Highlights of Meituan's ACL 2024 Papers: Speculative Decoding, Graph‑Structured Decoding, DolphCoder, and Instruction Fine‑tuning

This article selects four papers from Meituan's technical team that were accepted at ACL 2024, covering training‑cost optimization, speculative decoding, code‑generation optimization, and instruction fine‑tuning.

ACL is the premier international conference on computational linguistics and natural language processing, organized by the Association for Computational Linguistics. It is a CCF‑A class venue and the most influential conference in the field.

Meituan will have a booth (No. 11) at the ACL 2024 venue. A live stream and paper discussion will be held on August 12 at 17:00. Attendees are welcome to make appointments.

01. Speculative Decoding via Early‑exiting for Faster LLM Inference with Thompson Sampling Control Mechanism (Long Paper) – PDF

The rapid growth of large language models (LLMs) raises inference cost. The authors propose Early‑Exiting Speculative Decoding (EESD), which adds an early‑exit after N layers to generate draft tokens, uses self‑distillation to improve them, and employs a Thompson‑sampling controller to decide the number of draft tokens per step. Experiments on 13B and 70B models show significant speed‑up without quality loss.

02. Graph‑Structured Speculative Decoding (Long Paper) – PDF

This work extends speculative decoding by generating multiple draft hypotheses and organizing them in a directed acyclic graph (DAG). The graph allows efficient prediction and merging of repeated token sequences, reducing the draft model’s computation. Applied to models up to 70B parameters, the method (GSD) improves generation speed by 1.73‑1.96× over standard speculative decoding.

03. DolphCoder: Echo‑Locating Code Large Language Models with Diverse and Multi‑Objective Instruction Tuning (Long Paper) – PDF

To boost code‑generation performance of large code LLMs, the paper introduces DolphCoder, a diversified instruction‑tuned model that incorporates self‑assessment objectives. It learns varied instruction targets and combines code‑evaluation goals, achieving superior results on HumanEval and MBPP benchmarks. The study shows that diverse responses and better evaluation of code solutions enhance both code creation and correctness.

04. Learning or Self‑aligning? Rethinking Instruction Fine‑tuning (Long Paper) – PDF

The authors propose a knowledge‑perturbation analysis framework to disentangle the effects of instruction fine‑tuning (IFT) on large models. Experiments reveal that IFT mainly aligns existing internal knowledge rather than injecting new knowledge, and that preserving internal knowledge consistency is crucial for successful fine‑tuning.

Live broadcast will walk through the ACL papers. Meituan’s booth (No. 11) and technical experts will be present for交流 and Q&A.