Research map / Continual learning / Continual fine-tuning of LLMs
Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis
Yujie Zheng, Zhuo Li, Shengtao Zhang, Jiaqian Wang, Junjie Sheng, Junchi Yan, Weinan Zhang, Ying Wen, Bo Tang, Muning Wen
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 value-driven Q-based retrieval instead of heuristic similarity retrieval (ablation) | EvoKernel (GPT-5.2) · Ascend C KernelBench L2 | correctness (%): 67% -> 77% | Improved |
| add cross-task shared memory over per-operator refinement (ablation) | EvoKernel (GPT-5.2) · Ascend C KernelBench L2 | accuracy (%): 3.0 -> 76.0 | Improved |
| add cross-task shared memory over per-operator refinement (ablation) | EvoKernel (GPT-5.2) · Ascend C KernelBench L1 | accuracy (%): +49 pp Acc | Improved |
| add L1 memory before L2 (curriculum transfer) versus L2 from scratch | EvoKernel (GPT-5.2) · Ascend C KernelBench L2 | correctness (%): 34 -> 64 | Improved |
| use L1 then L2 ordering instead of mixed L1+L2 stream | EvoKernel (GPT-5.2) · Ascend C KernelBench L2 | correctness (%): 53 -> 64 | Improved |
| add GPT-5.2 memory to DeepSeek-V3.2 | DeepSeek-V3.2 · Held-out 50 operators | correctness (%): 6% -> 58% | Improved |
| add GPT-5.2 memory to Qwen3-Coder-30B | Qwen3-Coder-30B · Held-out 50 operators | correctness (%): 4% -> 32% | Improved |
| use EvoKernel instead of Refinement on weaker backbone DeepSeek-V3.2 | DeepSeek-V3.2 · Ascend C KernelBench | compilation rate (%): 35.0 -> 29.0 | Hurt |
| use EvoKernel instead of Pass@k on weaker backbone Qwen3-Coder-30B | Qwen3-Coder-30B · Ascend C KernelBench | accuracy (%): 4.0 -> 5.5 | Improved |
| use EvoKernel instead of Codex agent | GPT-5.2 · Ascend C KernelBench | accuracy (%): 46.0 -> 83.0 | Improved |
| use stronger backbone GPT-5.2 versus weaker models for experience-driven method | EvoKernel · Ascend C KernelBench | accuracy (%): 4.0 -> 83.0 | Mixed |
| add iterative refinement after first correct draft | EvoKernel · Ascend C KernelBench (159 operators) | speedup over first draft: median 3.60x | Improved |
Closest papers by meaning
- DRTriton: Large-Scale Synthetic Data Driven Reinforcement Learning for Triton Kernel Generation (NeurIPS 2026 · similarity 0.86)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · similarity 0.86)
- EGG: An Expert-Guided Agent Framework for Kernel Generation (ICML 2026 · similarity 0.85)
- Procedural Memory Distillation: Online Reflection for Self-Improving Language Models (NeurIPS 2026 · similarity 0.85)
- Adaptive Kernel Design for Bayesian Optimization Is a Piece of CAKE with LLMs (NeurIPS 2025 · similarity 0.85)
- Compiler-R1: Towards Agentic Compiler Auto-tuning with Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
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.
