The LightReasoner paper from Hong Kong University shows that small language models can guide large models on critical reasoning steps, achieving up to 90% faster inference and significant accuracy gains across multiple math benchmarks.
This guide walks through setting up a Kunlun P800 XPU host, preparing Docker containers, deploying Qwen3‑8B/‑32B/‑VL models with vLLM‑Kunlun, benchmarking performance, and running full‑parameter DPO training using LLaMA‑Factory, providing scripts, configuration files, and troubleshooting tips for AI engineers.
This article critically examines R1‑Zero‑style training by analyzing foundation models and reinforcement learning, uncovering pre‑training and optimization biases, proposing an unbiased Dr. GRPO method, and demonstrating a minimalist 7B‑model recipe that achieves new state‑of‑the‑art performance on AIME 2024.
This article presents the PLUG 270‑billion‑parameter Chinese language model and introduces a parameter‑efficient sparsity training (PST) framework that combines unstructured and structured pruning with low‑rank decomposition to dramatically reduce model size while preserving downstream performance.
The article explains how overly long prompts dilute a large‑model’s attention, causing slower responses and contradictory outputs, and introduces a three‑step prompt‑decay audit—density measurement, slimming, and versioned output—that cuts response time from 1.8 s to 0.6 s, triples logical density, and reduces hallucinations by 60 %.
Microsoft’s research shows that a 4‑billion‑parameter small model, Terminus‑4B, can act as an execution sub‑agent for terminal tasks, trimming token consumption by about 30% while preserving performance on demanding SWE‑Bench benchmarks, demonstrating a practical alternative to costly large models.
DECS, a novel training framework introduced by researchers from Fudan, Shanghai Jiao Tong, and the Shanghai AI Lab, theoretically exposes the flaws of length‑penalty rewards and, through token‑level reward decoupling and dynamic batch scheduling, reduces inference token counts by over 50% while improving accuracy across multiple benchmarks.
EMNLP 2025 in Suzhou showcases Taobao's booth featuring four cutting‑edge AI papers that introduce a novel embedding compression framework, an automatic iterative refinement method for complex instruction generation, a knowledge infusion scaling law for large language models, and a video caption optimization approach for text‑to‑video generation.
This guide details how to replace native Transformers inference with the high‑performance vLLM engine, leveraging PagedAttention, continuous batching, tensor parallelism, and OpenAI‑compatible APIs to achieve 3‑4× higher throughput, lower latency, and scalable multi‑GPU deployments for production‑grade large language models.
The article analyzes Qwen3.6‑35B‑A3B’s MoE architecture, showing how its 30 B active parameters outperform larger dense models across programming, agent, and multimodal benchmarks, and examines the flagship Qwen3.6‑Max‑Preview’s substantial gains in world knowledge, instruction following, and third‑party rankings.
MiniCPM‑SALA introduces a hybrid sparse‑linear attention architecture that reduces quadratic compute and memory costs, achieves state‑of‑the‑art performance on long‑context benchmarks, and delivers up to 3.5× faster inference than full‑attention models on sequences up to 1 million tokens.
This talk presents Alibaba DAMO Academy’s recent work on compressing large pretrained language models, covering task‑adaptive AdaBERT, data‑augmented L2A, and meta‑knowledge distillation Meta‑KD, describing their motivations, architectures, NAS‑based search, loss designs, and experimental results across multiple NLP tasks.
Alibaba Cloud’s AI platform PAI recently saw two papers accepted at EMNLP2024—VideoCLIP‑XL, which enhances video‑text representation for long descriptions using a large video‑long‑description dataset and novel pre‑training tasks, and TAPIR, a curriculum‑planning framework that distills instruction‑following abilities of large language models—while also releasing associated models, datasets, and integration tools for users.
SpecExit combines early‑exit and speculative decoding to let large reasoning models detect when they have almost finished thinking, trimming redundant chain‑of‑thought steps, reducing over‑thinking by 72% and achieving up to 2.5× faster end‑to‑end inference without noticeable accuracy loss.
This article introduces four advanced model‑merging algorithms—Task Vector, SLERP, TIES, and DARE—explains their underlying principles, compares their strengths, and demonstrates a practical merge of Mistral‑7B, WizardMath‑7B and CodeLlama‑7B using the open‑source MergeKit toolkit.
Ant Group's Ling‑2.6‑1T, a 1‑trillion‑parameter LLM built for token efficiency and fast‑thinking, outperforms on elite reasoning and agentic benchmarks, offers easy local deployment via vLLM or SGLang, provides a quantized 3.6‑bit version, and includes practical usage tips for developers and knowledge workers.
TaH, a selective latent‑iteration method for small language models, identifies and avoids unnecessary token‑level loops, cutting about 93% of extra iterations while delivering a stable 3.0%‑6.8% accuracy gain across nine math, QA, and code benchmarks.
The article explains how On-Policy Distillation (OPD) combines on‑policy sampling with dense teacher feedback via reverse KL to address low signal density, distribution shift, and capability interference in large‑model post‑training, and compares implementations by Qwen3, GLM‑5, MiMo‑V2 and DeepSeek‑V4.
The article analyzes the poor generalization of supervised fine‑tuning (SFT) for large language models, reveals its gradient as a high‑variance inverse‑probability policy gradient, proposes a one‑line Dynamic Fine‑Tuning correction, and shows substantial gains on challenging math and offline RL benchmarks.
Nous Research introduces Token Superposition Training, which groups tokens into bags, averages their embeddings, and predicts token groups without altering model architecture or adding compute, achieving up to 2.5× faster pre‑training while maintaining standard inference.
