📊 Results:
✅ MDPO matches previous SOTA with 60× fewer updates and improves over SOTA +9.6% on MATH-500, +54.2% on Countdown when trained with the same number of gradient updates.
✅ MDPO + RCR consistently outperforms all baselines
We see this as a step toward more sampling and data-efficient DLMs.
✅ MDPO matches previous SOTA with 60× fewer updates and improves over SOTA +9.6% on MATH-500, +54.2% on Countdown when trained with the same number of gradient updates.
✅ MDPO + RCR consistently outperforms all baselines
We see this as a step toward more sampling and data-efficient DLMs.
August 20, 2025 at 9:40 AM
📊 Results:
✅ MDPO matches previous SOTA with 60× fewer updates and improves over SOTA +9.6% on MATH-500, +54.2% on Countdown when trained with the same number of gradient updates.
✅ MDPO + RCR consistently outperforms all baselines
We see this as a step toward more sampling and data-efficient DLMs.
✅ MDPO matches previous SOTA with 60× fewer updates and improves over SOTA +9.6% on MATH-500, +54.2% on Countdown when trained with the same number of gradient updates.
✅ MDPO + RCR consistently outperforms all baselines
We see this as a step toward more sampling and data-efficient DLMs.
🔹 Running Confidence Remasking (RCR) – a training-free decoding strategy that allows revising low-confidence tokens flexibly.
August 20, 2025 at 9:40 AM
🔹 Running Confidence Remasking (RCR) – a training-free decoding strategy that allows revising low-confidence tokens flexibly.
🔹 Masked Diffusion Policy Optimization – the first policy gradient method that optimizes MDLMs as a sequential decision-making process. MDPO exploits the nice property that MDLMs yield text completions at every inference step and optimizes the model with with intermediate-step rewards.
August 20, 2025 at 9:40 AM
🔹 Masked Diffusion Policy Optimization – the first policy gradient method that optimizes MDLMs as a sequential decision-making process. MDPO exploits the nice property that MDLMs yield text completions at every inference step and optimizes the model with with intermediate-step rewards.
👉 Rigid Remasking: The remasking schedules used in inference usually 'freeze' tokens once predicted and not remasked immediately, making it impossible to revise low-confidence tokens predicted at early steps.
💡 To address these, we propose:
💡 To address these, we propose:
August 20, 2025 at 9:40 AM
👉 Rigid Remasking: The remasking schedules used in inference usually 'freeze' tokens once predicted and not remasked immediately, making it impossible to revise low-confidence tokens predicted at early steps.
💡 To address these, we propose:
💡 To address these, we propose:
👉 Training–Inference Divide: MDLMs are trained to predict all randomly masked tokens in a single pass, while at inference, MDLMs follow a model-dependent, confidence-guided unmasking schedule that progressively reveals structure of the generated sequence.
August 20, 2025 at 9:40 AM
👉 Training–Inference Divide: MDLMs are trained to predict all randomly masked tokens in a single pass, while at inference, MDLMs follow a model-dependent, confidence-guided unmasking schedule that progressively reveals structure of the generated sequence.
Masked diffusion language models (MDLMs) are rising as powerful alternatives to autoregressive LMs, yet face two fundamental but overlooked problems:
August 20, 2025 at 9:40 AM
Masked diffusion language models (MDLMs) are rising as powerful alternatives to autoregressive LMs, yet face two fundamental but overlooked problems: