Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | https://arxiv.org/abs/2510.23868 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866918500448075776 |
|---|---|
| author | Wang, Zhichao |
| author_facet | Wang, Zhichao |
| contents | This paper investigates whether reward matching is a viable alternative to reward maximization methods for on-policy RL of LLMs. Group-relative Implicit Fine-Tuning (GIFT) is proposed, combining GRPO-style group sampling, DPO-style implicit reward, and UNA-style MSE between implicit and explicit advantages. By applying z-score standardization, the intractable partition function $Z(x)$ in the DPO implicit reward is canceled, and the KL coefficient $β$ is eliminated from the RLHF and RLVR objective. The population minimizers of $\mathcal{L}_{\text{GIFT}}$ are characterized in closed form: they coincide exactly with the GRPO/RLHF solution family $π^{*}_β(y|x)\proptoπ_{\text{ref}}(y|x)e^{\frac{1}βr_ϕ(x,y)}$, with a prompt-dependent, variance-determined KL coefficient $β(x)=\frac{σ_ϕ(x)}{\hatσ_θ(x)}$. GIFT therefore solves the same parametric policy family as GRPO while replacing GRPO's externally tuned scalar $β$ with a prompt-adaptive $β(x)$ optimized endogenously by matching reward distributions. Empirically, on 7B-32B backbones, GIFT converges faster than GRPO, DAPO and GSPO and overfits less on RLVR (GSM8K, MATH, AIME) and produces higher length-controlled win rates on RLHF (AlpacaEval, Arena-Hard). All proofs and detailed background are deferred to the appendix. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_23868 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | GIFT: Group-Relative Implicit Fine-Tuning Integrates GRPO with DPO and UNA Wang, Zhichao Machine Learning Computation and Language This paper investigates whether reward matching is a viable alternative to reward maximization methods for on-policy RL of LLMs. Group-relative Implicit Fine-Tuning (GIFT) is proposed, combining GRPO-style group sampling, DPO-style implicit reward, and UNA-style MSE between implicit and explicit advantages. By applying z-score standardization, the intractable partition function $Z(x)$ in the DPO implicit reward is canceled, and the KL coefficient $β$ is eliminated from the RLHF and RLVR objective. The population minimizers of $\mathcal{L}_{\text{GIFT}}$ are characterized in closed form: they coincide exactly with the GRPO/RLHF solution family $π^{*}_β(y|x)\proptoπ_{\text{ref}}(y|x)e^{\frac{1}βr_ϕ(x,y)}$, with a prompt-dependent, variance-determined KL coefficient $β(x)=\frac{σ_ϕ(x)}{\hatσ_θ(x)}$. GIFT therefore solves the same parametric policy family as GRPO while replacing GRPO's externally tuned scalar $β$ with a prompt-adaptive $β(x)$ optimized endogenously by matching reward distributions. Empirically, on 7B-32B backbones, GIFT converges faster than GRPO, DAPO and GSPO and overfits less on RLVR (GSM8K, MATH, AIME) and produces higher length-controlled win rates on RLHF (AlpacaEval, Arena-Hard). All proofs and detailed background are deferred to the appendix. |
| title | GIFT: Group-Relative Implicit Fine-Tuning Integrates GRPO with DPO and UNA |
| topic | Machine Learning Computation and Language |
| url | https://arxiv.org/abs/2510.23868 |