Reproduce “TESS 2: A Large-Scale Generalist Diffusion Language Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jaesung Tae, Hamish Ivison, Sachin Kumar, Arman Cohan
TESS 2 introduces a large-scale generalist diffusion language model adapted from an autoregressive base, achieving competitive performance with AR baselines on instruction-following tasks and demonstrating faster sampling for long sequences. The model integrates a reward-based classifier guidance mechanism, which improved AlpacaEval win rates by up to 3-4 percentage points at inference time.
Key insights
- Powerful autoregressive models can be effectively adapted into competitive generalist diffusion LMs through a specific recipe, surpassing training diffusion models from scratch.
- Diffusion LMs offer unique advantages, including scalable inference-time compute for quality control and modular alignment via reward guidance for steering generations.
- While TESS 2 performs comparably to AR models on general QA, achieving competitive performance on complex reasoning tasks often requires domain-specific finetuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.