Haochen You

Vector quantization (VQ) underpins modern generative and representation models by turning continuous latents into discrete tokens. Yet hard nearest-neighbor assignments are non-differentiable and are typically optimized with heuristic straight-through estimators, which couple the update step size to the quantization gap and train each code in isolation, leading to unstable gradients and severe codebook under-utilization at scale. In this paper, we introduce GRIT-VQ (Generalized Radius and Integrated Transform-Vector Quantization), a unified surrogate framework that keeps hard assignments in the forward pass while making VQ fully differentiable. GRIT-VQ replaces the straight-through estimator with a radius-based update that moves latents along the quantization direction with a controllable, geometry-aware step, and applies a data-agnostic integrated transform to the codebook so that all codes are updated through shared parameters instead of independently. Our theoretical analysis clarifies the fundamental optimization dynamics introduced by GRIT-VQ, establishing conditions for stable gradient flow, coordinated codebook evolution, and reliable avoidance of collapse across a broad family of quantizers. Across image reconstruction, image generation, and recommendation tokenization benchmarks, GRIT-VQ consistently improves reconstruction error, generative quality, and recommendation accuracy while substantially increasing codebook utilization compared to existing VQ variants.

With the rise of Large Language Models (LLMs), significant progress has been made in research related to code automation within the field of software engineering. Among these developments, the code generation task has attracted considerable attention from both academia and industry due to its broad applicability. However, most existing studies focus primarily on improving the correctness of the code generated by LLMs, while overlooking the efficiency attributes of the code. As a result, LLMs often produce correct code without considering human preferences for efficiency. Efficient code not only conserves resources but also enhances the user experience. For online optimization scenarios, we find that the fixed clipping mechanism of Group Relative Policy Optimization (GRPO) can lead to entropy collapse, causing premature convergence to suboptimal solutions. To address this, we propose DC-GRPO. The method employs a multi-level reward that accounts for compilation success, functional correctness, and execution efficiency to alleviate sparse-reward issues. In addition, a dynamic clipping mechanism based on a four-quadrant analysis of advantage and generation probability adapts the clipping bounds via a nonlinear mapping, encouraging low-probability/high-reward responses while suppressing high-probability/low-reward ones. In our experiments on a 1.5B-parameter code model, DC-GRPO achieves an eff@5 of 13.4%, an improvement of 1.2 percentage points over standard GRPO, and shows consistent gains over other baselines on code-oriented LLMs. Our anonymous codes are available at https://anonymous.4open.science/r/DC-GRPO-9E0D/.