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

Efficiently Aligning Draft Models via Parameter- and Data-Efficient Adaptation

NeurIPS 2026 · 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 EDA, which adapts speculative-decoding draft models to fine-tuned target models by pairing a frozen shared expert with a lightweight private expert, regenerating training data with the target, and selecting high-value samples via PCA/Mahalanobis scoring.

Problem

Speculative decoding draft models lose acceptance length when the target LLM is domain fine-tuned, and retraining a full draft model per target is costly.

Approach

Decouple the draft FFN into a frozen shared expert (from a general draft) and a trainable private expert combined by learned gating; regenerate training data with the fine-tuned target to match drafting objectives; and select training samples by a training-free representation-shift (PCA + Mahalanobis, high-quantile) score.

Key results

• On Qwen2.5-Math-7B math tasks at T=0, EDA (Ours) reaches average acceptance length 5.19 and 3.27x speedup vs. 4.69 for Full-FT.
• Overhead on math transfer: EDA trains 127 MB and 2.0 h vs. 462 MB and 5.1 h for re-training, with tau 4.79 vs. 4.22.
• Using 50% of the data selected by PCA-based score, EDA reaches tau 4.79 on GSM8K, close to full-data 4.82.

Surprising findings

• Full-FT and LoRA sometimes underperform the base-data EDA variant and in some medical MMLU cells the speedup is inconsistent with acceptance length gains.

Limitations

• Evaluation limited to Qwen2.5-7B-family target and draft pairs.
• Gains at temperature 1 are smaller and absolute acceptance lengths remain modest on some tasks.
• Relies on access to target model for self-generation, and on a general reference set for the Mahalanobis statistics.

Experimental setup

Datasets: ShareGPT DeepMath-103K Magicoder-Evol-Instruct-110K Medical dataset GSM8K AIME2024 SVAMP Hendrycks-MATH MathQA HumanEval APPS BigCodeBench

Benchmarks: GSM8K AIME2024 SVAMP Hendrycks-MATH MathQA HumanEval APPS BigCodeBench HumanEval+ MBPP MedMCQA MedQA-USMLE

Metrics: average acceptance length decoding speedup

Baselines: Training-Free draft Full-FT LoRA

Builds on: EAGLE LoRA Mahalanobis distance

Closest papers by meaning

• Learning To Draft: Adaptive Speculative Decoding with Reinforcement Learning (ICLR 2026 · similarity 0.92)
• PARD: Accelerating LLM Inference with Low‑Cost PARallel Draft Model Adaptation (ICLR 2026 · similarity 0.91)
• Steering Pretrained Drafters During Speculative Decoding (AAAI 2026 · similarity 0.91)
• Speculative Decoding with CTC-based Draft Model for LLM Inference Acceleration (NeurIPS 2024 · similarity 0.90)
• Bridging Draft Policy Misalignment: Group Tree Optimization for Speculative Decoding (ICLR 2026 · similarity 0.90)
• Not-a-Bandit: Provably No-Regret Drafter Selection in Speculative Decoding for LLMs (ICLR 2026 · similarity 0.90)