Apple ML Research: how diffusion models learn to select tokens without manual heuristics

Apple ML Research: How Diffusion Models Learn to Select Tokens Without Manual Heuristics

Apple ML Research published in July 2026 a study on training token revelation policies for diffusion language models: instead of manual heuristics that require tuning and show unstable results, the proposal is to train the token selection process itself.

What Are Diffusion Language Models

Diffusion language models (dLLMs) differ fundamentally from familiar autoregressive architectures like GPT or Claude. An autoregressive model builds text sequentially — token by token from left to right. A diffusion model works differently: it starts with completely "masked" text and over several iterations gradually "reveals" positions. According to Apple ML Research, modern dLLMs on many practical tasks are already comparable in quality to autoregressive models — while promising advantages in inference speed. Since the model can reveal multiple tokens in a single step, theoretically higher throughput is achievable with lower computational costs.

Why Token Selection Decides Everything

At each step of diffusion, the model must answer the question: which of the remaining masked tokens should be revealed next? This seemingly technical question directly affects both the quality of generated text and the speed of operation. A naive approach — random token selection — turns out to be the worst strategy. Apple researchers found that the confidence thresholding heuristic gives significantly better results: the model prioritizes opening positions where its confidence (probability score) is highest.

Key facts about current heuristics:

• Confidence thresholding improves sample quality compared to random revelation
• The strategy increases token throughput — the number of tokens per unit time
• The disadvantage is the need for manual parameter tuning for each task
• Heuristic performance is unstable and varies depending on the application

What Apple Proposes Instead of Manual Rules

The central idea of the work is to replace fixed heuristics with a learnable revelation policy. Instead of a developer manually selecting threshold values and tuning parameters for each scenario, the model itself learns the optimal token selection strategy.

"Such heuristics have disadvantages: they require manual tuning, and their performance..." — from the annotation of the

Apple ML Research study.

The principle is not new to machine learning: replacing human heuristics with automatically optimized strategies is the standard path from "manual tuning" to "learning from data." Applied to diffusion models, such an approach potentially discovers token revelation strategies that a human developer would not have thought of in advance, and does not require recalibration when switching to new tasks.

What This Means

Diffusion language models attract attention as one of the promising alternative approaches to large language model architecture — especially from the perspective of inference efficiency. If token revelation policies can be trained automatically and consistently outperform manual heuristics, this will eliminate the key practical barrier to applying dLLMs in industrial conditions, where generation cost and speed are critical.

Frequently Asked Questions

How Do Diffusion Language Models Differ From GPT?

Autoregressive models like GPT generate text sequentially — one token at a time. Diffusion language models start with completely masked text and over several steps iteratively "reveal" it, allowing parallel processing of multiple positions and potentially accelerating inference.

What Is Confidence Thresholding in the Context of dLLMs?

This is a heuristic where at each diffusion step, the model first opens the position where its confidence is highest. Apple ML Research shows this improves both text quality and generation speed compared to random revelation — but requires manual parameter tuning.