DiVR
An independent verifier for diffusion language models that guides regeneration decisions by predicting the final correctness of a partially generated sequence. It is plug-and-play.
Timeline (last 90 days)
DiVR (Diffusion Verifier) is a standalone verifier used to predict whether a partially generated sequence will ultimately reach the correct answer.
DiVR's training data comes from the final correctness of the model's own trajectories, that is, letting the model generate a large number of trajectories and recording whether each trajectory ultimately succeeds, then using these results to train the verifier.
At inference time, DiVR decides when to regenerate a block and which candidate to keep.
As a plug-and-play external module, DiVR does not modify the generator itself, so it is compatible with different generators without retraining the generative model.
The experiments use LLaDA-8B-Base on five reasoning benchmarks, and DiVR raises the average accuracy of the fine-tuned generator from 35.3% to 43.4%.
DiVR costs 1.9-3.5 times that of greedy decoding.
Search-based baseline methods cost 3.8-7.4 times as much but still achieve lower accuracy than DiVR.
DiVR is complementary to post-training methods such as supervised fine-tuning (SFT) and reinforcement learning (RL).
DiVR can generalize to other generators, including Dream.
The core insight of DiVR is that error correction for diffusion language models should not rely solely on token-level confidence, but should be verified starting from sequence-level correctness.
All Facts (10)
DiVR (Diffusion Verifier) is a standalone verifier used to predict whether a partially generated sequence will ultimately reach the correct answer.
50%UnverifiedDiVR's training data comes from the final correctness of the model's own trajectories, that is, letting the model generate a large number of trajectories and recording whether each trajectory ultimately succeeds, then using these results to train the verifier.
50%UnverifiedAt inference time, DiVR decides when to regenerate a block and which candidate to keep.
50%UnverifiedAs a plug-and-play external module, DiVR does not modify the generator itself, so it is compatible with different generators without retraining the generative model.
50%UnverifiedThe experiments use LLaDA-8B-Base on five reasoning benchmarks, and DiVR raises the average accuracy of the fine-tuned generator from 35.3% to 43.4%.
50%UnverifiedDiVR costs 1.9-3.5 times that of greedy decoding.
50%UnverifiedSearch-based baseline methods cost 3.8-7.4 times as much but still achieve lower accuracy than DiVR.
50%UnverifiedDiVR is complementary to post-training methods such as supervised fine-tuning (SFT) and reinforcement learning (RL).
50%UnverifiedDiVR can generalize to other generators, including Dream.
50%UnverifiedThe core insight of DiVR is that error correction for diffusion language models should not rely solely on token-level confidence, but should be verified starting from sequence-level correctness.
50%