arXiv AI

Not All Preferences Deserve Gradients: Understanding Gradient Utility in Offline Reasoning Alignment

The paper argues that in offline preference optimization for reasoning models, applying gradients uniformly to all chosen–rejected pairs is inefficient and can be harmful. It introduces the concept of gradient utility, showing that a pair’s contribution depends on both informativeness and stability, and that high-gradient samples often lie in high‑curvature regions, causing noisy updates. To address this, the authors propose SAGE (Stability‑Aware Gradient Efficiency), which maintains difficulty‑stratified candidate pools and selects only high‑utility pairs for backpropagation, resulting in smoother optimization and better performance on mathematical reasoning benchmarks.

arXiv Machine Learning
5d ago

Gradients Know What Outcomes Don't: Unlocking Reinforcement Learning for LLM Reasoning with Gradient-Aligned Rewards

The paper introduces Gradient-Aligned Reward (GAR), a reinforcement learning technique that uses truncated backpropagation to generate a compact gradient vector for each rollout and compares it to an expert-anchor gradient via cosine similarity. This dense, reasoning-aware reward improves large language model chain-of-thought reasoning on math benchmarks and transfers to other tasks without domain‑specific data, while adding less than 9% computational overhead. GAR outperforms existing baselines such as GRPO on Qwen3-4B and Qwen3-8B models.

By Leqi Zheng, Jinbo Su, Fang Niu, Chaokun Wang, Weiping Wang, Jiajun Zhang, Shannan Yan, Jie Wu, Zhaolu Kang, Rong Fu, Hang Zhang
Hugging Face Trending Papers
Jun 3

BiasGRPO: Stabilizing Bias Mitigation in High-Variance Reward Landscapes via Group-Relative Policy Optimization

Mitigating social bias in Large Language Models (LLMs) presents a distinct alignment challenge: unlike verifiable tasks, bias lacks a single ground truth, creating a high-variance, subjective reward landscape. Previous preference-based fine-tuning methods have major trade-offs: Direct Preference Optimization (DPO) is limited by the lack of exploration inherent in offline training, while Proximal Policy Optimization (PPO) can lead to training instability due to potentially unreliable critic estimates.

arXiv Machine Learning
Jun 2

Drifting Preference Optimization for One-Step Generative Models

arXiv:2606. 02521v1 Announce Type: new Abstract: One-step text-to-image generators are attractive for deployment because they generate an image with a single forward pass, but preference finetuning them remains difficult: standard alignment methods often rely on policy likelihoods, denoising trajectories, differentiable reward gradients, or test-time optimization.

By Zhou Jiang, Yandong Wen, Zhen Liu