Adaptive Inference Hyperparameter Optimization
A method that dynamically adjusts inference hyperparameters per prompt using a lightweight reward model to prevent performance drops caused by brittle static configurations.
Source paper
Generated by AI. The paper’s authors did not write these.
A method that dynamically adjusts inference hyperparameters per prompt using a lightweight reward model to prevent performance drops caused by brittle static configurations.
A method that aligns autoregressive language models to diffusion frameworks using logit distillation to preserve the teacher model's probabilistic priors and prevent severe performance regressions.
A system that directly optimizes diffusion generation models on human preference data by formulating the preference loss over noise prediction errors without requiring complex reinforcement learning setups.
This paper introduces dLLM, a unified, open-source framework designed to standardize the training, inference, and evaluation of Diffusion Language Models (DLMs). It provides modular components for reproducing existing models (like LLaDA and Dream), recipes for converting autoregressive models and BERT encoders into DLMs with minimal compute, and a comprehensive evaluation pipeline to benchmark performance and speed.
The framework is built around three core abstractions: a modular Trainer supporting Masked and Block Diffusion objectives via HuggingFace adapters; a Sampler interface that decouples inference logic (including Fast-dLLM acceleration) from model architecture; and an Evaluation harness aligned with official benchmarks. The authors demonstrate the framework by releasing checkpoints of BERT-based chat models and converted Qwen models trained via Supervised Fine-Tuning (SFT).