The report summarizes NeurIPS 2025 in San Diego, highlighting four NIRC papers on noise‑robust 3D human pose estimation, LVLM video‑anomaly understanding, and hand‑object reconstruction, and discusses broader industry trends such as feed‑forward generation and large‑scale pre‑training showcased by leading AI companies.
A recent study reveals that the widely used KL regularization in LLM reinforcement learning (RLVR) is mathematically biased, leading to unstable training and poorer generalization, and shows that moving the KL term back to the reward with a simple K1 estimator can boost out‑of‑domain performance by up to 20%.
Recursive Language Models (RLM) offer a cost‑effective alternative to expanding LLM context windows by storing prompts as variables and enabling recursive calls, allowing models to process over 100,000 tokens, with experiments showing superior performance and lower median costs compared to baseline approaches.
Tencent’s HyperAI team introduces WorldPlay, an open‑source real‑time interactive world model that achieves 24 FPS 720p video generation while preserving long‑term geometric consistency through dual‑action representation, dynamic context memory reconstruction, and a novel context‑forcing distillation, and also showcases Maya1 emotional TTS and RFdiffusion3 protein design models.
This paper introduces SeDi, a semantics‑ and distribution‑aware cross‑tokenizer knowledge distillation framework that aligns teacher and student token spaces via bipartite graph components and top‑K re‑encoding, achieving state‑of‑the‑art performance and lower exposure bias on multiple LLM benchmarks.
AI researchSemantic Alignmentcross-tokenizer distillation
This article introduces Bottom‑up Policy Optimization (BuPO), a novel reinforcement‑learning framework that treats large language models as collections of internal layer and modular policies, revealing distinct inference entropy patterns in Llama and Qwen‑3 and demonstrating superior performance on complex mathematical reasoning benchmarks.
Google’s 2025 research recap highlights eight major breakthroughs—from the Gemini 3 series achieving unprecedented multimodal reasoning and efficiency, to AI‑driven advances in scientific discovery, creative generation, quantum computing, climate resilience, and responsible AI safety—showcasing how intelligent agents are reshaping products, research, and global challenges.
This week’s AI paper roundup reviews six recent research works—including RecGPT‑V2, Nemotron 3 Nano, FrontierScience benchmark, AutoGLM, Deeper‑GXX, and QwenLong‑L1.5—highlighting advances in large‑language‑model‑driven recommendation, Mixture‑of‑Experts models, expert‑level scientific reasoning, GUI‑based foundation agents, graph neural network deepening, and ultra‑long‑context inference.
UnityVideo, a new vision framework from HKUST, CUHK, Tsinghua and Kuaishou, unifies training across depth, flow, pose, segmentation and RGB modalities, achieving faster convergence, higher video quality, zero‑shot generalization and stronger physical reasoning compared with existing single‑modality video generators.
This article surveys the latest research on Unified Multimodal Foundations (UFM), explaining why integrating understanding and generation across text, image, video, and audio is essential for AGI, and detailing modeling paradigms, encoding/decoding strategies, training pipelines, benchmarks, and real‑world applications.
This article walks through the open‑source DeerFlow framework—its multi‑agent architecture, core features, and a step‑by‑step implementation using the Go‑based LangGraphGo library, covering planner, researcher, reporter and podcast nodes, state‑graph design, CLI/web modes, and deployment instructions.
The article examines the chaotic rise of “world models” in AI, tracing their origins from early mental‑model theory to modern representation‑ and generation‑based approaches, and argues that the current hype reflects a broader shift away from large language models toward embodied, physics‑grounded intelligence.
FFGo reveals that the first frame of text‑to‑video models acts as a conceptual memory buffer storing visual entities, and by using a few‑shot LoRA trained on only 20‑50 curated examples with a special transition prompt, it reliably activates multi‑object fusion, enabling high‑quality, controllable video customization without model architecture changes.
Flow‑GRPO introduces online reinforcement learning into flow‑matching models by converting deterministic ODE sampling to stochastic SDE sampling and reducing denoising steps, raising SD‑3.5‑Medium's GenEval accuracy from 63% to 95%—surpassing GPT‑4o—and demonstrating strong gains in complex composition, text rendering, and human‑preference alignment across multiple generative tasks.
The LocDiff framework introduces a spherical‑harmonics Dirac‑delta encoding and a conditional Siren‑UNet diffusion model that enables accurate worldwide image geolocation without relying on predefined grids or external image libraries, outperforming prior methods in precision, generalization, and computational efficiency.
The Free Transformer paper introduces a decoder architecture that injects random latent variables to condition generation, breaking traditional GPT constraints and achieving notable performance gains on reasoning‑heavy benchmarks such as HumanEval+, MBPP, GSM8K, MMLU, and CSQA.
The award‑winning industrial graph learning system developed by Peking University and Alibaba Mama combines novel dynamic graph embedding and GNN techniques, scales to millions of merchants, and has driven over 12% ad ROI improvement while publishing dozens of top‑conference papers.
AI researchAdvertising OptimizationGraph Neural Networks
VLM‑FO1 introduces a generate‑plus‑reference paradigm that replaces coordinate generation with region token referencing, adding plug‑in modules such as a proposal generator, a hybrid fine‑grained encoder, and a region‑language connector to give any pretrained visual language model accurate, fine‑grained perception while preserving its original capabilities.
The paper introduces RoboMonkey, a framework that applies a generate‑and‑verify paradigm and test‑time scaling to Vision‑Language‑Action models, showing that increasing sampling and verification at inference dramatically reduces action error across multiple VLA architectures, and presents scalable verifier training, synthetic data augmentation, and efficient deployment strategies.
OpenAI’s recent restructuring created the OpenAI Foundation, pledged $25 billion to health and AI‑resilience research, outlined a multi‑year AGI timeline, announced plans for AI hardware, and set milestones for an AI research intern by next September and a fully autonomous AI researcher by 2028.
This article analyzes the rStar2‑Agent paper, revealing how Agentic Reinforcement Learning, the GRPO‑RoC algorithm, a high‑throughput code‑execution service, and a three‑stage training recipe let a modest 14‑billion‑parameter model surpass much larger LLMs on challenging math benchmarks.
From Oct 11‑17 2025, this digest presents four recent AI papers on time‑series forecasting: FIRE introduces a frequency‑domain decomposition with independent amplitude‑phase modeling and adaptive weighting; CauchyNet leverages holomorphic activations for compact, data‑efficient learning; the EvoRate framework quantifies learnability via mutual information; and CoRA adds covariate‑aware adaptation to foundation models, all reporting significant accuracy gains and enhanced interpretability.
AI researchcovariate-aware adaptationdeep learning
This curated list showcases Meituan’s latest large‑model breakthroughs and academic papers up to October 2025, spanning LLM system optimizations, multimodal generation, evaluation benchmarks, quantization techniques, and reinforcement‑learning‑driven improvements, offering researchers valuable insights and resources across the AI landscape.
Google DeepMind hired Shunyu Yao, a Tsinghua physics prodigy and former Anthropic researcher, whose rapid transition from theoretical physics to AI highlights the intense workload, values clash, and the accelerating pace of large‑model research.
This article analyzes a recent paper that unifies supervised fine‑tuning (SFT) and reinforcement learning (RL) for large language models under a single gradient estimator, introduces the Unified Policy Gradient Estimator (UPGE) and the Hybrid Post‑Training (HPT) algorithm, and demonstrates their superior performance on math reasoning benchmarks.
This article reviews Huawei Noah's Ark Lab's exploration of large language models in recommendation systems, covering background challenges, the KAR and Uni-CTR projects, experimental results, and future research directions for open, knowledge‑driven recommendation pipelines.
AI researchHuaweiknowledge-assisted recommendation
Manzano introduces a hybrid vision tokenizer and a three‑stage training recipe that let a 3‑billion‑parameter multimodal LLM achieve state‑of‑the‑art results on both image‑understanding benchmarks and text‑to‑image generation, while scaling smoothly to larger sizes and minimizing task conflict.
FSDrive introduces a spatio‑temporal chain‑of‑thought approach that enables visual language models to generate future driving scenes as images, improving trajectory planning accuracy and safety by eliminating cross‑modal gaps and enforcing physical constraints in autonomous driving.
This article surveys recent advances in applying reinforcement learning to large reasoning models, outlining the historical background, key breakthroughs like OpenAI o1 and DeepSeek‑R1, current challenges in reward design and scalability, and future research directions toward more capable AI systems.
Tongyi DeepResearch, an open‑source AI model and framework, achieves SOTA on multiple Deep Research benchmarks by combining fully open‑source models, frameworks, and data pipelines, and introduces novel agentic pre‑training, fine‑tuning, and reinforcement‑learning methods to enable complex multi‑step reasoning and real‑world applications.
AI researchagentic reinforcement learningopen-source
DeepSeek‑R1, the first mainstream large language model to pass peer review in Nature, was trained for $294,000 using 648 H800 GPUs, and its RL‑enhanced version, DeepSeek‑R1‑Zero, achieved up to 86.7% pass@1 on AIME 2024, outperforming human averages across math, coding, and science tasks.
Ant Group and Renmin University unveiled LLaDA‑MoE, the industry’s first native MoE‑based diffusion language model trained on 20 TB of data, achieving performance comparable to Qwen2.5 while delivering several‑fold faster inference, and the model will be fully open‑sourced to accelerate global AI research.
Shunyu Yao, a Princeton‑trained AI researcher who rose through Tsinghua’s elite Yao class and OpenAI, is known for pioneering works like Tree of Thoughts, SWE‑bench, and ReAct, and now focuses on building general‑purpose agents and exploring the “second half” of AI development.
This article examines how large language model reasoning has progressed from linear Chain‑of‑Thought prompting to parallel Tree‑of‑Thought and flexible Graph‑of‑Thought approaches, highlighting each method’s mechanism, strengths, limitations, computational costs, and the broader shift toward cognitive‑centric AI research.
The article outlines ARC Lab’s evolution from its 2019 founding as an internal corporate research unit to a high‑impact AI team that pursues difficult multimodal understanding and generation problems, measures success through a technology‑impact funnel, publishes 30‑40 top‑tier papers annually, and translates research into open‑source tools and products that drive academic, industry, business, and societal value.
The article analyzes Denny Zhou's Stanford CS25 lecture on large language model reasoning, explaining how intermediate token generation, chain‑of‑thought prompting, self‑consistency, reinforcement‑learning fine‑tuning, and answer aggregation together unlock powerful reasoning capabilities beyond traditional greedy decoding.
This article examines the evolution of LangChain's Open Deep Research project from a monolithic LLM pipeline to a multi‑agent system, highlighting the role of context engineering, architectural trade‑offs, practical code examples, and best‑practice guidelines for building scalable, token‑efficient AI solutions.
The ERNIE‑4.5‑VL visual‑language model breaks single‑modality limits by delivering breakthrough image, video, and text understanding across more than 100 languages, offering lightweight yet competitive performance against models like Qwen2.5‑VL, supporting 128K context, dual “thinking” modes, and extensive deployment resources.
The Kwai Keye team presents Thyme, a novel multimodal reasoning framework that lets large language models generate and safely execute Python code for image manipulation and complex calculations, achieving significant performance gains over existing vision‑language models across perception, reasoning, and hallucination‑reduction benchmarks.
This article examines the evolution of large language model architectures over the past seven years, comparing key design choices such as Multi‑Head Latent Attention, Grouped‑Query Attention, Mixture‑of‑Experts, sliding‑window attention, normalization placement, and optimizer variants across models like DeepSeek V3, OLMo 2, Gemma 3, Llama 4, Qwen 3, SmolLM 3, and Kimi 2.
This article reviews a recent study that demonstrates how a highly compressed one‑dimensional tokenizer, using only 32 discrete tokens and gradient‑based test‑time optimization, can generate high‑quality images without training a separate generative model, and explores its methodology, findings, applications, and limitations.
This article reviews the most influential papers in large language model research since 2017, covering foundational works such as the Transformer, GPT‑3, BERT, scaling laws, and recent innovations like FlashAttention, Mamba, and QLoRA, highlighting their core contributions and impact on AI development.
MetaQuery introduces learnable queries that connect a frozen multimodal LLM with diffusion models, enabling knowledge‑enhanced image generation, reconstruction, and editing while preserving state‑of‑the‑art multimodal understanding, and achieves new SOTA results across multiple benchmarks.
Orthus, a new unified multimodal model presented at ICML 2025, leverages an autoregressive Transformer backbone with separate language and diffusion heads to enable lossless image‑text interleaved generation, outperforming existing models on both understanding and generation benchmarks while remaining computationally efficient.
AI researchImage Generationautoregressive transformer
The article examines how Meta’s CEO Mark Zuckerberg’s aggressive talent‑buying and commercial focus have sidelined Turing‑award winner Yang Li‑kun, detailing the restructuring of Meta’s AI labs, the clash over research directions, and the broader dilemma of balancing academic innovation with business imperatives in the AI industry.
A former OpenAI engineer shares a candid, unfiltered account of the company's fast‑paced growth, bottom‑up research culture, engineering practices, product decisions, and the intense seven‑week sprint that delivered Codex, offering valuable insights for AI researchers, product managers, and tech leaders.
The article analyzes Skywork OR1's technical report, detailing how small‑scale teams use GRPO‑based reinforcement learning with multi‑stage training, advantage‑mask variants, high‑temperature sampling, adaptive entropy loss, and resource‑allocation tricks to improve large language model performance while avoiding premature entropy collapse.
This article provides a comprehensive technical analysis of Meta's Llama series, tracing the evolution from Llama‑1 through Llama‑2 to Llama‑3, detailing model architectures, training data pipelines, optimization methods, benchmark results, and the broader impact on the open‑source AI community.
This article introduces VMBench, a perception‑aligned video motion generation benchmark with a five‑dimensional metric suite and meta‑guided prompt generation, and LD‑RPS, a zero‑shot unified image restoration framework using latent diffusion and recurrent posterior sampling, detailing their motivations, innovations, experiments, and future directions.
iQIYI’s AI research team secured three paper acceptances—two at ACL 2025 (including a main conference and a Findings paper) and one at INTERSPEECH 2025—covering long‑context large language model evaluation, Chinese novel summarization, and efficient Thai speech recognition, with links to each work.
EvoSearch, a test‑time evolutionary search method, dramatically improves image and video generation by increasing inference compute without extra training, outperforming existing scaling techniques on diffusion and flow models while maintaining robustness and diversity across multiple benchmarks.
This talk reviews the rapid evolution of generative large‑model AI from rule‑based systems to massive pre‑training, examines the current bottlenecks in continual learning and knowledge discovery, and proposes large‑scale meta‑learning—especially context‑based reinforcement learning (ICRL)—as a path toward truly autonomous, self‑learning agents.
Join Baidu senior algorithm engineer Ji Zhi at the DataFun Summit 2025 to explore how generative AI transforms ad recommendation recall, covering item representation, evolving solution architectures, long‑sequence challenges, and practical insights for building efficient large‑model recommendation systems.
The article surveys 117 recent multimodal‑fusion papers, classifies them into improvement‑based and combination‑based approaches, highlights representative works such as TimeXL, OGP‑Net, MMR‑Mamba and FusionSight, and provides a free collection of papers, classic models and code repositories for researchers.
Ant Research introduces GCA, a causal retrieval‑based grouped cross‑attention mechanism that end‑to‑end learns to fetch relevant past chunks, dramatically reducing memory usage and achieving over 1000× length generalization on long‑context language modeling tasks, with near‑constant inference memory and linear training cost.
Kuaishou presented twelve peer‑reviewed papers at CVPR 2025 covering video quality assessment, large‑scale video datasets, dynamic 3D avatar reconstruction, 4D scene simulation, controllable video generation, scaling laws for diffusion transformers, multimodal foundations, and more, highlighting the company's leading research in computer vision and AI.
The article introduces DiSA, a training‑free diffusion step annealing technique that dramatically speeds up autoregressive image generation by reducing diffusion steps in later generation phases while preserving high visual quality, and validates the method across several state‑of‑the‑art AR‑Diffusion models.
The article compares autoregressive and diffusion language models, detailing their mathematical foundations, training and inference pipelines, performance trade‑offs such as speed, coherence and diversity, and explores hybrid approaches and emerging research directions for more efficient and controllable text generation.
The article examines the current dominance of diffusion models in commercial video generation, contrasts them with autoregressive methods, and details how the open‑source MAGI‑1 model combines both paradigms to achieve longer, more controllable video synthesis while addressing scalability and quality challenges.
The Mass framework, developed by Google and Cambridge University, automates multi‑agent system design by jointly optimizing prompts and topologies through three staged processes, demonstrating significant performance gains over existing methods across various tasks while highlighting the importance of coordinated prompt‑topology optimization.
dots.llm1, an open‑source 142‑billion‑parameter Mixture‑of‑Experts language model from hi lab, achieves Qwen2.5‑72B‑level performance after training on 11.2 T high‑quality tokens, and the release includes full models, intermediate checkpoints, and detailed training pipelines for the research community.
AI researchTraining Efficiencylarge language model
This article systematically explains the concepts of Deep Search and Deep Research, contrasts them with traditional Retrieval‑Augmented Generation, reviews leading commercial and open‑source solutions, details their architecture for code retrieval, and outlines future plans for specialized code‑search agents.
This article critically examines seven high‑profile reinforcement‑learning papers for large language models, exposing flawed baseline evaluations, unrealistic settings, and modest actual improvements despite bold claims of dramatic performance gains.
The KwaiPilot team released the KwaiCoder‑AutoThink‑preview model, which introduces a novel automatic‑thinking training paradigm and a process‑supervised reinforcement‑learning method called Step‑SRPO, enabling the model to dynamically switch between thinking and non‑thinking modes, reduce inference cost, and achieve up to 20‑point gains on code and math benchmarks while handling large‑scale codebases.
The TailoredBench framework dramatically reduces large‑language‑model evaluation cost and error by using a global probe set, model‑specific source selection, extensible K‑Medoids clustering, and calibration, achieving up to 300× speedup and a 31.4% MAE reduction across diverse benchmarks.
The article explores how DeepSeek R1 and long‑thinking chains have revived interest in machine reasoning, tracing the evolution of natural‑language models, defining reasoning as logical knowledge composition, and outlining future research directions in efficient reasoning architectures and deep‑thinking applications.
Anthropic has introduced a voice mode for Claude, enabling English users to speak and type interchangeably with five voice personalities, while a new 3D AI startup, SpAItial, showcases photorealistic room generation and researchers present INTUITOR, a confidence‑driven training method that improves AI reasoning.
The article analyzes the Token‑Shuffle method, which reduces visual token redundancy to allow high‑resolution (2048×2048) autoregressive image generation, detailing its architecture, training pipeline, experimental results, efficiency gains, and comparisons with diffusion and other AR models.
AI researchautoregressive modelshigh-resolution image generation
The article presents MeanFlow, a novel one‑step generative modeling framework that replaces instantaneous velocity with an average‑velocity field, achieving a record‑low FID of 3.43 on ImageNet 256×256 with a single function evaluation and outperforming both prior single‑step and multi‑step diffusion models.
This article reviews ten recent papers that tackle the over‑thinking problem in large language models by shortening chain‑of‑thought reasoning, introducing dynamic early‑exit, adaptive thinking triggers, and reinforcement‑learning‑based training, showing how models can maintain or improve accuracy while dramatically reducing token usage and latency.
The article introduces three BGE vector models—BGE‑Code‑v1, BGE‑VL‑v1.5, and BGE‑VL‑Screenshot—detailing their architectures, open‑source resources, benchmark results on CoIR, Code‑RAG, MMEB, and MVRB, and their impact on code and multimodal retrieval research.
The award‑winning project from Tsinghua University and JD Retail introduces re‑parameterization model design, cross‑scene adaptive learning, and platform‑aware compression to overcome accuracy‑efficiency trade‑offs in visual deep learning, achieving over 20% accuracy gains and more than 50% inference speedup in real‑world e‑commerce deployments.
DreamO, a collaborative effort by ByteDance and Peking University, introduces a unified 400M‑parameter framework built on Flux‑1.0‑dev that enables simultaneous control of identity, style, appearance, and virtual try‑on, offering open‑source, low‑cost, and fast image customization comparable to commercial large models.
The article introduces Group Policy Gradient (GPG), a reinforcement‑learning framework that eliminates surrogate loss functions and critic models, directly optimizes the original objective, reduces bias and variance, and achieves state‑of‑the‑art performance on both single‑modal and multimodal tasks.
The paper introduces G3PT, a groundbreaking autoregressive 3D generation model that employs a Cross‑Scale Querying Transformer and multi‑scale tokenization to produce high‑quality meshes from a single image, outperforming diffusion‑based methods and revealing a scaling law for 3D generation.
A new study introduces the lightweight “Canon” layer for large language models, showing how it improves information flow, inference depth, and scalability across Transformers, linear attention, and state‑space architectures, while offering a controlled synthetic pre‑training benchmark for deeper architectural analysis.
Recent empirical research by Tsinghua’s LeapLab and Shanghai Jiao Tong University reveals that reinforcement‑learning‑based fine‑tuning (RLVR) improves sampling efficiency but does not extend the fundamental reasoning abilities of large language models beyond their base capabilities, as demonstrated across mathematics, code, and visual reasoning benchmarks.
Microsoft Research introduced BitNet b1.58 2B4T, a native 1‑bit large language model with 2 billion parameters trained on 4 trillion tokens, achieving only 0.4 GB non‑embedding memory, 0.028 J decoding energy, and 29 ms CPU latency while matching full‑precision performance.
This reflective essay traces reinforcement learning’s decade‑long evolution through four stages—early algorithmic foundations, application‑driven growth, problem‑construction focus, and speculative future—while critiquing the expanding definition and its impact on research and industry.
OpenAI has launched GPT‑4.1, a multimodal AI model that expands context windows to one million tokens, improves coding and instruction following, offers cheaper Mini and Nano variants, and signals a shift in its release roadmap, including plans to retire GPT‑4 and delay GPT‑5.
The article argues that AI has entered its second half, where reinforcement learning finally generalizes thanks to large‑scale language pretraining and reasoning, shifting focus from building ever better models to redefining problems, evaluation methods, and real‑world utility.
This article reviews the breakthrough image‑generation capabilities of GPT‑4o, showcases diverse examples, and offers a detailed speculation on its underlying autoregressive architecture, tokenization methods, VQ‑VAE/GAN advances, and training strategies that could explain its performance.
At the ICLR 2025 live session in Singapore, Ant Group showcased four cutting‑edge papers—CodePlan, Animate‑X, Group Position Embedding, and OmniKV—demonstrating advances in large‑language‑model reasoning, universal character animation, layout‑aware document understanding, and efficient long‑context inference.
AI researchDocument Understandinglarge language models
The article introduces Meta's newly open‑sourced Llama 4 series—including Scout with a 1 billion‑token context window, Maverick with 400 billion parameters, and the upcoming Behemoth teacher model—detailing their expert‑mix architecture, the NoPE positional‑encoding removal, training pipelines, performance benchmarks, and infrastructure improvements for large‑scale AI research.
The article presents the abstracts, publication types, links, and research areas of seventeen Ant Group papers accepted at ICLR 2025, covering topics such as embodied robot co‑design, efficient distributed training for large language models, optimization via LLMs, character animation, interactive frame interpolation, KV‑cache management, and privacy‑preserving Transformers.
Google researchers discovered a linear relationship between brain activity recorded during natural conversation and the internal embeddings of a speech‑to‑text large language model, revealing that acoustic and lexical representations from the model can accurately predict neural responses in both language comprehension and production.
This article provides an in‑depth analysis of the NOVA model, a non‑quantized autoregressive video generation framework that combines frame‑by‑frame temporal prediction with set‑by‑set spatial prediction, uses diffusion loss for token estimation, and achieves state‑of‑the‑art results on multiple video and image benchmarks.
MMGDreamer introduces a mixed‑modality graph and a dual‑branch diffusion model that combine text, image, and relational cues to generate highly realistic, geometrically controllable 3D indoor scenes, outperforming prior methods on multiple quantitative and qualitative benchmarks.
Alibaba's newly released Qwen2.5‑VL‑32B multimodal model delivers state‑of‑the‑art visual and textual performance, offering human‑aligned responses, superior mathematical reasoning, fine‑grained image understanding, and efficient deployment features that make it a compelling tool for developers and AI researchers alike.
This article provides an in‑depth review of alignment algorithms for multimodal large language models, covering application scenarios, dataset construction methods, evaluation benchmarks, current challenges, and future research directions, while summarizing contributions from leading academic institutions.
AI researchalignment algorithmsdataset construction
This article surveys the latest multimodal large language model research, dissecting the design, training strategies, and performance trade‑offs of models such as Llama 3.2, Molmo, NVLM, Qwen2‑VL, Pixtral, MM1.5, Emu3, and Janus, and highlights the challenges of fair cross‑model evaluation.
AI researchModel Training Strategiescross-attention
Researchers He Kaiming, Yann LeCun and colleagues propose a 9‑line Dynamic Tanh (DyT) layer that replaces LayerNorm/RMSNorm in Transformers, showing comparable or superior accuracy across vision, language, speech and DNA tasks while also reducing inference latency on modern GPUs.
The article recounts a JD retail engineer’s rapid growth through tackling low‑resource large‑model training, developing a margin‑based dynamic data selection method (DynaMS) that earned an ICLR paper, and sharing practical insights on aligning business needs with cutting‑edge AI research.
The paper presents FAR, a frequency‑aware autoregressive framework that predicts image tokens from low‑frequency to high‑frequency components using a continuous tokenizer, and demonstrates its efficiency and quality on ImageNet and text‑to‑image benchmarks compared with existing AR and VAR methods.
This presentation reviews the development, technical details, and societal impact of DeepSeek's R1 model, explaining its reasoning capabilities, training pipeline, comparisons with other models, and future directions for AI research and product applications.
The paper introduces a novel attention‑distillation loss and a guided‑sampling scheme that together enable diffusion models to faithfully transfer visual features from reference images, dramatically speeding synthesis and surpassing prior plug‑and‑play attention methods across style transfer, text‑to‑image generation, and texture synthesis tasks.
Compiled by Tsinghua University experts, this guide systematically analyzes the DeepSeek‑R1 inference model and DeepResearch platform, offering multi‑model comparisons, real‑world case studies, and end‑to‑end AI‑driven solutions from data collection to report generation for researchers.
This article examines how to build and improve reasoning‑capable large language models, explains the definition and use‑cases of reasoning models, details DeepSeek‑R1’s training pipeline, compares four key enhancement methods—including inference‑time scaling, pure RL, SFT + RL, and distillation—and offers budget‑friendly advice.
The DeepSeek technical overview details the evolution from the dense 67 B V1 model through the 236 B MoE‑based V2 and 671 B V3 with FP8 training, to the RL‑only R1 series that learns reasoning without supervision, highlighting innovations such as Grouped‑Query Attention, Multi‑Head Latent Attention, load‑balancing‑free MoE, Multi‑Token Prediction, and knowledge distillation, and reporting state‑of‑the‑art benchmark results and open‑source reproduction projects.