Research map / Large language models / Instruction and fine-tuning

RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models

ICML 2025 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RoSTE, a quantization-aware supervised fine-tuning method that combines straight-through-estimator QAT with an adaptive per-layer choice between Walsh-Hadamard and identity rotations to reduce activation outliers, yielding 4-bit W4A4KV4 fine-tuned LLMs.

Problem

Fine-tuning followed by post-training quantization of LLMs yields suboptimal low-bit (W4A4KV4) accuracy because outliers in weights and activations inflate quantization error, and fine-tuning and quantization are treated as separate steps.

Approach

Jointly perform quantization-aware SFT with STE while adaptively selecting, per layer and once per outer iteration, between a random Walsh-Hadamard rotation and identity based on a layer-wise weight/activation quantization error surrogate; a theoretical bound links prediction error to weight quantization error.

Key results

• On Exp.2, RoSTE at W4A4KV4 averages 31.69 on Llama 3.1 8B Tulu 3 fine-tuning, +2.56 over best baseline SpinQuant (29.13), but 10.47 below full-precision SFT (42.16).
• On Exp.1 with Pythia-6.9B W4A4KV4, RoSTE reaches 23.66 ROUGE average vs 20.65 for STE and 6.22 for SpinQuant, close to FP16 SFT at 24.72.
• Ablation on Pythia-1B W4A4KV4: adaptive rotation gives 23.07 ROUGE vs 22.37 with no rotation and 13.09 with complete rotation.

Measured effects reported in the paper

Surprising findings

• Applying Hadamard rotation to every layer (complete rotation) substantially hurts performance (13.09 vs 22.37 with no rotation on Pythia-1B), so adaptive per-layer selection is necessary.
• Rotation sometimes introduces new outliers, and the last layers of Pythia do not benefit from rotation while Llama generally does.

Limitations

• Rotation choice is restricted to identity or Walsh-Hadamard per layer with a layer-wise heuristic instead of solving the full bilevel problem; K=1 outer iteration used.
• Large gap remains to full-precision SFT on Llama 3.1 8B (10.47 average points) and on several tasks like MMLU-Pro and GSM8K.
• Theoretical analysis relies on a linearized quadratic model with interpolation and Gram matrix assumptions.

Experimental setup

Datasets: Reddit TL;DR summarization Tulu 3 SFT mixture TruthfulQA MMLU-Pro BigBenchHard AGIEval GSM8K MATH

Benchmarks: Reddit TL;DR test (ROUGE) TruthfulQA MMLU-Pro BigBenchHard AGIEval GSM8K MATH

Metrics: ROUGE-1 ROUGE-2 ROUGE-L ROUGE-LSum accuracy exact match

Baselines: RTN GPTQ QuaRot SpinQuant STE LLM-QAT QLoRA DuQuant

Builds on: QuaRot SpinQuant straight-through estimator GPTQ

Closest papers by meaning

• LoftQ: LoRA-Fine-Tuning-aware Quantization for Large Language Models (ICLR 2024 · similarity 0.90)
• OWQ: Outlier-Aware Weight Quantization for Efficient Fine-Tuning and Inference of Large Language Models (AAAI 2024 · similarity 0.89)
• QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.89)
• Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.89)
• Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 2025 · similarity 0.89)
• FisherSFT: Data-Efficient Supervised Fine-Tuning of Language Models Using Information Gain (ICML 2025 · similarity 0.89)