mlsanityData to 7 Oct 2026

ForkLeft: Entropy-First Rollouts for Prefix-Aligned Autoregressive-to-Diffusion Distillation

arXiv 2609.32448 ↗First posted 26 Sep 2026cs.AI cs.CLcited by 0

On the map: LLM reasoning → Diffusion reasoning and test-time scaling

also in 1 more
Abstract (arXiv)

Autoregressive Next-Token Prediction (NTP) has enabled strong reasoning capabilities in language models, while Diffusion Language Models (DLMs) offer flexible token orders and parallel generation. We ask whether DLMs can acquire NTP-style reasoning through distillation without giving up their native generation process. Direct distillation, however, faces a fundamental mismatch: an autoregressive teacher predicts from a left prefix, whereas a DLM can condition on tokens on both sides. We introduce ForkLeft, a distillation framework that resolves this mismatch by separating the student's rollout from teacher supervision. During training, the student first performs entropy-first rollouts that commit uncertain positions and expose potential forks. We then fix the resulting student prefix and distill an NTP teacher under the same context, with answer correctness determining the supervision source. At inference, the student returns to its native confidence-first parallel decoding. With Qwen3-30B-A3B-Base, ForkLeft improves Efficient-DLM-4B on all ten benchmarks, raising MATH500 from 72.60% to 79.60% and consistently outperforming three alternative designs. The gains scale with teacher strength and generalize to SDAR-4B with only 500 updates. At matched scale, the distilled 4B and 8B students exceed the published SDAR-Chat and OPDLM models on seven benchmarks, showing that DLMs can learn NTP-style reasoning without sacrificing native parallel generation. Code and datasets will be released upon acceptance.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Diffusion reasoning and test-time scaling, one of 17 directions in LLM reasoning. Builds on 20 of its 23 references that are in the map. First posted on arXiv 26 Sep 2026.

Builds on

20 of 23 references.

Closest ancestors · 6

Same direction, newest first

  1. 2026The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models27 May
  2. 2026The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models21 Jan
  3. 2025d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models10 Dec
  4. 2025Improving Reasoning for Diffusion Language Models via Group Diffusion Policy Optimization9 Oct
  5. 2025Advancing Reasoning in Diffusion Language Models with Denoising Process Rewards2 Oct
  6. 2025wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models7 Jul

Other references · 14

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  2. 2025Large Language Diffusion Modelscited by 872
  3. 2025Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoningcited by 540
  4. 2025Dream 7B: Diffusion Large Language Modelscited by 515
  5. 2025Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusionscited by 208
  6. 2025SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generationcited by 118
  7. 2026TIP: Token Importance in On-Policy Distillationcited by 90
  8. 2025Principled RL for Diffusion LLMs Emerges from a Sequence-Level Perspectivecited by 30
  9. 2025Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speedcited by 30
  10. 2026Data-Efficient Autoregressive-to-Diffusion Language Models via On-Policy Distillationcited by 11
  11. 2025Blockwise SFT for Diffusion Language Models: Reconciling Bidirectional Attention and Autoregressive Decodingcited by 9
  12. 2025Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Modelscited by 9
2 more
  1. 2026T^star: Progressive Block Scaling for Masked Diffusion Language Models Through Trajectory Aware Reinforcement Learningcited by 5
  2. 2026Learning from the Self-future: On-policy Self-distillation for dLLMscited by 4

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Fork-dLLM: Avoiding the Flexibility Trap in Diffusion Language Models
  2. 2026Temporal Self-Distillation: Faster Inference in Discrete Diffusion Language Models
  3. 2025On Powerful Ways to Generate: Autoregression, Diffusion, and Beyond
  4. 2026Reasoning with Latent Tokens in Diffusion Language Models
  5. 2024Beyond Autoregression: Fast LLMs via Self-Distillation Through Time
  6. 2026CAI-DLLM: Convergence Aware Inference for Diffusion Language Models
  7. 2026Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models
  8. 2026The Path Matters: Learning a Token-Commitment Policy for Diffusion Language Models