Research map / Architectures and efficiency / RNNs and normalization
LightningRL: Breaking the Accuracy–Parallelism Trade-off of Block-wise dLLMs via Reinforcement Learning
Yanzhe Hu, Yijie Jin, Pengfei Liu, Kai Yu, Zhijie Deng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add token-level NLL loss on correct trajectories (ablation) | LightningRL · GSM8K | accuracy (%): 80.7 -> 90.3 | Improved |
| add per-reward decoupled normalization (ablation) | LightningRL · GSM8K | accuracy (%): 85.3 -> 90.3 | Improved |
| add TPF-aware filtering (dynamic sampling) (ablation) | LightningRL · GSM8K | accuracy (%): 87.2 -> 90.3 | Improved |
| add TPF-aware filtering (dynamic sampling) (ablation) | LightningRL · GSM8K | AUP: 454.5 -> 507.5 | Improved |
| use sequence-level policy/KL and token-level NLL reduction (Seq-Tok-Tok) instead of pure token-level (ablation) | LightningRL · GSM8K | accuracy (%): 80.0 -> 90.3 | Improved |
| use Seq-Seq-Seq reduction instead of Seq-Tok-Tok (ablation) | LightningRL · GSM8K | AUP: 507.5 -> 467.1 | Hurt |
| apply LightningRL RL post-training to SDAR-8B-b32 | SDAR-8B-b32 · average of four benchmarks | AUP: 211.2 -> 437.8 | Improved |
| apply LightningRL to SDAR-8B-b32 | SDAR-8B-b32 · MBPP | TPF: 2.44 -> 11.10 | Improved |
| apply LightningRL to SDAR-8B-b32 | SDAR-8B-b32 · HumanEval | accuracy (%): 73.5 -> 72.6 | Hurt |
| apply LightningRL to SDAR-8B-b32 | SDAR-8B-b32 · MATH500 | accuracy (%): 63.6 -> 63.0 | No clear effect |
| use LightningRL instead of TraceRL | SDAR-8B-b32 · MBPP | TPF: 2.50 -> 11.10 | Improved |
| add decoupled reward normalization (ablation) | LightningRL · GSM8K training | collapse ratio | Improved |
Closest papers by meaning
- AREAL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning (NeurIPS 2025 · similarity 0.89)
- TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts (NeurIPS 2026 · similarity 0.87)
- Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening (ICML 2026 · similarity 0.87)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
