Diffusion language models (dLLMs), such as LLaDA and Dream, have become competitive with autoregressive (AR) LLMs in generation quality while supporting native parallel decoding. A standard acceleration strategy is block-wise decoding, where each forward pass predicts a block of length B and commits high-confidence tokens. Confidence-threshold commitment adapts which tokens are accepted, but a fixed B still constrains the prediction scope. Existing horizon policies use manually chosen sizes, volatility statistics, or candidate-boundary scores. We instead use the model's own per-step confidence: in-window confidence typically follows a context-dependent cliff, providing a direct signal for adapting the look-ahead horizon. In this work, we propose PACE-dLLM, which refits this parametric cliff periodically in closed form, sets the next horizon at a shared confidence anchor, and retains confidence-threshold token commitment. Under a saturated-yield abstraction with iid, policy-independent monotone profiles, the oracle horizon is the smallest attaining maximal per-pass yield and improves the asymptotic forward-pass rate when fixed blocks undershoot it with positive probability. On four reasoning and code benchmarks across two open-source dLLM backbones, PACE-dLLM achieves the best average accuracy among the evaluated decoders on both backbones, with average throughput speedups of 5.23x on LLaDA and 3.06x on Dream (up to 8.52x on LLaDA MATH) over the unaccelerated semi-AR baseline, advancing the quality-throughput Pareto frontier.
Diffusion language models predict many masked positions at each denoising step, but their efficiency and output quality depend on which predictions enter the evolving context. Confidence is informative but incomplete for this decision: at matched confidence, proposals with different identity and support histories exhibit different one-step revision rates. We propose Trajectory-Aware Commit Gating (TACG), a training-free decoder that separates the current token proposal from the decision to commit it. TACG scores temporal support for the current proposal by contrasting its logits with an exponential moving average of earlier logits, requires the proposal identity to persist through a History Gate, and promotes a capped number of additional positions per step. The gate decides which positions to commit and when; every written token remains the contemporaneous base proposal. On complete four-task LLaDA evaluations under the stated single-seed protocol, measured task-specific operating points reach 51.02\% mean accuracy with 67.82 mean model evaluations, compared with 48.00\% and 76.59 for Native Confidence. Under a matched per-event commit-count replay, history-guided selection exceeds confidence ranking by 8.3 accuracy points at identical per-sample NFE. History-reference interventions alter commitment timing while preserving the base-proposal write rule. Experiments on Dream and LLaDA2-Mini characterize quality--cost behavior across backbones and generation lengths. The code is publicly available at https://github.com/Clarence-CV/TACG-DLLM.
확산(diffusion) 및 흐름 기반(flow-based) 모델은 이미지와 동영상 등의 도메인에서 연속 데이터를 생성하는 사실상의 접근법이 되었다. 그 성공은 언어 모델링(language modeling)에 이를 적용하려는 관심을 점점 더 끌어들이고 있다. 이미지 도메인의 대응 모델과 달리, 오늘날 선도적인 확산 언어 모델(diffusion language models, DLMs)은 주로 이산 토큰(discrete tokens) 위에서 동작한다. 본 논문에서는 이산 도메인에 대한 최소한의 적응만으로 연속 DLM을 효과적으로 만들 수 있음을 보인다. 우리는 연속 시간 Flow Matching에 기반한 연속 임베딩 공간에서의 확산 모델 클래스인 Embedded Language Flows(ELF)를 제안한다. 기존 DLM과 달리 ELF는 마지막 시간 단계까지 주로 연속 임베딩 공간 내에 머무르며, 이 단계에서 공유 가중치(shared-weight) 네트워크를 사용하여 이산 토큰으로 매핑한다. 이 구성은 이미지 도메인 확산 모델에서 확립된 기법, 예를 들어 classifier-free guidance(CFG)를 적응시키기 쉽게 만든다. 실험 결과 ELF는 선도적인 이산 및 연속 DLM을 크게 능가하며, 더 적은 샘플링 단계로 더 나은 생성 품질을 달성한다. 이 결과는 ELF가 효과적인 연속 DLM을 향한 유망한 경로를 제공함을 시사한다.