Research map / Architectures and efficiency / Pruning and efficient tuning
Arbitrary-Order Block SignSGD for Memory-Efficient LLM Fine-Tuning
Yijie Zhou, Shi Pu
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 depth-biased block selection (vs updating all blocks as in BAdam) | ABSignSGD · OpenMathInstruct-2 runtime | runtime (h): 3.32 -> 2.66 | Improved |
| replace Adam with SignSGD as the base optimizer under block updates (ablation) | block-coordinate optimizer · Qwen3-1.7B OpenMathInstruct-2 | training loss: BAdam impaired vs Adam; ABSignSGD fastest | Improved |
| use SignSGD instead of SGD in block-coordinate updates (ablation) | BlockSGD · Qwen3-1.7B OpenMathInstruct-2 | training loss: faster iteration-wise convergence than SGD | Improved |
| vary block selection rule (depth-biased, deep-to-shallow, uniform random) (ablation) | ABSignSGD · Qwen3-1.7B OpenMathInstruct-2 | training loss: similar convergence across DS, UR, DB | No clear effect |
| add majority-vote aggregation over 4 agents (ablation) | ABSignSGD · Qwen3-8B OpenMathInstruct-2 | training loss: tracks non-MV baseline closely | No clear effect |
| scale MV agents from 1 to 32 (ablation) | ABSignSGD-MV · Qwen3-8B OpenMathInstruct-2 | training loss: global batch size 4 to 128 | No clear effect |
| use ABSignSGD instead of best baseline (BAdam) for math fine-tuning | Qwen3-8B · math benchmarks average | accuracy: 70 -> 76 | Improved |
| use ABSignSGD instead of BAdam for instruction tuning | Qwen3-8B · MT-Bench average | score: 5.72 -> 6.18 | Improved |
| use ABSignSGD instead of Apollo for instruction tuning | Llama3-8B · MT-Bench average | score: 5.67 -> 5.99 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GaLore — GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- BAdam: A Memory Efficient Full Parameter Optimization Method for Large Language Models (NeurIPS 2024 · similarity 0.87)
- Sign-Based Optimizers Are Effective Under Heavy-Tailed Noise (NeurIPS 2026 · similarity 0.87)
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 2024 · similarity 0.87)
- S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity (NeurIPS 2024 · similarity 0.87)
- Gradient Weight-normalized Low-rank Projection for Efficient LLM Training (AAAI 2025 · similarity 0.86)
- ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization (NeurIPS 2024 · similarity 0.86)
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.
