---
title: Papers with code
url: https://www.ml-quant.com/library/pwc/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-09-26
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
---


# Papers with code

Papers that shipped their code, from the Papers with Code feed (2023-25).

- [SkillsVote: Governance for Agent Skills](https://github.com/MemTensor/skills-vote) (2026-05-20): Governance for Agent Skills: SkillsVote is a system that helps long-term AI agents manage and improve their skills over time.
- [MMSkills: Multimodal Skills for Agents](https://github.com/DeepExperience/MMSkills) (2026-05-20): Multimodal Skills for Agents: Multimodal procedural knowledge frameworks enhance visual agents by merging text and visuals for improved decision-making.
- [SelfDistilled RL: Enhancing Training with Self-Distillation](https://github.com/ZJU-REAL/SDAR) (2026-05-20): Enhancing Training with Self-Distillation: SDAR optimizes training for multiturn agents in reinforcement learning by boosting positive feedback and minimizing negative feedback.
- [AI Research Trust](https://github.com/worldbench/awesome-ai-auto-research) (2026-05-20): AI excels in structured tasks but needs human oversight in scientific research to handle new ideas and judgment effectively.
- [Olympiad Reasoning Scaling](https://github.com/Simplified-Reasoning/SU-01) (2026-05-20): A new systematic approach boosts reasoning models, making them top contenders in math and physics competitions through advanced learning methods.
- [Dynamic NPC Steering](https://github.com/INV-WZQ/ReactiveGWM) (2026-05-20): ReactiveGWM improves game play by separating player controls from NPC actions, enhancing strategic options in various games.
- [DexJoCo for Manipulation](https://github.com/brave-eai/dexjoco) (2026-05-20): DexJoCo provides a benchmark and toolkit for evaluating dexterous manipulation skills, along with a budget-friendly data collection system.
- [SkillClaw: Skill Evolution](https://github.com/AMAP-ML/SkillClaw) (2026-04-16): Skill Evolution: SkillClaw enhances multiuser AI systems by leveraging group interactions to strengthen shared abilities.
- [ClawGUI: Unified GUI](https://github.com/ZJU-REAL/ClawGUI) (2026-04-16): Unified GUI: ClawGUI is an open-source tool that streamlines the development of GUI agents using unified reinforcement learning across different platforms.
- [DDTree: Speculative Decoding](https://github.com/liranringel/ddtree) (2026-04-16): Speculative Decoding: DDTree improves speculative decoding by generating draft trees from data distributions and validating multiple paths at once.
- [Strips as Tokens](https://github.com/Xrvitd/SATO) (2026-04-16): SATO introduces a new way to order tokens in transformers that improves mesh generation by preserving edge flow using triangle strip sequences.
- [Introspective Consistency](https://github.com/Introspective-Diffusion/I-DLM) (2026-04-16): Introspective Diffusion Language Models enhance autoregressive models by refining their output consistency through advanced decoding and optimized methods.
- [HabitatGS Navigation](https://github.com/zju3dv/habitat-gs) (2026-04-16): HabitatGS enhances HabitatSim with 3D Gaussian Splatting for realistic visuals and dynamic avatars, improving AI agent navigation and generalization.
- [KnowUBench: Mobile Agent Evaluation](https://github.com/ZJU-REAL/KnowU-Bench) (2026-04-16): Mobile Agent Evaluation: KnowUBench assesses how well personalized mobile agents can understand user preferences and provide helpful assistance in real-world graphical user interfaces.
- [KnowRL: LLM Reasoning Enhancement](https://github.com/Hasuer/KnowRL) (2026-04-16): LLM Reasoning Enhancement: KnowRL improves the reasoning abilities of language models through a framework that uses reinforcement learning to provide better, guided interactions.
- [OnPolicy Distillation in Language Models](https://github.com/thunlp/OPD) (2026-04-16): Effective distillation in large language models depends on matching thought processes between teacher and student models, with teachers needing to impart new skills.
- [Parallel Decoding for Diffusion Models](https://github.com/czg1225/DMax) (2026-04-16): DMax introduces a new technique for diffusion language models that reduces mistakes in parallel decoding.
- [Autonomous ML with AiScientist](https://github.com/AweAI-Team/AiScientist) (2026-04-16): AiScientist develops a system that boosts long-term machine learning research by improving coordination and project management.
- [Benchmarking LLMs for Human Behavior Simulation](https://github.com/icip-cas/OmniBehavior) (2026-04-16): The OmniBehavior benchmark reveals that large language models have difficulty mimicking complex human behaviors due to biases and limited diversity.
- [Latent Space as Foundation](https://github.com/YU-deep/Awesome-Latent-Space) (2026-04-03): Latent space improves language models by creating a continuous representation that minimizes redundancy and boosts efficiency.
- [Unified Multimodal Processing](https://github.com/meituan-longcat/LongCat-Next) (2026-04-03): The Discrete Native Autoregressive framework enables integrated handling of various data types through a common discrete space and innovative visual transformer design.
- [Generative World Renderer: Enhanced AAA Game Rendering](https://github.com/ShandaAI/AlayaRenderer) (2026-04-03): Enhanced AAA Game Rendering: A new dataset from AAA games enhances rendering quality and better evaluation methods that match human perception.
- [SKILL0: RL for Skill Internalization](https://github.com/ZJU-REAL/SkillZero) (2026-04-03): RL for Skill Internalization: SKILL0 empowers LLM agents to autonomously learn and execute tasks, boosting their effectiveness with a flexible training process.
- [GEMS Multimodal Generation Framework with Memory](https://github.com/lcqysl/GEMS) (2026-04-03): GEMS introduces a multimodal framework that helps agents refine their skills and memory, leading to improved performance across different tasks.
- [dLLM Diffusion Framework](https://github.com/ZHZisZZ/dllm) (2026-03-04): A new open-source framework has been created to make it easier and more consistent to develop and customize diffusion language models.
- [MobilityBench: Route Planning Evaluation](https://github.com/AMAP-ML/MobilityBench) (2026-03-04): Route Planning Evaluation: MobileBench offers a scalable way to evaluate route-planning agents using large language models with real-world anonymized query testing.
- [Trinity of World Model Consistency](https://github.com/openraiser/awesome-world-model-evolution) (2026-03-04): To improve World Models, three key principles—modal, spatial, and temporal—are identified, along with a benchmark for assessing multimodal learning systems.
- [BeyondSWE: Evaluating Code Agents](https://github.com/AweAI-Team/BeyondSWE) (2026-03-04): Evaluating Code Agents: BeyondSWE and SearchSWE are new benchmarks that improve the evaluation of coding agents by including real-life reasoning and knowledge.
- [ARLArena: Stable Reinforcement Learning](https://github.com/WillDreamer/ARL-Arena) (2026-03-04): Stable Reinforcement Learning: The ARLArena framework boosts reinforcement learning training stability through the SAMPO method for more consistent policy optimization.
- [MediXR1: Medical RL for Clinical Reasoning](https://github.com/mbzuai-oryx/MediX-R1) (2026-03-04): Medical RL for Clinical Reasoning: MediXR1 is a flexible reinforcement learning framework for medical language models, enhancing clinical reasoning with varied rewards and assessments.
- [Flash Frontier: Sparse Agentic Intelligence](https://github.com/stepfun-ai/Step-3.5-Flash) (2026-02-12): Sparse Agentic Intelligence: Step 3.5 Flash is an AI model designed for advanced intelligence using efficient parameters and attention mechanisms.
- [InternAgent 1.5: Autonomous Discovery System](https://github.com/InternScience/InternAgent) (2026-02-12): Autonomous Discovery System: InternAgent1.5 is a system that integrates computational modeling with experimental research for autonomous scientific discovery.
- [CodeWorld: GUI Agents for Visual States](https://github.com/AMAP-ML/Code2World) (2026-02-12): GUI Agents for Visual States: Code2World enables autonomous GUI agents to predict visual outcomes by generating renderable code, improving navigation.
- [SkillRL: Evolving RL Agents](https://github.com/aiming-lab/SkillRL) (2026-02-12): Evolving RL Agents: SkillRL enhances large language models by enabling skill discovery and policy evolution while conserving computational resources.
- [UniAudio 2.0: Unified Audio Model](https://github.com/yangdongchao/UniAudio2) (2026-02-12): Unified Audio Model: Researchers developed ReasoningCodec for improved audio processing and UniAudio 2.0, a strong model for various audio tasks using a large dataset.
- [WeakDriven Learning: Improving Agents with Weak Checkpoints](https://github.com/chenzehao82/Weak-Driven-Learning) (2026-02-12): Improving Agents with Weak Checkpoints: WMSS boosts large language models by using weak checkpoints to close learning gaps for better results.
- [Agent World Model: Advancing RL in Synthetic Environments](https://github.com/Snowflake-Labs/agent-world-model) (2026-02-12): Advancing RL in Synthetic Environments: Large language models in synthetic settings exceed traditional models in adapting to new situations.
- [ROCKET: Model Compression via Knapsack Optimization](https://github.com/mts-ai/ROCKET) (2026-02-12): Model Compression via Knapsack Optimization: ROCKET introduces an efficient model compression technique that treats the process like solving a knapsack problem.
- [Prism: Enhanced Block-Sparse Attention for LLMs](https://github.com/xinghaow99/prism) (2026-02-12): Enhanced Block-Sparse Attention for LLMs: Prism enhances long-context LLM performance by optimizing which blocks to select in blocksparse attention.
- [LOCAbench: Benchmarking Language Agents in Extreme Contexts](https://github.com/hkust-nlp/LOCA-bench) (2026-02-12): Benchmarking Language Agents in Extreme Contexts: LOCAbench is a benchmark designed to evaluate language agents handling long-context tasks.
- [AI Safety Framework](https://github.com/AI45Lab/AgentDoG) (2026-02-02): AI agents require better safety protocols when using tools and interacting with their surroundings.
- [SelfEvolving Systems](https://github.com/YunjueTech/Yunjue-Agent) (2026-02-02): Adaptive agents that learn from feedback can adapt more effectively to changing environments and enhance knowledge sharing.
- [Enhancing Math Reasoning](https://github.com/AMAP-ML/MathForge) (2026-02-02): MathForge enhances AI's mathematical reasoning by adjusting question difficulty and reformulating problems.
- [ASTRA: Automated Decision Framework](https://github.com/LianjiaTech/astra) (2026-02-02): Automated Decision Framework: ASTRA trains language models with synthetic data to improve their ability to make complex decisions.
- [AdaReasoner: Visual Reasoning Tool](https://github.com/ssmisya/AdaReasoner) (2026-02-02): Visual Reasoning Tool: AdaReasoner enables multimodal models to learn how to use tools effectively for enhanced visual reasoning through scalable data and adaptive methods.
- [Token Filtering for Models](https://github.com/neilrathi/token-filtering) (2026-02-02): Token filtering during pretraining minimizes undesirable traits in language models and enhances performance while managing noisy labels.
- [SelfDistillation in RL](https://github.com/lasgroup/SDPO) (2026-02-02): SelfDistillation Policy Optimization (SDPO) enhances reinforcement learning by incorporating textual feedback to make training language models more efficient and accurate.
- [Scalable Conditional Memory](https://github.com/deepseek-ai/Engram) (2026-01-16): Conditional memory in Transformer models improves knowledge retrieval and reasoning by efficiently managing data sparsity.
- [Unified Multimodal Retrieval](https://github.com/QwenLM/Qwen3-VL-Embedding) (2026-01-16): The Qwen3VLEmbedding and Qwen3VLReranker models form a precise multimodal search system using cross-attention techniques.
- [Advancements in 3D Orientation](https://github.com/SpatialVision/Orient-Anything-V2) (2026-01-16): Orient Anything V2 improves understanding of 3D orientation through new asset synthesis and rotation prediction methods.
- [Optimized Code Generation](https://github.com/QuantaAlpha/EvoControl) (2026-01-16): The Controlled SelfEvolution method boosts code generation by employing genetic evolution and better exploration strategies.
- [Automated Research Evaluation](https://github.com/Infinity-AILab/DeepResearchEval) (2026-01-16): DeepResearchEval uses adaptable agents to automate complex research tasks, verifying facts without relying on citations.
- [Linear Attention Enhancement](https://github.com/DAGroup-PKU/MHLA) (2026-01-16): MultiHead Linear Attention boosts performance by balancing representational diversity and computational efficiency while increasing expressive capabilities.
- [Synthetic Data for Programming](https://github.com/JieWu02/X-Coder) (2026-01-16): Code LLMs trained on synthetic data excel in competitive programming compared to traditional models, reducing dependence on real datasets.
- [SelfEvolving Reasoning Agents](https://github.com/facebookresearch/drzero) (2026-01-16): A data-free self-evolution framework enables large language models to improve reasoning by generating their own questions, achieving performance similar to supervised learning.
- [Stable Sinkhorn-Knopp Iterations](https://github.com/FFTYYY/mhc-lite) (2026-01-16): Hyperconnections with dynamic residual matrices improve convergence stability through a new reparameterization technique that ensures precise doubly stochasticity.
- [Agentic AI Adaptation Framework](https://github.com/pat-jj/Awesome-Adaptation-of-Agentic-AI) (2025-12-28): The paper presents a framework for improving AI systems by adapting agents and tools, focusing on design strategies and challenges to boost AI performance.
- [Evaluating LLMs for Science](https://github.com/InternScience/SGI-Bench) (2025-12-28): A new framework for Scientific General Intelligence (SGI) is introduced, enhancing existing models with TestTime Reinforcement Learning to improve their scientific capabilities.
- [Uncovering Language Model Policies](https://github.com/Trae1ounG/BuPO) (2025-12-28): The research analyzes large language model policies by dissecting them into modular parts, revealing reasoning patterns and advocating for Bottom-up Policy Optimization to enhance complex reasoning tasks.
- [StepGUI: Optimizing GUI Automation](https://github.com/stepfun-ai/gelab-zero) (2025-12-19): Optimizing GUI Automation: A new training system boosts GUI automation by making it more efficient, accurate, and private for real-world use.
- [AI Agents: Memory Research Overview](https://github.com/Shichun-Liu/Agent-Memory-Paper-List) (2025-12-19): Memory Research Overview: The survey examines agent memory research, detailing its types, functions, and future research directions.
- [Jacobi Forcing: Parallel Decoding Efficiency](https://github.com/hao-ai-lab/JacobiForcing) (2025-12-19): Parallel Decoding Efficiency: Jacobi Forcing enables quicker and more efficient parallel decoding of transformer models while maintaining performance.
- [LitePT: Efficient 3D Model](https://github.com/prs-eth/LitePT) (2025-12-19): Efficient 3D Model: LitePT is a 3D point cloud model that uses a combination of convolutions and attention to enhance efficiency.
- [ErrorFree Attention: Optimized Language Model](https://github.com/declare-lab/EFLA) (2025-12-19): Optimized Language Model: ErrorFree Linear Attention (EFLA) is a quick and dependable attention method that outperforms DeltaNet in language processing tasks.
- [Universal Reasoning: Enhanced Transformer](https://github.com/zitian-gao/URM) (2025-12-19): Enhanced Transformer: The Universal Reasoning Model enhances Universal Transformers to improve reasoning abilities on ARCAGI challenges.
- [DeepCode: Doc-to-Code Synthesis](https://github.com/HKUDS/DeepCode) (2025-12-14): Doc-to-Code Synthesis: DeepCode is an autonomous framework that improves the process of turning documents into code, surpassing human experts with its advanced optimization techniques.
- [GRAPE: Positional Encoding Framework](https://github.com/model-architectures/GRAPE) (2025-12-14): Positional Encoding Framework: GRAPE is a new framework for positional encoding that combines rotations and logit biases to enhance the performance of existing methods like RoPE and ALiBi.
- [Textto-3D Reinforcement Learning](https://github.com/Ivan-Tang-3D/3DGen-R1) (2025-12-14): ARDR1 is a groundbreaking model that uses reinforcement learning to generate 3D content from text, featuring new reward systems and optimization methods.
- [Procedural Terrain Diffusion](https://github.com/xandergos/terrain-diffusion) (2025-12-14): Terrain Diffusion leverages diffusion models and InfiniteDiffusion to produce realistic, infinitely expandable environments that can be accessed quickly.
- [Evolving Cost-Effective Deep Learning Rubrics](https://github.com/rlresearch/dr-tulu) (2025-12-01): Reinforcement Learning with Evolving Rubrics (RLER) enhances the affordable training of deep learning models for lengthy tasks and outperforms current methods.
- [Efficient Memory](https://github.com/VectorSpaceLab/general-agentic-memory) (2025-12-01): GAM is a framework that improves memory efficiency and task performance through JIT compilation, a lightweight memorizer, and reinforcement learning.
- [Collaborative Representation](https://github.com/Gen-Verse/LatentMAS) (2025-12-01): LatentMAS boosts collaboration among LLM agents by utilizing latent space representations, enhancing reasoning quality while reducing computational expenses.
- [MiroThinker — Scalable Research Agent](https://github.com/MiroMindAI/MiroThinker) (2025-11-19): MiroThinker v1.0 — An open-source agent that boosts reasoning by training models to interact more with tools and environments, enabling long multi-step workflows and strong benchmark results.
- [P RL Physics Olympiad Champion](https://github.com/PRIME-RL/P1) (2025-11-19): P A family of open-source, reinforcement-learned models that excel at physics reasoning (winning top medals at IPhO 2025) and also perform well on math and coding.
- [PartXMLLM — 3D Part-aware LLM](https://github.com/AiEson/Part-X-MLLM) (2025-11-19): PartXMLLM: Turns a colored 3D point cloud and a text prompt into a compact program that lists part-level boxes, labels, and edit commands to drive geometry tools.
- [LoopTool — Robust LLM Call Loop](https://github.com/Rednote-ExperienceAI-Lab/LoopTool) (2025-11-19): LoopTool: An automated loop that repeatedly generates tool-using examples, refines data and models, and retrains LLMs to improve their tool-use and overall performance.
- [DeepEyesV2: Agentic Multimodal](https://github.com/Visual-Agent/DeepEyes) (2025-11-12): Agentic Multimodal: DeepEyesV2 — a two-stage-trained multimodal agent that learns to use external tools to solve real-world reasoning tasks.
- [Diffusion LMs: Low-Data Learners](https://github.com/JinjieNi/MegaDLMs) (2025-11-12): Low-Data Learners: Diffusion language models — they predict tokens via iterative denoising and flexible token order, outperforming autoregressive models with limited data and keeping advantages as they scale.
- [Adaptive Verifiable Environments for LM Reasoning](https://github.com/Zhiyuan-Zeng/RLVE) (2025-11-12): RLVE Gradually increases problem difficulty during training so language models learn stronger reasoning, outperforming fixed environments and standard RL.
- [Retrofitted Recurrence for Deeper LM Reasoning](https://github.com/mcleish7/retrofitting-recurrence) (2025-11-12): Recurrence curriculum: Convert pretrained nonrecurrent models to depth-recurrent ones and train with an increasing-recurrence curriculum to get better results for the same compute.
- [Kimi Linear: Hybrid Linear Attention](https://github.com/MoonshotAI/Kimi-Linear) (2025-11-04): Hybrid Linear Attention: Kimi Linear: a new hybrid linear‑attention model that beats full attention while using far less memory, decoding much faster, and with code and checkpoints open‑sourced.
- [FP Fixes Training–Inference Mismatch in RL Fine-tuning](https://github.com/sail-sg/Precision-RL) (2025-11-04): Switching RL fine‑tuning from BF16 to FP16: fixes numerical mismatches between training and inference, making training more stable and faster.
- [Toolathlon: Long‑Horizon Agent Benchmark](https://github.com/hkust-nlp/Toolathlon) (2025-11-04): Long‑Horizon Agent Benchmark: Toolathlon — a benchmark across 32 apps and 604 tools that tests agents on long, realistic workflows and shows current models often fail to complete them.
- [CALM Continuous Next‑Vector Models](https://github.com/shaochenze/calm) (2025-11-04): CALM — a method that compresses multiple tokens into one continuous vector and predicts those vectors instead of tokens, cutting generation steps and making generation much faster.
- [DeepAgent — Autonomous Tool Agent](https://github.com/RUC-NLPIR/DeepAgent) (2025-10-27): DeepAgent — an end-to-end agent that finds and uses tools on its own and compresses long interaction histories so it solves complex, long-horizon tasks better.
- [LightMem — Efficient Memory for LLMs](https://github.com/zjunlp/LightMem) (2025-10-27): LightMem — a lightweight three-stage memory that filters, groups, and consolidates past interactions to improve LLM accuracy while greatly cutting token, API, and runtime costs.
- [DeepAnalyze — Auto Data-Science LLM](https://github.com/ruc-datalab/DeepAnalyze) (2025-10-27): DeepAnalyze8B: an 8B-parameter model trained progressively to run full data-science pipelines and produce analyst-quality research reports.
- [PBSAttn — Block-Sparse Attention](https://github.com/xinghaow99/pbs-attn) (2025-10-27): PBSAttn: rearranges attention into block-sparse patterns to speed long-context LLM processing while keeping accuracy close to full attention.
- [EDR Steerable Multi-Agent Analytics](https://github.com/SalesforceAIResearch/enterprise-deep-research) (2025-10-27): EnterpriseDeepResearch (EDR): a multiagent system that plans, searches, uses tools, visualizes results, and reflects to automatically produce high-quality enterprise research reports.
- [Transformer Explainer](https://github.com/poloclub/transformer-explainer) (2025-07-25): Despite the significant impact of Transformers on machine learning, their workings remain unclear to many people.
- [Superhuman Reasoning](https://github.com/alibaba-nlp/webagent) (2025-07-25): The article discusses the challenge of surpassing human cognitive limitations in the training of Large Language Models (LLMs).
- [Workflow Generation Models](https://github.com/aidc-ai/comfyui-copilot) (2025-07-25): The article presents a curated dataset of 4K workflows used to create comprehensive reasoning data, including aspects like node selection, workflow planning, and code-level workflow representation.
- [GLM4 vs Qwen2.5VL7B](https://github.com/thudm/glm-4.1v-thinking) (2025-07-25): The article discusses a model that performs better than Qwen2. 5VL7B in 28 public benchmarks and equals or surpasses Qwen2. 5VL72B in 18 benchmarks.
- [NVIDIA Audio Flamingo](https://github.com/NVIDIA/audio-flamingo) (2025-07-25): The article emphasizes the importance of enhancing large language models to understand audio, including non-speech sounds and non-verbal speech, for various real-world applications.
- [Streaming 4D VGT](https://github.com/wzzheng/streamvggt) (2025-07-25): The article introduces a streaming 4D visual geometry transformer, similar to autoregressive large language models, designed to support interactive and real-time applications.
- [IQLearn: Inverse softQ](https://github.com/Div99/IQ-Learn) (2025-07-25): Inverse softQ: The article highlights the availability of human or expert data in sequential decision-making tasks like robotics control and game playing, which offers valuable task-related information.
- [Benchmarking Autonomous Agents](https://github.com/agi-inc/agisdk) (2025-07-25): The article presents REAL, a tool for assessing multiturn agents' performance on simulations of actual websites.
- [Mathematical Deep Learning Introduction](https://github.com/introdeeplearning/book) (2025-07-25): The book offers a beginner-friendly guide to understanding deep learning algorithms.
- [HackSynth LLM Agent](https://github.com/aielte-research/HackSynth) (2025-07-25): The article introduces HackSynth, an agent based on Large Language Model for independent penetration testing.
- [PhysX Asset Generation](https://github.com/ziangcao0312/PhysX) (2025-07-25): The article explores the shift of 3D modeling from a virtual environment to a physical one.
- [Dynamic 4D Scene Forecasting](https://github.com/lzzzzzm/ii-world) (2025-07-25): The article introduces a method that divides scene tokenization into two parts: intrascene and interscene tokenizers.
- [Conv3x3 Designs in Diffusion Models](https://github.com/yuchuantian/dic) (2025-07-25): The article emphasizes the exceptional performance of diffusion models in tasks related to visual generation.
- [AutoSteering Trajectories](https://github.com/lileyang1210/ant) (2025-07-25): The article points out problems with anchorfree methods leading to visual artifacts and anchorbased methods relying on heuristic selection of anchor concepts.
- [Safety Vulnerability of Diffusion LLMs](https://github.com/zichenwen1/dija) (2025-07-25): The article presents DIJA, a system that uses adversarial interleaved masktext prompts to control the text generation processes of dLLMs.
- [GumTreeDiff](https://github.com/GumTreeDiff/gumtree) (2025-07-17): The first article explores the idea of software evolution via an edit script applied to a source code file.
- [cmriat L0](https://github.com/cmriat/l0) (2025-07-17): The second article presents LZero (L0), a training pipeline designed for general-purpose agents.
- [apple Embedding Atlas](https://github.com/apple/embedding-atlas) (2025-07-17): The third article discusses the application of embedding projections in visualizing large datasets and models.
- [yongliu20](https://github.com/yongliu20/UniLSeg) (2025-07-17): The fourth article targets achieving universal segmentation at any semantic level.
- [Compressing Language Models](https://github.com/HKUDS/SepLLM) (2025-07-17): The research indicates that data between separator tokens can be effectively compressed into the tokens themselves without significant loss of information.
- [Energy Transformers for Learning](https://github.com/alexiglad/EBT) (2025-07-17): Enhanced Bridge Transformers (EBTs) perform better than existing models in most tasks, suggesting superior generalization, even if pretraining performance is similar or worse.
- [Robust Stereo Matching with SMoEStereo](https://github.com/cocowy1/smoe-stereo) (2025-07-17): SMoEStereo is a new framework that adapts VFMs for stereo matching, combining LowRank Adaptation (LoRA) and MixtureofExperts (MoE) modules.
- [Evaluating Conversational Agents](https://github.com/sierra-research/tau2-bench) (2025-07-17): Current benchmarks for conversational AI agents only create scenarios where the AI interacts, while the user only provides information passively.
- [TCSinger 2: Zeroshot Singing Voice Synthesis](https://github.com/aaronz345/tcsinger2) (2025-07-17): Zeroshot Singing Voice Synthesis: TCSinger 2 is a new multitask, multilingual, zero-shot SVS model with style transfer and control, designed to address existing challenges.
- [PyTorch: Deep Learning Library](https://github.com/pytorch/pytorch) (2025-07-03): Deep Learning Library: The article explores the trade-off between user-friendliness and performance in deep learning platforms.
- [Absorbing States in Reinforcement Learning](https://github.com/Unity-Technologies/ml-agents) (2025-07-03): The research indicates that the complexity of supervised learning tasks grows with the number of absorbing states, but attention can adjust to different input sizes.
- [Quantized Training of Decision Trees](https://github.com/Microsoft/LightGBM) (2025-07-03): The article emphasizes the widespread success of Gradient Boosting Decision Trees in diverse machine learning tasks recently.
- [LLM Alignment Survey](https://github.com/bradyfu/awesome-multimodal-large-language-models) (2025-07-03): The paper offers an in-depth analysis of alignment algorithms used in Multi-Level Language Models (MLLMs).
- [Study of Mamba Language Models](https://github.com/NVIDIA/Megatron-LM) (2025-07-03): The hybrid model outperforms or equals the Transformer in 23 tasks involving long-context.
- [Hyperparameter Optimization Framework](https://github.com/optuna/optuna) (2025-07-03): The article outlines strategies for creating robust software that fulfills certain requirements.
- [Efficient Inference Engine](https://github.com/alibaba/MNN) (2025-07-03): The article explores the growing trend of implementing deep learning models on mobile devices.
- [RAGAs Framework](https://github.com/explodinggradients/ragas) (2025-07-03): The article presents RAGAs, a new framework for assessing Retrieval Augmented Generation pipelines without a reference.
- [GPT Bias Study](https://github.com/rfordatascience/tidytuesday) (2025-07-03): The study examines the effectiveness of different GPT detectors using English writing samples from both native and non-native speakers.
- [ML Serving Desiderata](https://github.com/SeldonIO/seldon-core) (2025-07-03): The article emphasizes the crucial role of inference in the infrastructure of machine learning software.
- [Enhanced minimap2 Alignment](https://github.com/lh3/minimap2) (2025-07-03): The paper highlights recent enhancements to minimap2, a flexible aligner for nucleotide sequences.
- [Hyperbolic Data Distillation](https://github.com/Guang000/Awesome-Dataset-Distillation) (2025-07-03): Dataset distillation condenses large datasets for deep learning without compromising model performance.
- [SymbolicAI Logic Framework](https://github.com/ExtensityAI/symbolicai) (2025-07-03): A new framework uses in-context learning to create and assess explainable computational graphs.
- [Deep Learning for SingleCell Analysis](https://github.com/scverse/scvi-tools) (2025-07-03): The article reviews the latest advancements in classical and deep learning methods, outlining their strengths and weaknesses.
- [Efficient Prediction of Equivariant Hamiltonian Matrices](https://github.com/divelab/AIRS) (2025-07-03): Efficient SO2-equivariant operations have been introduced, removing the need for SO3 tensor products and enhancing feature updates and message passing.
- [Experts Algorithms Theory](https://github.com/deepseek-ai/DeepEP) (2025-06-25): The article highlights the significant advancements in artificial intelligence, focusing on the progress of large foundational models.
- [Language Models and CDNs](https://github.com/lmcache/lmcache) (2025-06-25): The article underscores the increasing demand for comprehensive knowledge to back the fast-growing use of large language models.
- [Efficiency of Muon Pretraining](https://github.com/KellerJordan/Muon) (2025-06-25): The article demonstrates how the second-order optimizer, Muon, enhances the balance between computation and time compared to AdamW.
- [HighQuality Song Generation](https://github.com/tencent-ailab/songgeneration) (2025-06-25): The article presents a new technique using Direct Preference Optimization to enhance musicality and adherence to instructions.
- [AquaSignal: Underwater Analysis](https://github.com/codelion/adaptive-classifier) (2025-06-25): Underwater Analysis: AquaSignal is a scalable system designed to process and classify underwater acoustic signals, focusing on reducing noise and detecting new signals.
- [SoundMind: Audio Logic](https://github.com/xid32/soundmind) (2025-06-25): Audio Logic: The use of large language models in audio processing, particularly in large audio/language models, is significantly underexplored.
- [OPTBENCH: LLM Agent Evaluation](https://github.com/oliverleexz/opt-bench) (2025-06-25): LLM Agent Evaluation: Large Language Models have shown remarkable capabilities in handling a wide range of tasks.
- [LogitsBased OOD Detection](https://github.com/dvlab-research/logits-based-finetuning) (2025-06-25): The research thoroughly examines the contributors to Out-Of-Distribution (OOD) detection, emphasizing the potential of reconstruction-based tasks in understanding inherent data distributions.
- [Online Chinese Handwritten Writer Retrieval](https://github.com/scut-dlvclab/dolphin) (2025-06-25): The paper presents OLIWER, a large dataset of over 670,000 Chinese handwritten phrases from 1,731 individuals, to address the issue of data scarcity.
- [InfLLM v2: Longcontext Processing](https://github.com/openbmb/minicpm) (2025-06-18): Longcontext Processing: The article introduces InfLLM v2, a new model with a trainable sparse attention mechanism designed for quicker processing of long-context data.
- [TradingAgents: Financial Trading Framework](https://github.com/tauricresearch/tradingagents) (2025-06-18): Financial Trading Framework: The second article explores the progress in automated problem-solving using societies of agents driven by large language models (LLMs).
- [RWKV Goose: State Evolution](https://github.com/fla-org/flash-linear-attention) (2025-06-18): State Evolution: The third article presents RWKV7 Goose, a new sequence modeling architecture that ensures consistent memory usage and inference time per token.
- [TabM: Tabular DL Advancement](https://github.com/yandex-research/tabm) (2025-06-18): Tabular DL Advancement: The article explores different deep learning structures for managing and learning from structured data, including basic and advanced models like Transformers.
- [MEIA: Multimodal Perception](https://github.com/hcplab-sysu/causalvlr) (2025-06-18): Multimodal Perception: The article presents the Multimodal Embodied Interactive Agent (MEIA), a system that can convert complex tasks described in everyday language into a series of actionable steps.
- [SurveyForge: Auto Survey Writing](https://github.com/alpha-innovator/surveyforge) (2025-06-18): Auto Survey Writing: The article highlights the crucial role of review articles in scientific research, especially given the fast-paced increase in research publications.
- [PixelsDB Data Analytics](https://github.com/pixelsdb/pixels) (2025-06-18): The article explores a serverless query engine that performs queries and provides diverse pricing based on performance service levels.
- [RFUAV Dataset for UAV Detection](https://github.com/kitoweeknd/RFUAV) (2025-06-18): The article presents the RFUAV dataset, a baseline preprocessing method, and tools for model evaluation.
- [MASLab MultiAgent Systems Codebase](https://github.com/masworks/maslab) (2025-06-18): The article unveils MASLab, a complete codebase for LLM-based MAS designed to tackle specific challenges.
- [PreTraining Framework for Agentic Search](https://github.com/alibaba-nlp/masksearch) (2025-06-18): The Retrieval Augmented Mask Prediction (RAMP) task improves Large Language Models' retrieval and reasoning skills by teaching them to use search tools during the pretraining stage.
- [Spiking Graph Convolution Networks](https://github.com/zulunzhu/spikinggcn) (2025-06-18): Graph Convolutional Networks (GCNs) excel due to their superior capacity to learn and process graph information.
- [Evaluating Language Models in Games](https://github.com/fuxiailab/civagent) (2025-06-18): Large Language Models (LLMs) are evolving quickly, with potential applications as digital employees like analysts, teachers, and programmers.
- [Autonomous Knowledge Graph Construction](https://github.com/hkust-knowcomp/autoschemakg) (2025-06-18): AutoSchemaKG is a novel framework that enables the autonomous creation of knowledge graphs, removing the requirement for predefined schemas.
- [AutoAgent: Automated LLM Framework](https://github.com/hkuds/autoagent) (2025-06-11): Automated LLM Framework: AutoAgenta is a new framework that enables users to build and launch Language Learning Model agents using natural language only.
- [Sequential Models in Data Vault](https://github.com/sdv-dev/SDV) (2025-06-11): The Sequential Synthetic Data Vault (SDV) was utilized to produce synthetic data, and its quality was compared to a non-sequential model named CTGAN.
- [AReaL: RL System for Language Reasoning](https://github.com/inclusionai/areal) (2025-06-11): RL System for Language Reasoning: Majority of large-scale Reinforcement Learning systems for Language Learning Models alternate between generation and training in a batch setting, with each batch being generated by the same model.
- [Godel Machine](https://github.com/jennyzzt/dgm) (2025-06-11): The Godel machine is a self-enhancing AI that can make beneficial self-modifications.
- [SEW Code Generation](https://github.com/evoagentx/evoagentx) (2025-06-11): Large Language Models (LLMs) have proven effective in generating code.
- [RKV Cache Compression](https://github.com/zefan-cai/r-kv) (2025-06-11): The Redundancy-aware KV Cache Compression (RKV) technique focuses on redundant tokens in reasoning models.
- [MoonCast Generation](https://github.com/jzq2000/mooncast) (2025-06-11): Advances in text-to-speech synthesis have greatly enhanced the production of high-quality short utterances for specific speakers.
- [Enhancing Speech Extraction](https://github.com/wanghelin1997/solospeech) (2025-06-11): The article critiques the limitations of generative models for Time Series Embedding (TSE), particularly in relation to perceptual quality and comprehension.
- [Automated Failure Attribution](https://github.com/mingyin1/agents_failure_attribution) (2025-06-11): The paper presents a new research area centered on automated identification of failures in Large-scale Learning Multiagent (LLM) systems.
- [Incentivizing Biological Reasoning](https://github.com/bowang-lab/bioreason) (2025-06-11): The article emphasizes the challenges in deriving significant biological understanding from intricate genomic data through Artificial Intelligence.
- [Earth System Forecasting](https://github.com/easylearningscores/triton_ai4earth) (2025-06-11): Current AI models used for predicting Earth system dynamics are unstable during long-term simulations.
- [WINA for Language Models](https://github.com/microsoft/wina) (2025-06-11): The growing computational needs of large language models require efficient inference and activation strategies.
- [Efficient LongContext LLM Training](https://github.com/wenhaoli-xmu/seco) (2025-06-11): Despite their excellent document processing capabilities, the high training costs of long-context language models often limit their use in custom applications.
- [KVzip: Cache Compression](https://github.com/snu-mllab/kvzip) (2025-06-11): Cache Compression: Transformer-based language models store context as key-value pairs during the inference process.
- [sktime ROCKET](https://github.com/sktime/sktime) (2025-06-04): Current methods for classifying time series data are accurate but computationally complex, making them inefficient for large datasets.
- [microsoft RDAgentQuant](https://github.com/microsoft/rd-agent) (2025-06-04): Predicting asset returns is challenging due to the high dimensionality, nonstationarity, and persistent volatility of financial markets.
- [paper2poster](https://github.com/paper2poster/paper2poster) (2025-06-04): A new benchmark and metric suite for poster generation assesses visual quality, textual coherence, aesthetics, informational criteria, and the poster's effectiveness in conveying the main content of the paper.
- [Alita Generalist Agent](https://github.com/charlesq9/alita) (2025-06-04): The article explores how Alita uses general-purpose components to independently develop and improve its capabilities from open source, aiding in scalable decision-making.
- [CodeEnhanced Reasoning](https://github.com/dayuyang1999/awesome-code-reasoning) (2025-06-04): The article discusses how code and reasoning work together in large language models, with code providing a structure for reasoning and reasoning turning high-level objectives into executable steps, improving code intelligence.
- [Pareto-Optimal AI](https://github.com/datarobot/syftr) (2025-06-04): The article emphasizes the role of Retrieval-Augmented Generation pipelines in using large language models for proprietary or constantly changing data.
- [AudioTrust: Trustworthiness of ALLMs](https://github.com/jusperlee/audiotrust) (2025-06-04): Trustworthiness of ALLMs: AudioTrust is a new evaluation framework and benchmark specifically created for assessing the trustworthiness of Automatic Language Learning Models (ALLMs).
- [BizFinBench: Financial Benchmark for LLMs](https://github.com/hithink-research/bizfinbench) (2025-06-04): Financial Benchmark for LLMs: An assessment shows unique ability patterns in numerical calculation, reasoning, information extraction, and prediction recognition across different models.
- [Training a Search Agent via RL](https://github.com/pat-jj/s3) (2025-06-04): Retrieval-augmented generation (RAG) systems allow large language models (LLMs) to utilize external knowledge during the inference process.
- [ToolStar LLMBrained MultiTool](https://github.com/dongguanting/tool-star) (2025-06-04): The article explores the difficulties in applying the RL algorithm to improve multitool collaborative reasoning in large language models.
- [Phi Report](https://github.com/codelion/pts) (2025-06-04): The article presents phi4, a language model with 14 billion parameters that prioritizes data quality during training.
- [Oneshot Entropy Minimization](https://github.com/zitian-gao/one-shot-em) (2025-06-04): The article shows that training over 13,000 large language models with entropy minimization can significantly improve performance using only one unlabeled data and a 10-step optimization.
- [ChartGalaxy Dataset](https://github.com/chartgalaxy/chartgalaxy) (2025-06-04): The article highlights the importance of a dataset in enhancing infographic chart comprehension, setting code generation standards, and facilitating example-based infographic chart creation.
- [Building AI Agents](https://github.com/mem0ai/mem0) (2025-05-21): The article explores the difficulties in ensuring consistency in extended dialogues due to the limited context windows of Large Language Models.
- [Evaluating Language Models in Healthcare](https://github.com/openai/simple-evals) (2025-05-21): The article presents HealthBench, a freely available benchmark for evaluating the efficiency and safety of large language models in the healthcare sector.
- [Open Source Moxin7B Report](https://github.com/moxin-org/moxin-llm) (2025-05-21): The article emphasizes the recent major changes and increasing popularity of Large Language Models.
- [Unified Multimodal Models](https://github.com/jiuhaichen/blip3o) (2025-05-21): The team has created BLIP3o, a set of sophisticated multimodal models, through innovative training methods and datasets.
- [Humanlike Episodic Memory](https://github.com/em-llm/EM-LLM-model) (2025-05-21): Large language models (LLMs) show great potential but struggle with handling extensive contexts, which impacts their consistency and precision over lengthy sequences.
- [Scaling Atomistic Potential](https://github.com/zmyybc/alphanet) (2025-05-21): Molecular dynamics simulations need a special combination of precision and scalability to tackle major issues in catalysis and materials design.
- [Novel Alphas in Quantitative Investment](https://github.com/codelion/openevolve) (2025-05-21): The paper presents a new type of alphas for improving the modeling of scalar vector and matrix features.
- [MultiAgent Simulation Scaling for Portfolio Construction](https://github.com/gta0804/mass) (2025-05-21): The article discusses the growing interest in LLM-based multiagent systems for their potential in simulation and performance improvement.
- [Parallel Scaling Law for Language Models](https://github.com/qwenlm/parscale) (2025-05-21): The research uses P diverse transformations on the input, runs the model's forward passes simultaneously, and dynamically combines the P outputs.
- [Novel Alphas in Quantitative Investment](https://github.com/qwenlm/worldpm) (2025-05-21): The article investigates the presence of scaling laws in preference modeling, comparing them to those in language modeling, and how they relate to model and dataset sizes.
- [MultiAgent Simulation Scaling for Portfolio Construction](https://github.com/vainf/thinkless) (2025-05-21): The article examines the high performance of Reasoning Language Models that are capable of performing complex logical reasoning.
- [Parallel Scaling Law for Language Models](https://github.com/Pints-AI/Finetune-Bench-RAG) (2025-05-21): The article emphasizes the success of Retrieval-Augmented Generation (RAG) in improving the accuracy of large language models by using information from retrieved documents.
- [Microsoft BitNet](https://github.com/microsoft/bitnet) (2025-04-23): The creation of large language models such as BitNet b1.58 has increased interest in ternary language models.
- [Reinforcement Learning](https://github.com/natolambert/rlhf-book) (2025-04-23): The use of human feedback in reinforcement learning is now essential for developing advanced machine learning systems.
- [The AI Scientistv2](https://github.com/sakanaai/ai-scientist-v2) (2025-04-23): AI is playing a significant role in transforming the methods of scientific discoveries.
- [Tencent InstantCharacter Framework](https://github.com/tencent/instantcharacter) (2025-04-23): The article details the development of a massive character dataset with over 10 million samples for efficient framework training.
- [Lizonghang PRIMA: Breaking Barriers for LLMs](https://github.com/lizonghang/prima.cpp) (2025-04-23): Breaking Barriers for LLMs: The article emphasizes the exceptional performance of DeepSeek R1 and QwQ 32B in operating large language models on personal devices.
- [KRLabsOrg LettuceDetect: Hallucination Detection](https://github.com/KRLabsOrg/LettuceDetect) (2025-04-23): Hallucination Detection: The article discloses that despite utilizing external knowledge sources, Retrieval Augmented Generation (RAG) systems can still produce hallucinated answers.
- [SimpleAR: Visual Generation Framework](https://github.com/wdrink/simplear) (2025-04-23): Visual Generation Framework: The article presents SimpleAR, a straightforward visual generation framework that doesn't require complex architectural modifications.
- [Syzygy of Thoughts: Improving CoT](https://github.com/dlmaria/syzygy-of-thoughts) (2025-04-23): Improving CoT: The piece explores ChainofThought (CoT) prompting, a method that enhances the reasoning of large language models by simplifying problems into sequential steps, thereby minimizing errors.
- [VSLAMLAB: Visual SLAM Framework](https://github.com/alejandrofontan/vslam-lab) (2025-04-23): Visual SLAM Framework: The text underscores the difficulties in Visual Simultaneous Localization and Mapping (VSLAM) research, caused by disjointed toolchains, intricate system setups, and inconsistent assessment techniques.
- [DeepMath103K](https://github.com/zwhe99/deepmath) (2025-04-23): The article emphasizes the role of complex mathematical reasoning in enhancing artificial intelligence.
- [EndEndDiffusion](https://github.com/End2End-Diffusion/REPA-E) (2025-04-23): The article introduces a technique to enhance AI model training through representationalignment (REPA) loss.
- [rlhflow](https://github.com/rlhflow/minimal-rl) (2025-04-23): The article provides an in-depth analysis of the fundamental elements of GRPO from the viewpoint of reinforcement algorithms.
- [spcl](https://github.com/spcl/knowledge-graph-of-thoughts) (2025-04-23): The article underscores the significance of structured task-relevant knowledge in assisting low-cost models to solve intricate tasks.
- [Foundation Agents](https://github.com/foundationagents/awesome-foundation-agents) (2025-04-16): Large Language Models (LLMs) have significantly improved AI capabilities, enhancing reasoning, perception, and action in diverse areas.
- [More Generalization](https://github.com/bytedance/uno) (2025-04-16): The article presents a consistent data synthesis pipeline designed to tackle complex problems.
- [Language Model Quantization](https://github.com/hanguo97/flute) (2025-04-16): Quantizing large language models is a prevalent technique used to reduce their memory and computational costs.
- [LocAgent GraphGuided LLM Agents](https://github.com/gersteinlab/locagent) (2025-04-16): LocAgent improves codebase search by using directed heterogeneous graphs for a lightweight representation.
- [NdLinear Representation Learning](https://github.com/ensemble-core/ndlinear) (2025-04-16): NdLinear is suggested as an advanced alternative to standard linear layers for next-gen neural architectures.
- [MultiSWEbench Issue Resolving Benchmark](https://github.com/multi-swe-bench/multi-swe-bench) (2025-04-16): Issue resolving involves altering a codebase to create a patch that fixes a specific problem.
- [Generalist-Specialist Framework for Computer Agents](https://github.com/simular-ai/agent-s) (2025-04-09): Computer use agents are created to enhance human efficiency by automating digital tasks via direct interaction with user interfaces on computers and mobile devices.
- [Automatic Speech Recognition Model for Eastern Languages](https://github.com/dataoceanai/dolphin) (2025-04-09): The report introduces Dolphin, a large-scale multilingual automatic speech recognition model that enhances the Whisper architecture to support additional languages.
- [Benchmark Dataset & Debiasing Framework for Multimodal Reasoning](https://github.com/reml-group/fortisavqa) (2025-04-09): The article presents a two-step method to increase test diversity and improve robustness evaluation in question distributions, providing the corresponding code on GitHub.
- [Efficient Reasoning for Large Language Models](https://github.com/devoallen/awesome-reasoning-economy-papers) (2025-04-09): The article emphasizes the recent progress in Large Language Models (LLMs) that have enhanced their complex reasoning abilities, with the related code shared on GitHub.
- [Efficient Inference for Large Reasoning Models](https://github.com/yueliu1999/awesome-efficient-inference-for-lrms) (2025-04-09): The article details how Large Reasoning Models (LRMs) boost the reasoning capacity of Large Language Models (LLMs), demonstrating promising outcomes in complex tasks, with the associated code on GitHub.
- [LLMAggreFact Code Release](https://github.com/vectara/hallucination-leaderboard) (2025-04-02): The LLMAggreFact code, used for data synthesis and modeling, is now available for public use.
- [Rankify Python Toolkit](https://github.com/DataScienceUIBK/Rankify) (2025-04-02): The importance of retrieval reranking and retrieval-augmented generation (RAG) in information retrieval and question answering applications is highlighted.
- [Open Search Tool](https://github.com/sentient-agi/opendeepsearch) (2025-04-02): Open Search Tool, a new internet search tool, has outperformed its proprietary competitors.
- [FEC Fast Euclidean Clustering](https://github.com/yizhenlao/fec) (2025-04-02): The segmentation of point cloud data is vital for applications such as remote sensing, mobile robots, and autonomous vehicles.
- [Reverse Dialogue Inference](https://github.com/zhengl00/rex-got) (2025-04-02): The DCMCQ task is a new tool designed to better interpret and respond to user queries within the vast amount of online dialogic data.
- [RL Enhanced LLMs Survey](https://github.com/shuhewang1998/reinforcement-learning-enhanced-llms-a-survey) (2025-04-02): The paper explores the use of reinforcement learning to enhance the performance of large language models by providing output-based feedback.
- [LLM Agent Methodology Survey](https://github.com/luo-junyu/awesome-agent-papers) (2025-04-02): The article examines the emergence of intelligent agents, which are driven by advancements in large language models.
- [Optimal Stepsize for Sampling](https://github.com/bebebe666/optimalsteps) (2025-04-02): The article emphasizes the high-quality output of diffusion models, but points out their computational intensity due to inefficient step discretization.
- [CPPO Training Acceleration](https://github.com/lzhxmu/cppo) (2025-04-02): The paper introduces Completion Pruning Policy Optimization (CPPO), a method aimed at accelerating the training of reasoning models based on Group Relative Policy Optimization (GRPO).
- [Augmenting Language Models with Knowledge Base](https://github.com/microsoft/KBLaM) (2025-03-26): The paper presents KBLaM, a novel technique to improve Large Language Models using external knowledge.
- [Pose Diversity in Audiodriven Talking Portraits](https://github.com/chaolongy/kdtalker) (2025-03-26): The article explores the shortcomings of Keypoint-based methods in accurately capturing detailed facial features due to the restrictions of the 3D Morphable Model.
- [Multimodal ChainofThought Reasoning Survey](https://github.com/yaotingwangofficial/awesome-mcot) (2025-03-26): The study expands the application of chain-of-thought reasoning to multimodal contexts, underlining its growing prominence in research, particularly when combined with multimodal large language models.
- [Physical Common Sense to Embodied Reasoning](https://github.com/nvidia-cosmos/cosmos-reason1) (2025-03-26): The article details essential skills for Physical AI reasoning, focusing on physical common sense and embodied reasoning.
- [D Task Point Representations](https://github.com/facebookresearch/sonata) (2025-03-26): The paper discusses the effectiveness of a self-supervised point cloud model in performing various 3D tasks with limited data and minimal computation.
- [Large Language Model Reasoning](https://github.com/eclipsess/awesome-efficient-reasoning-llms) (2025-03-26): Large Language Models (LLMs) are highlighted for their impressive capabilities in managing complex tasks.
- [SelfSupervised Learning in Molecular Graphs](https://github.com/hansen7/molgrapheval) (2025-03-26): Graph Self-Supervised Learning (GSSL) is presented as a powerful tool for generating embeddings without the need for expert labelling, particularly useful in molecular graphs.
- [Reinforcement Learning in Small LLMs](https://github.com/knoveleng/open-rs) (2025-03-26): The article discusses the challenges of enhancing the reasoning abilities of Large Language Models (LLMs), especially in situations with limited resources and extensive datasets.
- [Single-View Scene Reconstruction with Geometric Affine Field](https://github.com/xianzuwu/Niagara) (2025-03-26): The article highlights the challenges in single-view 3D scene reconstruction, particularly in capturing fine geometric details and maintaining structural consistency in high-quality outdoor scene modeling.
- [Optimizing Instruction in Large Language Models](https://github.com/emcie-co/parlant) (2025-03-20): The article presents Attentive Reasoning Queries (ARQs), a new method that improves instruction-following in Large Language Models through the use of specific reasoning blueprints.
- [Training LLMs to Reason with Search Engines](https://github.com/petergriffinjin/search-r1) (2025-03-20): The article highlights the need for large language models to effectively gather external knowledge and up-to-date information for better reasoning and text generation.
- [Empowering LMMs with Strong Reasoning](https://github.com/tidedra/lmm-r1) (2025-03-20): The article explores the difficulties in enhancing reasoning in Large Multimodal Models due to the intricate relationship between visual perception and logical reasoning, particularly in compact 3B-parameter structures.
- [qihoo360 LightR1](https://github.com/qihoo360/light-r1) (2025-03-20): The LightR1 series proves the efficiency of training longCOT models from scratch, showcasing excellence in SFT data and releasing top-notch models from RL.
- [generteam GENERator](https://github.com/generteam/generator) (2025-03-20): Recent progress in genomic language models underscores their capability in decoding DNA sequences.
- [prseth Neural Fields](https://github.com/prs-eth/thera) (2025-03-20): A novel technique is presented for creating neural fields, enabling points to be queried with an adjustable Gaussian PSF, guaranteeing accurate antialiasing at any chosen output resolution.
- [Block Diffusion Interpolating Models](https://github.com/kuleshov-group/bd3lms) (2025-03-20): Diffusion language models offer advantages such as parallelized generation and controllability, but face challenges with likelihood modeling and are limited to fixed-length generation.
- [ReasonGraph Reasoning Paths Visualization](https://github.com/ZongqianLi/ReasonGraph) (2025-03-20): The complexity of Large Language Models (LLMs) and lack of effective visualization tools make their reasoning processes hard to analyze.
- [LLMAD: Algorithm Design Platform](https://github.com/optima-cityu/llm4ad) (2025-03-20): Algorithm Design Platform: LLM4AD is a newly developed Python platform specifically designed for algorithm design using large language models.
- [ELIZA Chatbot on Time Sharing System](https://github.com/rupertl/eliza-ctss) (2025-03-20): The world's first chatbot's entire stack, compatible with Unix-like operating systems, is now open source and was developed on the first timesharing system.
- [Reinforcement Learning Beats Fine-Tuning](https://github.com/xiaomi-research/r1-aqa) (2025-03-20): Reinforcement learning has been shown to improve the reasoning abilities of large language models and is increasingly used in visual multimodal tasks.
- [OLMo 2 Model Architecture Recipe](https://github.com/allenai/OLMo-core) (2025-03-20): The updated model architecture and training recipe provide enhanced training stability and per-token efficiency.
- [Slim Attention: Reduce Context Memory](https://github.com/openmachine-ai/transformer-tricks) (2025-03-20): Reduce Context Memory: For Whisper models and similar encoder-decoder transformers, slim attention reduces context memory by 8x, accelerating token generation by 5x for a batch size of 64.
- [Efficient Text-to-Speech Model](https://github.com/sparkaudio/spark-tts) (2025-03-12): Recent advancements in large language models have greatly improved the ability to synthesize speech without prior training.
- [Zeroshot Voice Conversion](https://github.com/Plachtaa/seed-vc) (2025-03-12): Zero-shot voice conversion aims to modify a source speech to imitate the voice of an unfamiliar speaker.
- [Visual FineTuning](https://github.com/liuziyu77/visual-rft) (2025-03-12): Reinforcement FineTuning in Large Reasoning Models like OpenAI o1 is useful in scenarios where there is a scarcity of data for fine-tuning.
- [DeepRetrieval: Query Generation](https://github.com/pat-jj/deepretrieval) (2025-03-12): Query Generation: The article emphasizes the role of information retrieval systems in managing and accessing large document collections.
- [Draft Thinking Chain](https://github.com/sileix/chain-of-draft) (2025-03-12): The article examines the effectiveness of Large Language Models in solving complex tasks through ChainofThought prompting.
- [MultiAgentBench: LLM Agent Evaluation](https://github.com/multiagentbench/marble) (2025-03-12): LLM Agent Evaluation: The article points out the shortcomings of current benchmarks in evaluating the performance of Large Language Models in multiagent coordination and competition.
- [Merlion: Time Series Library](https://github.com/salesforce/merlion) (2025-03-05): Time Series Library: The article introduces Merlion, a new open-source library for machine learning in time series analysis.
- [Sentientagi: Loyalty Training](https://github.com/sentient-agi/oml-1.0-fingerprinting) (2025-03-05): Loyalty Training: The piece discusses the challenges in developing open-source models that are accessible and governed by the community.
- [HybridFlow: RLHF Framework](https://github.com/hiyouga/easyr1) (2025-03-05): RLHF Framework: The article describes how traditional Reinforcement Learning (RL) can be represented as a dataflow with nodes and edges symbolizing neural network computations and data dependencies respectively.
- [Atom of Thoughts: Markov LLM Scaling](https://github.com/qixucen/atom) (2025-03-05): Markov LLM Scaling: The Atom of Thoughts (AoT) model is introduced, which breaks down each reasoning state transition into a dependency-based directed acyclic graph.
- [IndexTTS: ZeroShot TextToSpeech System](https://github.com/index-tts/index-tts) (2025-03-05): ZeroShot TextToSpeech System: The article presents IndexTTS, a large language model-based text-to-speech system known for its natural sound and zero-shot voice cloning features.
- [Sparse Attention](https://github.com/thu-ml/spargeattn) (2025-03-05): The first article explores the difficulties in creating a universal sparse attention that improves the speed and efficiency of different models.
- [Adversarial Training](https://github.com/nisplab/at-bsl) (2025-03-05): The second article demonstrates through experiments that data augmentation can significantly improve robustness.
- [Depth Distillation](https://github.com/Westlake-AGI-Lab/Distill-Any-Depth) (2025-03-05): The third article conducts a thorough examination of how different depth normalization strategies impact pseudolabel distillation.
- [AudioFLAN: Audiolanguage Models](https://github.com/lmxue/audio-flan) (2025-03-05): Audiolanguage Models: AudioFLAN is a versatile audio language model capable of comprehension and generation tasks across different audio domains without needing specific training.
- [BaichuanAudio: Speech Interaction Framework](https://github.com/baichuan-inc/baichuan-audio) (2025-03-05): Speech Interaction Framework: A proposed two-stage pretraining strategy aims to enhance audio modeling and preserve language understanding, avoiding intelligence loss during pretraining.
- [Emergent Misalignment: Finetuning Effects](https://github.com/emergent-misalignment/emergent-misalignment) (2025-03-05): Finetuning Effects: A study was carried out where a model was discreetly manipulated to generate insecure code unbeknownst to the user.
- [Test Time Scaling](https://github.com/novasky-ai/skythought) (2025-02-26): The possibility of enhancing test-time computation for Large Language Models in code generation is promising but needs more research, despite extensive studies in mathematics.
- [StepAudio Understanding](https://github.com/stepfun-ai/step-audio) (2025-02-26): The StepEvalAudio360 benchmark reveals that StepAudio performs exceptionally well in human evaluations, especially in instruction following.
- [MoBA Block Attention](https://github.com/moonshotai/moba) (2025-02-26): Enhancing the effective context length is vital for the progress of Large Language Models towards achieving Artificial General Intelligence.
- [SWELancer: RealWorld Freelance Software Engineering](https://github.com/openai/swelancer-benchmark) (2025-02-26): RealWorld Freelance Software Engineering: The article introduces SWELancer, a benchmark comprising over 1,400 freelance software engineering tasks from Upwork, valued at 1 million USD.
- [Magma: Multimodal AI Agents Foundation Model](https://github.com/microsoft/Magma) (2025-02-26): Multimodal AI Agents Foundation Model: The article introduces Magma, a foundational model designed for multimodal AI tasks in both digital and physical environments.
- [CrawLLM: Web Crawling for LLM Pretraining](https://github.com/cxcscmu/crawl4llm) (2025-02-26): Web Crawling for LLM Pretraining: The article reveals that a majority of web pages crawled for pretraining large language models are discarded due to substandard data quality.
- [Modular Generative Models](https://github.com/LTH14/fractalgen) (2025-02-26): The article introduces a novel approach to modularization by simplifying generative models into basic generative modules.
- [SongGen: Text-to-Song Transformer](https://github.com/liuzh-19/songgen) (2025-02-26): Text-to-Song Transformer: The authors aim to share their model weights, training code, annotated data, and preprocessing pipeline to promote community involvement and additional research.
- [Slamming Training: Speech Model on One GPU](https://github.com/slp-rl/slamkit) (2025-02-26): Speech Model on One GPU: The research demonstrates that their training method efficiently scales with enhanced computing power, delivering results similar to top SLMs but at a reduced computational expense.
- [RiskSensitive RL for Trading Agents](https://github.com/benstaf/finrl_deepseek) (2025-02-26): The article presents a novel trading agent that combines reinforcement learning with large language models to manage risk-sensitive trading.
- [Economical Inference with Multi-Head Attention](https://github.com/JT-Ushio/MHA2MLA) (2025-02-26): The paper discusses how Llama27B's KV cache size was reduced by 92.19%, with only a slight 0.5% drop in LongBench performance.
- [Optimizing Model Selection for Compound AI Systems](https://github.com/LLMSELECTOR/LLMSELECTOR) (2025-02-26): The piece introduces LLMSelector, a system for choosing models in compound systems, utilizing insights about end-to-end and per-module performance estimated by a large language model.
- [MatterGen: Generative Model for Materials Design](https://github.com/microsoft/mattergen) (2025-02-19): Generative Model for Materials Design: The article presents adapter modules that help fine-tune property constraints using a labeled dataset.
- [LIMO: Less is More for Reasoning](https://github.com/gair-nlp/limo) (2025-02-19): Less is More for Reasoning: The article shows that complex mathematical reasoning can be achieved with a minimal number of examples.
- [SEALRG: Scaling TestTime Compute with Latent Reasoning](https://github.com/seal-rg/recurrent-pretraining) (2025-02-19): Scaling TestTime Compute with Latent Reasoning: The article talks about the expansion of a prototype model to 3.5 billion parameters and 800 billion tokens.
- [Meta Audiobox: Unified Quality Assessment for Audio](https://github.com/facebookresearch/audiobox-aesthetics) (2025-02-19): Unified Quality Assessment for Audio: The article investigates the difficulty of measuring audio aesthetics in audio processing due to its subjective nature.
- [CodeIO: Reasoning Patterns](https://github.com/hkust-nlp/codeio) (2025-02-19): Reasoning Patterns: The article discusses the inherent reasoning capabilities of Large Language Models.
- [PIKERAG: Knowledge Generation](https://github.com/microsoft/pike-rag) (2025-02-19): Knowledge Generation: The piece highlights the limitations of Retrieval-Augmented Generation systems in meeting the varied needs of industrial applications.
- [Temporal Memory for Understanding](https://github.com/xid32/naacl_2025_twm) (2025-02-19): The article introduces Temporal Working Memory, a cognitive module designed to enhance the temporal modeling abilities of MFMs.
- [Agentic Reasoning for LLMs](https://github.com/theworldofagents/agentic-reasoning) (2025-02-19): The article presents Agentic Reasoning, a new framework that incorporates external tool-using agents to improve the reasoning of Large Language Models.
- [Qwen Turbo Models on Alibaba Cloud](https://github.com/qwenlm/qwen2.5) (2025-02-05): Alibaba Cloud Model Studio provides two exclusive models, Qwen2.5Turbo and Qwen2.5Plus, for its hosted solutions.
- [JanusPro Unified Multimodal Generation](https://github.com/deepseek-ai/janus) (2025-02-05): This work introduces JanusPro, an upgraded version of the previous Janus model.
- [Tulu 3: Advancing Language Model PostTraining](https://github.com/allenai/open-instruct) (2025-02-05): Advancing Language Model PostTraining: Posttraining refines behaviors and develops new skills in recent language models, but open-source methods are not as advanced as proprietary ones.
- [Test Scaling](https://github.com/simplescaling/s1) (2025-02-05): The Qwen2.532BInstruct language model has been enhanced with budget forcing, leading to a 27% improvement in solving competition math questions.
- [Specialized Tuning](https://github.com/deepseek-ai/esft) (2025-02-05): The study explores the PEFT method in LLMs with MixtureofExperts architecture, revealing that routing distribution is highly concentrated for specific tasks, but the activated experts' distribution varies greatly across tasks.
- [Graph Augmented Generation](https://github.com/deep-polyu/awesome-graphrag) (2025-02-05): The use of large language models in specialized fields is difficult due to the requirement for profound expertise, despite their remarkable abilities.
- [Selfadaptive LLMs](https://github.com/SakanaAI/self-adaptive-llms) (2025-01-23): The article discusses self-adaptive large language models (LLMs) which are designed to tackle the issues of high computational intensity and inflexibility in managing diverse tasks, common problems with traditional fine-tuning methods.
- [Engaging Learning with Language Models](https://github.com/stanford-oval/storm) (2025-01-08): Language model chatbots and generative search engines excel at answering specific queries but struggle to uncover unknown information.
- [TorchTitan: LLM Pretraining with PyTorch](https://github.com/pytorch/torchtitan) (2025-01-08): LLM Pretraining with PyTorch: Significant accelerations were achieved at various GPU scales by stacking training optimizations, proving the effectiveness of parallelism.
- [VITAMLLM: Interaction with Language Models](https://github.com/VITA-MLLM/VITA) (2025-01-08): Interaction with Language Models: Recent Multimodal Large Language Models have mainly focused on integrating visual and text modalities, overlooking the importance of speech in enhancing interaction.
- [LatentSync: Diffusion Models for Lip Sync](https://github.com/bytedance/LatentSync) (2025-01-08): Diffusion Models for Lip Sync: The experience with SyncNet can be applied to other lip sync and audio-driven portrait animation methods without altering the overall training framework.
- [Memory Layers: Parallelizable Implementation](https://github.com/facebookresearch/memory) (2025-01-08): Parallelizable Implementation: A fully parallelizable memory layer implementation is offered, demonstrating scaling laws with up to 128B memory parameters, pre-trained to 1 trillion tokens.
- [Next Token Prediction Survey](https://github.com/lmm101/awesome-multimodal-next-token-prediction) (2025-01-08): Recent studies show that Large Language Models (LLMs) can efficiently process tasks from various modalities by converting multimodal information into tokens to predict the next one based on context.
- [TangoFlux Text to Audio Generation](https://github.com/declare-lab/TangoFlux) (2025-01-08): TangoFlux, a Text-to-Audio generative model with 515M parameters, can produce up to 30 seconds of 44.1kHz audio in just 3.7 seconds using a single A40 GPU.
- [MMLUCF Multitask Language Benchmark](https://github.com/microsoft/mmlu-cf) (2025-01-08): A new benchmark aims to improve the understanding of world knowledge by Large Language Models (LLMs) to prevent both unintentional and malicious data leakage.
- [Selfsupervised Hierarchical Makeup Transfer](https://github.com/snowfallingplum/shmt) (2025-01-08): The first article presents a new technique called Self-supervised Hierarchical Makeup Transfer (SHMT) that improves makeup transfer using latent diffusion models.
- [CacheAugmented Generation for Knowledge Tasks](https://github.com/hhhuang/cag) (2025-01-08): The second article introduces a new concept called cache-augmented generation (CAG) that uses the extended context windows of large language models (LLMs) to avoid real-time retrieval.
- [Building Mixture of Experts](https://github.com/BenChaliah/Superposition-Transformer) (2025-01-08): The third article highlights the major issue of catastrophic forgetting when adjusting large language models (LLMs) to new tasks or domains.
- [DeepSeekV3 Report](https://github.com/deepseek-ai/deepseek-v3) (2025-01-01): DeepSeekV3 is a robust language model with 671B parameters, 37B of which are used for each token.