🔁 Easy tokens and hard tokens get the exact same compute. That, this paper argues, is why diffusion language models still trail autoregressive ones. Title: ALoDLM: Adaptively Looped Diffusion Language Models URL: 📘 Overview It adds an inner loop to each denoising step so individual tokens can exit at different recurrent depths. Across eleven benchmarks it beats not just comparable diffusion models but the autoregressive baselines too. 🔥 The problem it solves Within a single denoising step, some masked positions are easy and others need real computation. Standard DLMs apply one fixed depth to every position, wasting compute on the easy ones and starving the hard ones. Confidence-based deferral only half-fixes this, since a deferred token's latent computation is thrown away and recomputed from its mask embedding next step. 🛠 Methodology The Transformer splits into a prelude, a recurrent core looped up to four times, and a coda. ・Tokens whose predictive entropy falls below a threshold commit, then feed back as discrete context ・Unresolved tokens keep their accumulated latent state and keep refining ・Halting probabilities accumulate until a threshold ends the inner loop Because each commit changes the context, summing over exit schedules exactly is intractable, so exit depths become latent variables trained via a variational bound with unbiased gradients. 📊 Results ・Average scores of 65.5 at 1.7B and 80.3 at 8B, past Qwen3's 63.8 and 78.5 ・On GSM8K, roughly 2.7x the throughput of vLLM-served Qwen3-8B at comparable or better accuracy ・Raising the exit threshold at inference lifts loops per token from 1.6 to 2.34 and average score from 77.9% to 79.1% ・With no difficulty labels at all, it learned to spend the deepest computation on numeric tokens #DiffusionModels# #LLM#