fa0ed43cce14868cffc6fcd304269a4bf29ce65865f825ef08ce)
BibTeX
@misc{dcee15cfa0d685,
title = {Direction-Aware Token Credit Assignment with Reward-Model Gradients — arXiv:2610.32772},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2610.32772},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_ee15cfa0d685; sha256 49b9f7627a11fa0ed43cce14868cffc6fcd304269a4bf29ce65865f825ef08ce},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.32772. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
GRPO shares a response-level advantage across tokens, leaving unresolved how strongly each token should contribute to learning. Updates induced by the shared advantage reinforce or suppress the sampled token at each position, but these updates do not always point toward increasing reward at the token level. This motivates a work-based view of credit: just as work depends on the component of force along a displacement, token credit should reflect the local reward change along the policy update direction. We introduce Direction-Aware Token Credit Assignment (DATC), which maps frozen reward-model gradients into policy logit space and allocates weights by evaluating the alignment between the reward direction and GRPO's signed sampled-action update direction, while preserving the sign and mean of the response-level advantage. DATC outperforms GRPO in all four settings across two model families. 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/2610.32772 — 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 2026-10-26T19:49:03.426Z dalla pagina https://arxiv.org/list/cs.LG/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).