What’s New in Large Model Research? Top Meituan AI Papers Up to Oct 2025

This article curates the latest technical achievements and academic papers from Meituan’s technology team in the large‑model domain (up to October 2025), covering large language models, system and architecture optimization, multimodal understanding and generation, evaluation, and intelligent interaction.

Selected Meituan Papers

LongCat‑Flash‑Chat. arXiv:2509.01322. PDF, GitHub, HuggingFace.

LongCat‑Flash‑Thinking. arXiv:2509.18883. PDF, GitHub, HuggingFace.

Revisiting Scaling Laws for Language Models: The Role of Data Quality and Training Strategies. ACL 2025. PDF.

NeedleInATable: Exploring Long‑Context Capability of Large Language Models towards Long‑Structured Tables. NeurIPS 2025. PDF.

AgentRefine: Enhancing Agent Generalization through Refinement Tuning. ICLR 2025. PDF.

Earlier Tokens Contribute More: Learning Direct Preference Optimization from Temporal Decay Perspective. ICLR 2025. PDF.

TODO: Enhancing LLM Alignment with Ternary Preferences. ICLR 2025. PDF.

SEAS: Self‑Evolving Adversarial Safety Optimization for Large Language Models. AAAI 2025. PDF.

CogAtom: From Cognitive Atoms to Olympiad‑level Mathematical Reasoning in Large Language Models. EMNLP 2025. PDF.

Mitigating Tail Narrowing in LLM Self‑Improvement via Socratic‑Guided Sampling. NAACL 2025. PDF.

SCoder: Progressive Self‑Distillation for Bootstrapping Small‑Scale Data Synthesizers to Empower Code LLMs. EMNLP 2025 Findings. PDF.

FIRE: Flexible Integration of Data Quality Ratings for Effective Pretraining. EMNLP 2025. PDF.

Predictor‑Corrector Enhanced Transformers with Exponential Moving Average Coefficient Learning. NeurIPS 2024. PDF.

Learning or Self‑aligning? Rethinking Instruction Fine‑tuning. ACL 2024. PDF.

DolphCoder: Echo‑Locating Code Large Language Models with Diverse and Multi‑Objective Instruction Tuning. ACL 2024. PDF.

Mixture‑of‑Experts and Quantization

Unveiling Super Experts in Mixture‑of‑Experts Large Language Models. arXiv:2507.23279. PDF.

EPS‑MoE: Expert Pipeline Scheduler for Cost‑Efficient MoE Inference. arXiv:2410.12247. PDF.

FPTQ: Fine‑grained Post‑Training Quantization for Large Language Models. arXiv:2308.15987. PDF.

A Speed Odyssey for Deployable Quantization of LLMs. arXiv:2311.09550. PDF.

Flash Communication: Reducing Tensor Parallelization Bottleneck for Fast Large Language Model Inference. arXiv:2412.04964. PDF.

Speculative Decoding via Early‑exiting for Faster LLM Inference with Thompson Sampling Control Mechanism. ACL 2024. PDF.

Multimodal and Vision‑Language Research

Let Them Talk: Audio‑Driven Multi‑Person Conversational Video Generation. NeurIPS 2025. PDF.

InfiniteTalk: Audio‑driven Video Generation for Sparse‑Frame Video Dubbing. arXiv:2508.14033. PDF, HuggingFace.

HyperSeg: Towards Universal Visual Segmentation with Large Language Model. CVPR 2025. PDF.

A Token‑level Text Image Foundation Model for Document Understanding. ICCV 2025. PDF.

Efficient Self‑Supervised Video Hashing with Selective State Spaces. AAAI 2025. PDF.

Embracing Collaboration Over Competition: Condensing Multiple Prompts for Visual In‑Context Learning. CVPR 2025. PDF.

Denoising with a Joint‑Embedding Predictive Architecture. ICLR 2025. PDF.

Enhancing Multilingual Speech Recognition Through Language Prompt Tuning and Frame‑level Language Adapter. ICASSP 2024. PDF.

Lumen: Unleashing Versatile Vision‑centric Capabilities of Large Multimodal Models. NeurIPS 2024. PDF.

MobileVLM V2: Faster and Stronger Baseline for Vision Language Model. arXiv:2402.03766v. PDF.

UniViTAR: Unified Vision Transformer with Native Resolution. NeurIPS 2025. PDF.

CLIP‑IN: Enhancing Fine‑Grained Visual Understanding in CLIP via Instruction Editing Data and Long Captions. NeurIPS 2025. PDF.

Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy. NeurIPS 2025. PDF.

RoboTron‑Mani: All‑in‑One Multimodal Large Model for Robotic Manipulation. ICCV 2025. PDF.

RoboTron‑Drive: All‑in‑One Large Multimodal Model for Autonomous Driving. ICCV 2025. PDF.

RoboTron‑Nav: A Unified Framework for Embodied Navigation Integrating Perception, Planning, and Prediction. ICCV 2025. PDF.

Benchmarks and Evaluation

Q‑Eval‑100K: Evaluating Visual Quality and Alignment Level for Text‑to‑Vision Content. CVPR 2025. PDF, HuggingFace.

OIBench: Benchmarking Strong Reasoning Models with Olympiad in Informatics. arXiv:2506.10481. PDF, HuggingFace.

Ask, Fail, Repeat: Meeseeks, an Iterative Feedback Benchmark for LLMs' Multi‑turn Instruction‑following Ability. arXiv:2504.21625. PDF, HuggingFace.

CoreCodeBench: A Configurable Multi‑Scenario Repository‑Level Benchmark. arXiv:2507.05281. PDF, HuggingFace.

Leveraging Dual Process Theory in Language Agent Framework for Real‑time Simultaneous Human‑AI Collaboration. ACL 2025. PDF.

Hallu‑PI: Evaluating Hallucination in Multi‑modal Large Language Models within Perturbed Inputs. ACM MM 2024. PDF.

A Wolf in Sheep's Clothing: Generalized Nested Jailbreak Prompts can Fool Large Language Models Easily. NAACL 2024. PDF.

Reinforcement Learning and Optimization for LLMs

Enhancing Efficiency and Exploration in Reinforcement Learning for LLMs. EMNLP 2025. PDF.

SDGO: Self‑Discrimination‑Guided Optimization for Consistent Safety in Large Language Models. EMNLP 2025. PDF.

When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning. EMNLP 2025. PDF.

DenoiseRotator: Enhance Pruning Robustness for LLMs via Importance Concentration. NeurIPS 2025. PDF.

AMoPO: Adaptive Multi‑objective Preference Optimization without Reward Models and Reference Models. ACL 2025. PDF.

Don’t Half‑listen: Capturing Key‑part Information in Continual Instruction Tuning. ACL 2025. PDF.

PIPER: Benchmarking and Prompting Event Reasoning Boundary of LLMs via Debiasing‑Distillation Enhanced Tuning. ACL 2025. PDF.

CLAQ: Pushing the Limits of Low‑Bit Post‑Training Quantization for LLMs. PDF.

A Reasoner for Real‑World Event Detection: Scaling Reinforcement Learning via Adaptive Perplexity‑Aware Sampling Strategy. EMNLP 2025. PDF.