💬 NLP

dVoting: Fast Voting for dLLMs

This paper introduces dVoting, a training-free technique that leverages the parallel generation capabilities of Diffusion Large Language Models (dLLMs) to iteratively refine outputs through consistency-based token voting, significantly boosting reasoning performance across multiple benchmarks with minimal computational overhead.

Original authors: Sicheng Feng, Zigeng Chen, Xinyin Ma, Gongfan Fang, Xinchao Wang

Published 2026-02-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sicheng Feng, Zigeng Chen, Xinyin Ma, Gongfan Fang, Xinchao Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a tricky riddle, but instead of thinking about it alone, you gather a group of five friends to brainstorm.

The Old Way (Autoregressive Models):
Usually, AI models work like a single person writing a story one word at a time. If they make a mistake early on, they have to keep writing the whole story based on that mistake, or start over from scratch. It's slow and rigid.

The New Way (Diffusion Models / dLLMs):
The paper introduces a new type of AI called a Diffusion Large Language Model (dLLM). Think of this AI not as a writer, but as a sculptor.

It starts with a block of stone where the words are hidden (masked).
It can look at the whole block at once and guess what the words might be.
Crucially, it can erase and rewrite any part of the sculpture at any time. It doesn't have to write from left to right; it can fix the ending before it even finishes the beginning.

The Problem: The "Brute Force" Approach

To get the best answer, researchers tried a simple trick: Voting.
They asked the AI to generate the answer five times and then picked the answer that appeared most often (Majority Voting).

The Good: This works great! The AI gets smarter.
The Bad: It's incredibly expensive. Asking the AI to write the entire story five times is like hiring five different architects to build five full houses just to pick the best floor plan. It wastes a lot of time and money (computing power).

The Solution: DVOTING (The "Smart Editor")

The authors, Sicheng Feng and his team, realized something interesting while watching these five AI friends work: They agree on most of the story.

If you ask five people to write a story about "Janet's ducks," they will all agree on words like "Janet," "ducks," "lay," and "eggs." They only disagree on the tricky math parts or the final conclusion.

DVOTING is a method that stops the AI from wasting time rewriting the parts everyone already agrees on. Here is how it works, using a creative analogy:

The "Group Project" Analogy

Imagine a group of students working on a group essay.

Round 1: Everyone writes a draft.
The Check: The teacher (DVOTING) looks at all five drafts.
- Observation: "Hey, everyone wrote 'Janet's ducks lay 16 eggs' exactly the same way. That part is solid."
- Observation: "But on the math part, three students said '18 dollars,' one said '20,' and one said '15.' That part is shaky."
The Fix (Remasking): Instead of making the students rewrite the whole essay, the teacher says: "Keep the first sentence exactly as it is. Erase only the math part and rewrite just that."
Repeat: The students rewrite only the math part. The teacher checks again. If they still disagree, they rewrite just that part again. If they agree, the project is done.

Why is this a Big Deal?

Speed: By only rewriting the "uncertain" parts, the AI finishes the job much faster. It's like editing a document by only changing the typos instead of retyping the whole book.
Smarter: It actually gets better at reasoning because it focuses its energy on the hard parts where mistakes happen.
No Training Needed: The best part? You don't need to retrain the AI or teach it new skills. You just give it a new set of instructions on how to work (a "voting strategy"). It's like giving a smart employee a better workflow, not a new degree.

The Results

The paper tested this on hard math and science problems (like the ones you see in high school or college exams).

DVOTING improved the AI's accuracy significantly (by about 5% to 15% on difficult tests).
It did this while using much less computing power than the old "brute force" method. In fact, it was up to 22 times faster than previous methods that tried to do the same thing.

In a Nutshell

DVOTING is like a smart editor for AI. Instead of asking the AI to write the same story five times from scratch, it asks the AI to write it once, then only fix the parts where the AI is confused. It saves time, saves money, and gets the right answer more often. It's a simple, clever trick that makes these new "sculptor" AIs much more practical for real-world use.

1. Problem Statement

Diffusion Large Language Models (dLLMs) have emerged as a competitive alternative to autoregressive models, offering the unique ability to generate tokens at arbitrary positions in parallel. This capability theoretically enables parallel test-time scaling (allocating more computation during inference to improve reasoning) without the sequential bottlenecks of autoregressive models.

However, existing approaches to enhance reasoning in dLLMs face two main challenges:

Training-Heavy Methods: Current reasoning enhancements often rely on Reinforcement Learning (RL) or fine-tuning, which are computationally expensive and require in-domain training data.
Inefficient Test-Time Scaling: Existing test-time scaling strategies (e.g., standard majority voting, HEX, RFG) often incur substantial inference overhead. They typically generate full sequences multiple times and aggregate results, leading to significant redundancy. For instance, standard majority voting might generate 5 full sequences, whereas many tokens across these sequences are identical and could be reused.

The core problem is how to leverage the parallel decoding and remasking capabilities of dLLMs to achieve high-performance reasoning via test-time scaling while minimizing the computational cost and redundancy inherent in naive voting strategies.

2. Methodology: DVOTING

The authors propose DVOTING, a training-free, fast voting strategy that iteratively refines generations by focusing computation only on uncertain tokens.

Key Observations

The method is built upon two empirical observations derived from analyzing multiple sampling runs on the same prompt:

Observation A (Answer Consistency): Easier questions often yield consistent answers across samples early in the process. High-performing samples show high "voting consistency" (e.g., 4/5 or 5/5 agreement), while low-performing samples show low consistency.
Observation B (Token-Level Redundancy): Across multiple sampled sequences for the same input, a large fraction of token positions remain identical (consistent). The paper introduces a metric, Non-Unique Position Rate (NUPR@k), showing that ~50% of tokens are identical across at least 2 samples (NUPR@2), and ~20% across 3 samples. This indicates massive redundancy in full-sequence generation.

The DVOTING Algorithm

DVOTING leverages the remasking mechanism of dLLMs (the ability to mask and regenerate arbitrary tokens at arbitrary positions) to eliminate this redundancy. The process involves:

Initial Sampling: Generate a small batch of candidate sequences (up to a limit $n$ , e.g., 5).
Consistency Analysis: Analyze token-level agreement across the candidate samples.
- Preserved Tokens: Tokens that are consistent (identical) across samples or belong to samples with highly consistent predicted answers are preserved (kept as fixed).
- Remasked Tokens: Uncertain tokens (those with low agreement) are remasked (replaced with [MASK]).
Iterative Refinement: The model regenerates only the remasked tokens in parallel, conditioned on the preserved tokens. This reduces the search space and computational load compared to regenerating the full sequence.
Early Stopping: The process terminates early if the candidate answers converge to a consistent result or if the maximum sampling count is reached.
Final Aggregation: The final answer is determined via majority voting on the refined candidate answers.

Technical Details:

Parallel Decoding: Uses an entropy-threshold scheme (threshold $\alpha$ ) to commit tokens with low entropy immediately, further speeding up the process.
Training-Free: Requires no additional model training or fine-tuning; it operates purely on the inference pipeline.

3. Key Contributions

Novel Insight: Identified and quantified significant token-level redundancy in dLLM sampling, demonstrating that many tokens are repeatedly generated correctly across samples.
DVOTING Framework: Proposed the first efficient test-time scaling strategy for dLLMs that combines iterative remasking with voting. It shifts the paradigm from "generate full sequences and vote" to "preserve confident tokens and refine uncertain ones."
Performance-Efficiency Trade-off: Demonstrated that DVOTING achieves superior performance with significantly fewer inference steps compared to existing baselines (Majority Voting, HEX, RFG).
Generalization: Validated the method across different dLLM architectures (LLaDA, Dream) and model sizes, including RL-enhanced models, proving its robustness.

4. Experimental Results

The authors evaluated DVOTING on mathematical (GSM8K, MATH500), scientific (ARC-C), and general reasoning (MMLU, GPQA) benchmarks using LLaDA-8B-Instruct, Dream-7B, and LLaDA-1.5.

Key Findings:

Accuracy Gains:
- GSM8K: +6.22% to +7.66% improvement over the base model.
- MATH500: +4.40% to +7.20% improvement.
- ARC-C: +3.16% to +14.84% improvement.
- MMLU: +4.83% to +5.74% improvement.
Efficiency (Speedup):
- DVOTING achieves these gains with 1.1–4.4× speedup over standard Majority Voting.
- It is 5.5–22.1× faster than HEX (which uses diverse masking schedules).
- It is 1.0–1.8× faster than RFG (which requires an auxiliary model).
Comparison to RL: DVOTING (training-free) achieves performance comparable to or better than RL-enhanced models (like d1, wd1, IGPO) but without the training cost.
Pareto Frontier: DVOTING establishes a new performance-efficiency frontier, offering the highest "Benefits Per Cost" (BPC) metric among all tested methods.

5. Significance

Unlocking dLLM Potential: DVOTING demonstrates that the unique parallel decoding and remasking capabilities of dLLMs can be effectively harnessed for reasoning tasks, moving beyond simple text generation.
Cost-Effective Reasoning: By eliminating redundant token generation, DVOTING makes test-time scaling practical for real-world applications where inference costs are a bottleneck. It offers a viable alternative to expensive RL training for boosting reasoning capabilities.
Foundation for Future Work: This work establishes the first baseline for efficient test-time scaling in dLLMs, providing a framework for future research into adaptive inference strategies for diffusion-based language models.

In conclusion, DVOTING proves that simple, iterative refinement based on token consistency can significantly boost reasoning performance in diffusion language models while drastically reducing the computational overhead associated with traditional voting methods.