{
  "schema_version": "1.1",
  "id": "archive:https://arxiv.org/abs/2610.35804v1",
  "slug": "2610-35804v1-1rhe9ju",
  "url": "https://feed7.dev/p/2610-35804v1-1rhe9ju",
  "title": "Learning When to Commit: Schedule Distillation for Fast Diffusion Language Models",
  "why_included": "dSpark distills token reveal schedules for diffusion language models, aiming to cut decoding rounds while preserving task performance. Its reported speedups may interest builders exploring parallel generation for code and reasoning.",
  "summary": "dSpark learns which token reveals make other positions easier to predict, then distills that schedule into diffusion language models. A **KL-divergence regularizer** is used to help preserve generation quality. On GSM8K, the paper reports **10.12×** speedup for LLaDA-8B-Instruct and **5.12×** for Dream-7B-Instruct over various baselines.",
  "practical_implication": "Builders evaluating diffusion models for code or reasoning generation can track whether schedule learning reduces latency without hurting task results.",
  "agent_context": "dSpark learns which token reveals make other positions easier to predict, then distills that schedule into diffusion language models. A **KL-divergence regularizer** is used to help preserve generation quality. On GSM8K, the paper reports **10.12×** speedup for LLaDA-8B-Instruct and **5.12×** for Dream-7B-Instruct over various baselines.\n\nBuilders evaluating diffusion models for code or reasoning generation can track whether schedule learning reduces latency without hurting task results.\n\nThe reported speedups are for GSM8K and use various baselines; the material does not give a direct comparison of latency or quality on code-generation tasks.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2610.35804v1",
    "published_at": "2026-10-26T17:35:58.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "model",
  "domains": [
    "coding",
    "research"
  ],
  "topics": [
    "reasoning"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The reported speedups are for GSM8K and use various baselines; the material does not give a direct comparison of latency or quality on code-generation tasks."
  ],
  "connected_context": {
    "meaning": "dSpark targets diffusion-model latency by learning token reveal order and distilling it, with a KL regularizer to help preserve generation quality. The reported GSM8K speedups motivate a code-task check, but do not establish that the gains transfer to coding or reasoning workloads beyond that benchmark.",
    "corpus_size": 928,
    "generated_at": "2026-10-27T11:06:55.362Z",
    "connections": [
      {
        "title": "Learning When to Think: Adaptive Reasoning for Test-Time Compute Allocation",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.20256v1",
        "feed7_url": "https://feed7.dev/p/2608-20256v1-15u528w",
        "reason": "Both seek lower inference cost while retaining task performance: adaptive reasoning reduces allocated output tokens, while dSpark changes the token reveal schedule. Their reported evidence covers different tasks and mechanisms."
      },
      {
        "title": "Prefix Sliding for efficient test-time scaling",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.26070v1",
        "feed7_url": "https://feed7.dev/p/2608-26070v1-1sn2iyc",
        "reason": "Prefix Sliding and dSpark offer distinct inference-time efficiency levers: one limits retained context, while the other changes generation order. The supplied summaries do not establish how their gains combine."
      },
      {
        "title": "Minimally Invasive Steering of Language Models",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2609.30218v1",
        "feed7_url": "https://feed7.dev/p/2609-30218v1-0xtyd0y",
        "reason": "Both include a KL-based constraint alongside an inference-time intervention, but MISVO steers outputs toward a reward whereas dSpark distills a reveal schedule; their aims and evidence are distinct."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-10-26T17:35:58.000Z",
  "modified_at": "2026-10-26T17:35:58.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2610-35804v1-1rhe9ju",
    "json": "https://feed7.dev/p/2610-35804v1-1rhe9ju.json",
    "markdown": "https://feed7.dev/p/2610-35804v1-1rhe9ju.md"
  }
}