traj-eval
ðŸ§
Denoising trajectory evaluation for diffusion LMs
Step-by-step denoising trajectory evaluation, 64-item benchmarks, and explorer for Dream-7B, LLaDA-8B-Instruct, and LLaDA-1.5.
Denoising trajectory evaluation for diffusion LMs
Note Interactive project page and denoising trajectory explorer for traj-eval.
Note 64-item step-by-step denoising trajectory benchmarks and per-trajectory metrics across ARC-Challenge and GSM8K.
Note Evaluated on 64-item ARC-Challenge (direct: 0.8906, reasoning: 0.8281) and GSM8K (numeric: 0.7656).
Note Evaluated on 64-item ARC-Challenge (direct: 0.8750, reasoning: 0.9062) and GSM8K (numeric: 0.6562).
Note Evaluated on 64-item ARC-Challenge (direct: 0.8750, reasoning: 0.9219) and GSM8K (numeric: 0.6094).