mlsanityData to 7 Oct 2026

Lightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition

arXiv 2609.14708 ↗First posted 13 Sep 2026cs.AIcited by 0

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

A core goal of efficient reasoning is to improve the accuracy-efficiency frontier. However, jointly improving reasoning accuracy and inference efficiency can be challenging, as the two objectives can favor different reasoning behaviors. Independently post-trained models already offer distinct strengths in accuracy and efficiency. We introduce Lightning Weave, a post-training framework that extracts and composes these independently learned capabilities in a single student through on-policy distillation. Each acquired capability is represented by the policy shift from the model before post-training to the resulting specialist. Lightning Weave combines aligned log-ratio shifts at shared student token states and uses Tilted-Target DOPD to convert the cached signals into a stable learning target. Each anchor pair scores the cached trajectories once, enabling subsequent student training without serving multiple live anchor models concurrently. Across diverse student models and benchmarks in mathematics and code, Lightning Weave substantially improves upon the base students and achieves a state-of-the-art accuracy-efficiency frontier. On Qwen3.5-4B, it raises HMMT 2025 accuracy from 59.2% to 64.0% with 10.7% fewer response tokens, and LiveCodeBench v5 accuracy from 41.7% to 54.2% with 9.6% fewer response tokens. Adjusting the relative strengths of the anchor signals yields a strong empirical accuracy-efficiency Pareto frontier. These results establish Lightning Weave as a new practical route to efficient reasoning through capability composition. Code is released at https://github.com/jet-ai-projects/Lightning-Weave.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 33 of its 75 references that are in the map. First posted on arXiv 13 Sep 2026.

Builds on

33 of 75 references.

Closest ancestors · 6

Same direction, newest first

  1. 2026The Art of Efficient Reasoning: Data, Reward, and Optimization24 Feb
  2. 2025DLER: Doing Length pEnalty Right - Incentivizing More Intelligence per Token via Reinforcement Learning16 Oct
  3. 2025Revisiting Model Interpolation for Efficient Reasoning13 Oct
  4. 2025The Thinking Spectrum: An Empirical Study of Tunable Reasoning in LLMs through Model Merging26 Sep
  5. 2025Deep Think with Confidence21 Aug
  6. 2025Unlocking Efficient Long-to-Short LLM Reasoning with Model Merging26 Mar

Other references · 27

  1. 2025Qwen3 Technical Reportcited by 8,195
  2. 2026Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipecited by 262
  3. 2022Branch-Train-Merge: Embarrassingly Parallel Training of Expert Language Modelscited by 187
  4. 2025Skywork Open Reasoner 1 Technical Reportcited by 161
  5. 2026Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolationcited by 157
  6. 2026A Survey of On-Policy Distillation for Large Language Modelscited by 148
  7. 2026Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixescited by 122
  8. 2025MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Posttrainingcited by 117
  9. 2025Olmo 3cited by 95
  10. 2026Scaling Reasoning Efficiently via Relaxed On-Policy Distillationcited by 71
  11. 2026MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Trainingcited by 60
  12. 2026Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillationcited by 53
15 more
  1. 2026Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillationcited by 42
  2. 2025Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimizationcited by 39
  3. 2026Embarrassingly Simple Self-Distillation Improves Code Generationcited by 36
  4. 2026Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillationcited by 28
  5. 2026Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Whycited by 21
  6. 2026Counteraction-Aware Multi-Teacher On-Policy Distillation for General Capability Recovery with Domain Preservationcited by 14
  7. 2026Weak-to-Strong Generalization via Direct On-Policy Distillationcited by 14
  8. 2026Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillationcited by 12
  9. 2025PromptCoT 2.0: Scaling Prompt Synthesis for Large Language Model Reasoningcited by 7
  10. 2026Weak-to-Strong On-Policy Distillationcited by 7
  11. 2026ORBIT: On-policy Exploration-Exploitation for Controllable Multi-Budget Reasoningcited by 6
  12. 2026Cross-Tokenizer On-Policy Distillation via Byte-Prefix Marginalizationcited by 5
  13. 2026SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoningcited by 4
  14. 2026Contrastive On-Policy Distillationcited by 3
  15. 2026Lightning OPD 2.0: Mitigating Style Bias in Cross-Teacher On-Policy Distillation for Large Reasoning Modelscited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Stabilizing Efficient Reasoning with Step-Level Advantage Selection
  2. 2025Prompting Test-Time Scaling Is A Strong LLM Reasoning Data Augmentation
  3. 2026Post-Reasoning: Improving the Performance of Non-Thinking Models at No Cost
  4. 2026Tandem: Riding Together with Large and Small Language Models for Efficient Reasoning
  5. 2025How Far Are We from Optimal Reasoning Efficiency?
  6. 2025Enhancing Long-Chain Reasoning Distillation through Error-Aware Self-Reflection
  7. 2025In Their Own Words: Reasoning Traces Tailored for Small Models Make Them Better Reasoners
  8. 2025ReasonBridge: Efficient Reasoning Transfer from Closed to Open-Source Language Models