{"id":"arxiv:2603.05454","title":"Beyond Scattered Acceptance: Fast and Coherent Inference for DLMs via Longest Stable Prefixes","date":"2026-03-05","section":"LLMs: architectures, context, training","quality":0.057813,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/","source_url":"https://arxiv.org/abs/2603.05454","text_available":true,"authors":["Pengxiang Li","Joey Tsai","Hongwei Xue","Kunyu Shi"],"outgoing":[{"source":"arxiv:2603.05454","target":"arxiv:2502.09992","evidence":"id","id":"arxiv:2502.09992","title":"Large Language Diffusion Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a4863ed6f639560bd2feb6c66c13d1e3ac4cf42f8919a65c9818625cc6b6cde9/"}],"incoming":[],"text_parts":[{"start":0,"end":15483,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html"},{"start":15483,"end":31463,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html"},{"start":31463,"end":36188,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-3.html"}],"contents":[{"title":"Beyond Scattered Acceptance : Fast and Coherent Inference for DLMs via Longest Stable Prefixes","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"Abstract","offset":98,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"1 Introduction","offset":1755,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"2 Related Work","offset":5988,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"2.1 Diffusion Large Language Model","offset":6007,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"2.2 Acceleration Methods for Diffusion Language Models","offset":7812,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"3 Method","offset":10049,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"3.1 Preliminary","offset":10667,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"Forward Corruption Process.","offset":10980,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"Reverse Generation Process.","offset":12154,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"The Inefficiency of Conventional Schedulers.","offset":13276,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"3.2 The Longest Stable Prefix (LSP) Scheduler","offset":14485,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"Single-Pass Prediction and Stability Assessment.","offset":15427,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-1.html#text"},{"title":"Distinction from Blockwise Autoregressive Decoding.","offset":16405,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Approximate KV Caching for Prefix States.","offset":17209,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Targeted Block Sizing via Adaptive Thresholding.","offset":17892,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Structural Coherence via Boundary Snapping and Monotone Progress.","offset":19654,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4 Experiments","offset":21549,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4.1 Experimental Setup","offset":21567,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Models and Benchmarks.","offset":21595,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Baseline and LSP Configuration.","offset":22714,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4.2 Main Results and Analysis","offset":23998,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Global Planning in Creative Writing.","offset":25374,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4.3 Ablation Studies and Analysis","offset":26011,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4.3.1 The Critical Role of Adaptive Sizing","offset":26491,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Motivation.","offset":26540,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Analysis.","offset":27096,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4.3.2 Enhancing Coherence with Structural Snapping","offset":28424,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Motivation.","offset":28481,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Analysis.","offset":28968,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Qualitative Evidence.","offset":30048,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4.3.3 Prefix-First vs. Scattered: The Power of Topology","offset":30687,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Motivation.","offset":30749,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"Analysis.","offset":31274,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-2.html#text"},{"title":"4.3.4 Quantifying Repair Costs via Token Flip Rate","offset":32805,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-3.html#text"},{"title":"5 Conclusion","offset":33707,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-3.html#text"},{"title":"Limitations","offset":35539,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/part-3.html#text"}],"text_code_points":36188,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/303994b6470f9a9063fca1f6980da95344903a409e61de83c6b35e99146b6329/source.md"}