Learning When Diffusion Language Models Should Regenerate
DiVR raised average reasoning accuracy from 35.3% to 43.4% by choosing when to regenerate blocks of text, with the generator unchanged.
The verifier learns from the final correctness of the model’s own generation attempts, predicting whether a partial sequence will reach a correct answer. Sudoku experiments found that hidden states carry error signals token confidence can miss, and revising an isolated error often reproduces it when surrounding context stays fixed. Across five reasoning benchmarks with LLaDA-8B-Base, DiVR used 1.9–3.5 times the cost of greedy decoding while outperforming search baselines costing 3.8–7.4 times as much. The authors also report gains alongside supervised fine-tuning and reinforcement learning, and generalization to other generators, including Dream.
ArXiv · AI/CL/LG's note
The verifier learns from the final correctness of the model’s own generation attempts, predicting whether a partial sequence will reach a correct answer. Sudoku experiments found that hidden states carry error signals token confidence can miss, and revising an isolated error often reproduces it when surrounding context stays fixed. Across five reasoning benchmarks with LLaDA-8B-Base, DiVR used 1.9–3.5 times the cost of greedy decoding while outperforming search baselines costing 3.8–7.4 times as much. The authors also report gains alongside supervised fine-tuning and reinforcement learning, and generalization to other generators, including Dream.
ArXiv · AI/CL/LG's note
score 5