Answer First, Reason Later: Commitment Order in Diffusion LLMs

2026-08-06

AI Summary

This work addresses a critical failure mode in diffusion language models during reasoning tasks, where unordered token commitments often lead to premature answer generation that bypasses intermediate reasoning steps. The study is the first to attribute this issue to insufficient reachability of valid token commitments rather than erroneous termination beliefs. To resolve this, the authors propose a frontier-gated commitment mechanism that enforces sequential token commitment while preserving the efficiency of parallel decoding. Through trajectory log analysis, a 2×2 prompt-decoding ablation study, and controlled windowing strategies, the method substantially improves performance—raising accuracy on GSM8K from 0.528 to 0.852—and achieves consistent gains on Dream-7B and MATH-500 benchmarks. These results demonstrate that chain-of-thought reasoning is effective only when token commitments occur in a strictly ordered manner.

Abstract

Masked diffusion language models (dLLMs) can commit tokens in any order -- a freedom marketed as their core advantage over autoregressive decoding. We show that on reasoning tasks this freedom is instead the axis of failure. Logging every commitment during decoding of LLaDA-8B on GSM8K, we find that unconstrained (pure) decoding commits the final answer at 15-24% of the trajectory while half the reasoning region is still masked, and collapses to answer-only outputs on up to 90% of problems as the canvas grows. The cause is not the model's termination beliefs -- EOS "pressure" is nearly identical across decoders -- but reachability: whether the sampler may act on those beliefs at distant positions. A 2x2 prompt-decoder design shows that chain-of-thought helps only under ordered commitment (interaction +34.8 percentage points, 95% CI [26.8, 42.8]; without reasoning text the decoders are indistinguishable), an interaction we decompose into a collapse channel and an order channel and replicate on Dream-7B and MATH-500. A single-knob intervention -- frontier-gated commitment -- causally recovers the full gap (0.528 to 0.852) while preserving up to 4x parallel decoding, along a measured frontier whose optimal window flips from w=1 at full refinement to unconstrained at 8 tokens/step. Our results reframe existing window-style samplers, previously motivated by efficiency, as the minimal fix for a reasoning pathology they were never designed to address.

Problem

• diffusion language models
• reasoning collapse
• commitment order
• masked generation
• chain-of-thought

Innovation

• masked diffusion language models
• commitment order
• frontier-gated commitment
• reasoning collapse
• parallel decoding