Changes

Every new paper, count, summary and review, kept as history.

New countsNov 9, 2026Language Models are Few-Shot LearnersInfluential citations rose from 5,611 to 5,637, and citations rose from 65,934 to 66,355.

And 218 more papers with new counts

LoRA: Low-Rank Adaptation of Large Language Models (Microsoft). Influential citations rose from 3,500 to 3,547, and citations rose from 25,222 to 25,555.
The Llama 3 Herd of Models (Meta). Influential citations rose from 3,731 to 3,790, and citations rose from 20,541 to 20,834.
Efficient Memory Management for Large Language Model Serving with PagedAttention (UC Berkeley). Influential citations rose from 1,331 to 1,375, and citations rose from 9,683 to 9,965.
Qwen3 Technical Report (Alibaba (Qwen)). Influential citations rose from 1,433 to 1,469, and citations rose from 9,838 to 10,089.
Training language models to follow instructions with human feedback (OpenAI). Influential citations rose from 2,541 to 2,553, and citations rose from 25,614 to 25,810.
GPT-4 Technical Report (OpenAI). Influential citations rose from 3,383 to 3,397, and citations rose from 28,214 to 28,363.
Direct Preference Optimization: Your Language Model is Secretly a Reward Model (Stanford University). Hugging Face upvotes rose from 71 to 73, and influential citations rose from 2,369 to 2,389, and citations rose from 11,437 to 11,573.
Segment Anything (Meta). Influential citations rose from 2,186 to 2,203, and citations rose from 16,655 to 16,793.
Mamba: Linear-Time Sequence Modeling with Selective State Spaces (Carnegie Mellon University). Influential citations rose from 1,397 to 1,413, and citations rose from 10,017 to 10,151.
Qwen3-VL Technical Report (Alibaba (Qwen)). Influential citations rose from 632 to 659, and citations rose from 3,066 to 3,178.
LLaMA: Open and Efficient Foundation Language Models (Meta). Influential citations rose from 2,297 to 2,311, and citations rose from 22,457 to 22,575.
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities (Google). Influential citations rose from 1,165 to 1,191, and citations rose from 4,523 to 4,594.
SAM 3: Segment Anything with Concepts (Meta). Influential citations rose from 223 to 237, and citations rose from 1,386 to 1,466.
FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (Stanford University). Influential citations rose from 482 to 492, and citations rose from 5,749 to 5,830.
From Local to Global: A Graph RAG Approach to Query-Focused Summarization (Microsoft). Influential citations rose from 346 to 351, and citations rose from 2,488 to 2,561.
GPT-4o System Card (OpenAI). Influential citations rose from 1,042 to 1,051, and citations rose from 5,328 to 5,388.
DINOv3 (Meta). Hugging Face upvotes rose from 311 to 320, and influential citations rose from 336 to 344, and citations rose from 1,883 to 1,932.
Scaling Laws for Neural Language Models (OpenAI). Influential citations fell from 652 to 649, and citations rose from 9,533 to 9,592.
Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism (NVIDIA). Influential citations rose from 411 to 417, and citations rose from 3,352 to 3,387.
Gemma 4 Technical Report (Google DeepMind). Influential citations rose from 43 to 48, and citations rose from 321 to 356.
Agent World Model in the Wild: Computer-Use Agent Environment Synthesis (Microsoft). Hugging Face upvotes fell from 100 to 69.
Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context (Google DeepMind). Influential citations rose from 887 to 888, and citations rose from 4,162 to 4,190.
Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone (Microsoft). Influential citations rose from 375 to 379, and citations rose from 2,628 to 2,649.
The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (GitHub). Influential citations fell from 123 to 121, and citations rose from 885 to 906.
Nemotron-Labs-Diffusion-Uni: Elastic Self-Refining Unified Diffusion LLM with Cross-Paradigm On-Policy Distillation (NVIDIA). Hugging Face upvotes fell from 31 to 12.
ASPIRE: Agentic /Skills Discovery for Robotics (NVIDIA). Influential citations rose from 16 to 18, and citations rose from 97 to 114.
DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation (DeepSeek). Influential citations rose from 32 to 38, and citations rose from 64 to 77.
Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models (Cohere). Citations rose from 399 to 418.
dLLM: Simple Diffusion Language Modeling (UC Berkeley). Hugging Face upvotes fell from 153 to 136, and citations rose from 17 to 18.
SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks? (Scale AI). Influential citations rose from 70 to 73, and citations rose from 352 to 367.
On-Policy Self-Training: Data-Free Self-Improvement in Large Language Models (Google). Hugging Face upvotes fell from 48 to 31.
StreamVibe: Proactive Video Agents for Evolving User Intent (Google). Hugging Face upvotes fell from 95 to 78.
Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking (Alibaba (Qwen)). Citations rose from 299 to 315.
Retrieval Augmented Generation Compositionality (IBM). Hugging Face upvotes fell from 49 to 34.
SAM 3D: 3Dfy Anything in Images (Meta). Influential citations rose from 77 to 81, and citations rose from 332 to 343.
Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains (Scale AI). Influential citations rose from 68 to 70, and citations rose from 418 to 431.
Group Sequence Policy Optimization (Alibaba (Qwen)). Influential citations rose from 78 to 80, and citations rose from 802 to 814.
MolmoAct2: Action Reasoning Models for Real-world Deployment (Ai2). Influential citations rose from 11 to 12, and citations rose from 87 to 100.
CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation (Alibaba (Qwen)). Hugging Face upvotes fell from 38 to 25.
Eat Your Precision: Precision Objective Mitigates Reward Hacking in Rubric-based RL (Carnegie Mellon University). Hugging Face upvotes fell from 47 to 34.
On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability (Alibaba (Qwen)). Influential citations rose from 11 to 13, and citations rose from 65 to 76.
DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos (NVIDIA). Influential citations rose from 25 to 26, and citations rose from 168 to 179.
BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References (Alibaba (Qwen)). Hugging Face upvotes fell from 34 to 23.
OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory (Apple). Hugging Face upvotes fell from 112 to 101.
World-in-Context: Interactive Video World Modeling with In-Context Conditioning (Alibaba (Qwen)). Hugging Face upvotes fell from 33 to 22.
Qwen-Image-PE: Learning Visual Design for Text-to-Image Prompt Enhancement (Alibaba (Qwen)). Hugging Face upvotes fell from 33 to 23.
InteractiveWorld: A Verifiable Human-Centric Environment for Building Better Interactive MLLM Assistants (Meta). Hugging Face upvotes fell from 56 to 46.
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Anthropic). Citations rose from 647 to 657.
Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild? (Anthropic). Hugging Face upvotes fell from 27 to 18.
UserArena: Benchmarking Interactive User Simulation for Long-Horizon Shopping Agents (Alibaba (Qwen)). Hugging Face upvotes fell from 64 to 55.
ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning (NVIDIA). Hugging Face upvotes fell from 66 to 57.
Qwen3-ASR Technical Report (Alibaba (Qwen)). Influential citations rose from 14 to 17, and citations rose from 194 to 200.
GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization (NVIDIA). Influential citations rose from 47 to 49, and citations rose from 227 to 234.
Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention (NVIDIA). Influential citations rose from 7 to 9, and citations rose from 42 to 48.
Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning (NVIDIA). Influential citations rose from 7 to 9, and citations rose from 48 to 54.
WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks? (ServiceNow). Citations rose from 435 to 443.
The BrowserGym Ecosystem for Web Agent Research (ServiceNow). Influential citations fell from 18 to 17, and citations rose from 193 to 200.
Productivity Assessment of Neural Code Completion (GitHub). Influential citations fell from 36 to 35, and citations rose from 342 to 349.
Beyond Watching and Searching: Benchmarking Agents for Video-centered Deep Research Reports (Alibaba (Qwen)). Hugging Face upvotes fell from 26 to 19.
Playful Agentic Robot Learning (UC Berkeley). Influential citations rose from 3 to 5, and citations rose from 32 to 37.
LoRA Learns Less and Forgets Less (Databricks). Citations rose from 479 to 486.
Global MMLU: Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation (Cohere). Influential citations rose from 31 to 33, and citations rose from 231 to 236.
Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models (Alibaba (Qwen)). Citations rose from 66 to 72.
WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces (Microsoft). Influential citations rose from 5 to 8, and citations rose from 12 to 15.
SkillOpt: Executive Strategy for Self-Evolving Agent Skills (Microsoft). Citations rose from 167 to 173.
mHC: Manifold-Constrained Hyper-Connections (DeepSeek). Influential citations rose from 22 to 23, and citations rose from 103 to 108.
Agents' Last Exam (UC Berkeley). Influential citations rose from 3 to 4, and citations rose from 37 to 42.
Correction Training for Diffusion Models (UC Berkeley). Hugging Face upvotes fell from 47 to 42.
Self-Evolving Coding Agents via Harness-Policy Optimization (Microsoft). Hugging Face upvotes fell from 53 to 48.
SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation (Alibaba (Qwen)). Hugging Face upvotes fell from 35 to 30.
AgentFold: Long-Horizon Web Agents with Proactive Context Management (Alibaba (Qwen)). Influential citations rose from 7 to 8, and citations rose from 103 to 107.
V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning (Meta). Influential citations rose from 52 to 53, and citations rose from 157 to 161.
Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization (UC Berkeley). Influential citations rose from 2 to 3, and citations rose from 25 to 29.
Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (DeepSeek). Citations rose from 520 to 525.
Pixtral 12B (Mistral AI). Citations rose from 185 to 190.
Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation (NVIDIA). Citations rose from 58 to 63.
LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory (Google DeepMind). Influential citations rose from 4 to 5, and citations rose from 42 to 46.
SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer (NVIDIA). Citations rose from 65 to 70.
Qwen3-TTS Technical Report (Alibaba (Qwen)). Citations rose from 189 to 194.
Why Language Models Hallucinate (OpenAI). Citations rose from 386 to 391.
Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model (Microsoft). Citations rose from 13 to 17.
MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (Ai2). Influential citations rose from 5 to 7, and citations rose from 14 to 16.
PaperBanana: Automating Academic Illustration for AI Scientists (Google). Influential citations rose from 8 to 9, and citations rose from 33 to 36.
WorkArena++: Towards Compositional Planning and Reasoning-based Common Knowledge Work Tasks (ServiceNow). Influential citations rose from 6 to 7, and citations rose from 111 to 114.
StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents (Salesforce). Influential citations rose from 0 to 2, and citations rose from 4 to 6.
RDMA Point-to-Point Communication for LLM Systems (Perplexity). Influential citations rose from 3 to 4, and citations rose from 13 to 16.
Privileged Information Distillation for Language Models (ServiceNow). Citations rose from 101 to 105.
MultiChallenge: A Realistic Multi-Turn Conversation Evaluation Benchmark Challenging to Frontier LLMs (Scale AI). Influential citations rose from 21 to 22, and citations rose from 204 to 207.
MCP-Universe: Benchmarking Large Language Models with Real-World Model Context Protocol Servers (Salesforce). Influential citations rose from 23 to 24, and citations rose from 125 to 128.
EmbeddingGemma 2: Unified and Efficient Multimodal Representations (Google). Hugging Face upvotes rose from 31 to 34.
SAM Audio: Segment Anything in Audio (Meta). Influential citations rose from 11 to 12, and citations rose from 55 to 57.
Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures (DeepSeek). Citations rose from 113 to 116.
Fast-dLLM v2: Efficient Block-Diffusion LLM (NVIDIA). Citations rose from 155 to 158.
Efficiently Reconstructing Dynamic Scenes One D4RT at a Time (Google). Influential citations rose from 7 to 8, and citations rose from 72 to 74.
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories (Google). Citations rose from 6 to 9.
MAI-UI Technical Report: Real-World Centric Foundation GUI Agents (Alibaba (Qwen)). Citations rose from 106 to 109.
Retrospective Harness Optimization: Improving LLM Agents via Self-Preference over Trajectory Rollouts (Microsoft). Citations rose from 26 to 29.
Proact-VL: A Proactive VideoLLM for Real-Time AI Companions (Microsoft). Influential citations rose from 2 to 3, and citations rose from 17 to 19.
Reasoning Models Struggle to Control their Chains of Thought (OpenAI). Citations rose from 34 to 37.
Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs? (Microsoft). Influential citations rose from 16 to 17, and citations rose from 108 to 110.
NitroGen: An Open Foundation Model for Generalist Gaming Agents (NVIDIA). Citations rose from 26 to 29.
Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning (NVIDIA). Influential citations rose from 11 to 12, and citations rose from 110 to 112.
Multiplex Thinking: Reasoning via Token-wise Branch-and-Merge (Microsoft). Influential citations rose from 2 to 3, and citations rose from 17 to 19.
Mobile-Agent-v3.5: Multi-platform Fundamental GUI Agents (Alibaba (Qwen)). Citations rose from 88 to 91.
Qwen3-Coder-Next Technical Report (Alibaba (Qwen)). Citations rose from 92 to 95.
Flash-KMeans: Fast and Memory-Efficient Exact K-Means (UC Berkeley). Citations rose from 13 to 16.
TriAttention: Efficient Long Reasoning with Trigonometric KV Compression (NVIDIA). Influential citations rose from 6 to 7, and citations rose from 23 to 25.
Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (Alibaba (Qwen)). Influential citations rose from 17 to 18, and citations rose from 56 to 58.
AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces (Microsoft). Citations rose from 19 to 22.
Qwen-AgentWorld: Language World Models for General Agents (Alibaba (Qwen)). Citations rose from 25 to 28.
Qianfan-OCR: A Unified End-to-End Model for Document Intelligence (Baidu). Influential citations rose from 2 to 3, and citations rose from 23 to 25.
LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding (NVIDIA). Influential citations rose from 8 to 9, and citations rose from 36 to 38.
Sea Change in Software Development: Economic and Productivity Analysis of the AI-Powered Developer Lifecycle (GitHub). Influential citations rose from 7 to 8, and citations rose from 50 to 52.
Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning (Salesforce). Influential citations rose from 1 to 2, and citations rose from 3 to 5.
Do generative video models learn physical principles from watching videos? (Google DeepMind). Hugging Face upvotes fell from 33 to 31.
Co-Design Bench: Multimodal Collaborative Coding Evaluation for Agentic Models (Alibaba (Qwen)). Hugging Face upvotes fell from 25 to 23.
Sharp Monocular View Synthesis in Less Than a Second (Apple). Citations rose from 30 to 32.
Jet-RL: Enabling On-Policy FP8 Reinforcement Learning with Unified Training and Rollout Precision Flow (NVIDIA). Citations rose from 16 to 18.
Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models (NVIDIA). Citations rose from 35 to 37.
Towards Autonomous Mathematics Research (Google). Citations rose from 37 to 39.
SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation (NVIDIA). Influential citations rose from 4 to 5, and citations rose from 5 to 6.
Qwen-Music Technical Report (Alibaba (Qwen)). Influential citations rose from 0 to 1, and citations rose from 3 to 4.
jina-embeddings-v3: Multilingual Embeddings With Task LoRA (Jina AI). Citations rose from 219 to 221.
SANA-Streaming: Real-time Streaming Video Editing with Hybrid Diffusion Transformer (NVIDIA). Citations rose from 13 to 15.
Soft Adaptive Policy Optimization (Alibaba (Qwen)). Influential citations fell from 28 to 27, and citations rose from 93 to 94.
PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NVIDIA). Influential citations rose from 2 to 3, and citations rose from 9 to 10.
PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation (NVIDIA). Influential citations rose from 2 to 3, and citations rose from 8 to 9.
Lyra 2.0: Explorable Generative 3D Worlds (NVIDIA). Influential citations rose from 19 to 20, and citations rose from 49 to 50.
Learning to Discover at Test Time (Stanford University). Citations rose from 116 to 118.
MARS: Modular Agent with Reflective Search for Automated AI Research (Google). Citations rose from 29 to 31.
From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents (IBM). Citations rose from 32 to 34.
Unlimited OCR Works (Baidu). Influential citations rose from 2 to 3, and citations rose from 8 to 9.
ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration (NVIDIA). Citations rose from 50 to 52.
TiDAR: Think in Diffusion, Talk in Autoregression (NVIDIA). Citations rose from 41 to 43.
On Data Engineering for Scaling LLM Terminal Capabilities (NVIDIA). Citations rose from 42 to 44.
Latent Spatial Memory for Video World Models (Microsoft). Citations rose from 22 to 24.
FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution (UC Berkeley). Influential citations rose from 2 to 3, and citations rose from 7 to 8.
Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes (Meta). Citations rose from 5 to 7.
Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections (Snowflake). Influential citations rose from 1 to 2, and citations rose from 2 to 3.
LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation (NVIDIA). Citations rose from 35 to 37.
SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation (Snowflake). Citations rose from 19 to 21.
xLAM: A Family of Large Action Models to Empower AI Agent Systems (Salesforce). Citations rose from 129 to 131.
SWE-INTERACT: Reimagining SWE Benchmarks as User-Driven Long-Horizon Coding Sessions (Scale AI). Citations rose from 21 to 23.
Reward Hacking in Rubric-Based Reinforcement Learning (Scale AI). Citations rose from 31 to 33.
Making, not Taking, the Best of N (Cohere). Influential citations rose from 1 to 2, and citations rose from 6 to 7.
LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet (Scale AI). Citations rose from 184 to 186.
GTA1: GUI Test-time Scaling Agent (Salesforce). Citations rose from 118 to 120.
Future Optical Flow Prediction Improves Robot Control & Video Generation (Salesforce). Influential citations rose from 0 to 1, and citations rose from 9 to 10.
EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings (ServiceNow). Citations rose from 16 to 18.
DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents (Google). Hugging Face upvotes fell from 22 to 21.
Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigs (NVIDIA). Hugging Face upvotes fell from 27 to 26.
Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use (ServiceNow). Hugging Face upvotes rose from 36 to 37.
ShapeR: Robust Conditional 3D Shape Generation from Casual Captures (Meta). Influential citations rose from 3 to 4.
FineVision: Open Data Is All You Need (Hugging Face). Citations rose from 52 to 53.
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution (Google). Citations rose from 22 to 23.
Voxtral Realtime (Mistral AI). Citations rose from 9 to 10.
V-RGBX: Video Editing with Accurate Controls over Intrinsic Properties (Adobe). Citations rose from 6 to 7.
One Layer Is Enough: Adapting Pretrained Visual Encoders for Image Generation (Apple). Citations rose from 31 to 32.
Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning (Snowflake). Citations rose from 80 to 81.
Agent Learning via Early Experience (Meta). Citations rose from 78 to 79.
Reinforcement World Model Learning for LLM-based Agents (Microsoft). Citations rose from 23 to 24.
Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System (Alibaba (Qwen)). Citations rose from 22 to 23.
Physics of Language Models: Part 4.1, Architecture Design and the Magic of Canon Layers (Meta). Citations rose from 47 to 48.
Memory-V2V: Augmenting Video-to-Video Diffusion Models with Memory (Adobe). Citations rose from 2 to 3.
Transition Matching Distillation for Fast Video Generation (NVIDIA). Citations rose from 29 to 30.
Test-Time Training with KV Binding Is Secretly Linear Attention (NVIDIA). Citations rose from 15 to 16.
Robot Learning from a Physical World Model (Google DeepMind). Citations rose from 22 to 23.
Reinforced Attention Learning (Google). Citations rose from 7 to 8.
DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (Microsoft). Citations rose from 36 to 37.
TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking (Google). Citations rose from 7 to 8.
Toto 2.0: Time Series Forecasting Enters the Scaling Era (Datadog). Citations rose from 38 to 39.
Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification (NVIDIA). Citations rose from 10 to 11.
SAMA: Factorized Semantic Anchoring and Motion Alignment for Instruction-Guided Video Editing (Baidu). Citations rose from 8 to 9.
PixelDiT: Pixel Diffusion Transformers for Image Generation (NVIDIA). Citations rose from 102 to 103.
NVIDIA-labs OO Agents: Native Python Object-Oriented Agents (NVIDIA). Citations rose from 4 to 5.
Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models (NVIDIA). Citations rose from 52 to 53.
The Verification Horizon: No Silver Bullet for Coding Agent Rewards (Alibaba (Qwen)). Citations rose from 10 to 11.
Agent Lightning v1.0: Towards Harnessed Agentic RL (Microsoft). Citations rose from 10 to 11.
OneStory: Coherent Multi-Shot Video Generation with Adaptive Memory (Meta). Influential citations fell from 3 to 2.
NVIDIA Nemotron 3: Efficient and Open Intelligence (NVIDIA). Citations rose from 59 to 60.
Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation (Alibaba (Qwen)). Citations fell from 19 to 18.
Qwen-Image-2.0-RL Technical Report (Alibaba (Qwen)). Citations rose from 9 to 10.
PhyCritic: Multimodal Critic Models for Physical AI (NVIDIA). Citations rose from 16 to 17.
Mode Seeking meets Mean Seeking for Fast Long Video Generation (NVIDIA). Citations rose from 13 to 14.
Let the Expert Stick to His Last: Expert-Specialized Fine-Tuning for Sparse Architectural Large Language Models (DeepSeek). Citations rose from 25 to 26.
From AGI to ASI (Google). Citations rose from 8 to 9.
Fish Audio S2 Technical Report (Fish Audio). Citations rose from 23 to 24.
Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients (NVIDIA). Citations rose from 3 to 4.
Unified Latents (UL): How to train your latents (Google). Influential citations fell from 3 to 2.
TUNA: Taming Unified Visual Representations for Native Unified Multimodal Models (Meta). Citations rose from 37 to 38.
Embarrassingly Simple Self-Distillation Improves Code Generation (Apple). Citations rose from 39 to 40.
Recurrent-Depth VLA: Implicit Test-Time Compute Scaling of Vision-Language-Action Models via Latent Iterative Reasoning (Ai2). Citations rose from 20 to 21.
HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning (Alibaba (Qwen)). Citations rose from 6 to 7.
Stabilizing Reinforcement Learning with LLMs: Formulation and Practices (Alibaba (Qwen)). Citations rose from 48 to 49.
SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning (NVIDIA). Citations rose from 22 to 23.
Recursive Multi-Agent Systems (Stanford University). Citations rose from 11 to 12.
FIPO: Eliciting Deep Reasoning with Future-KL Influenced Policy Optimization (Alibaba (Qwen)). Citations rose from 30 to 31.
Experiential Reinforcement Learning (Microsoft). Citations fell from 19 to 18.
Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents (Alibaba (Qwen)). Citations rose from 17 to 18.
Nemotron-4 340B Technical Report (NVIDIA). Citations rose from 141 to 142.
Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels (Salesforce). Influential citations rose from 4 to 5.
To Code, or Not To Code? Exploring Impact of Code in Pre-training (Cohere). Citations rose from 60 to 61.
The IOL-AI Challenge: An Open Challenge towards Advancing Linguistic Reasoning (Cohere). Citations rose from 0 to 1.
Terminal Agents Suffice for Enterprise Automation (ServiceNow). Citations fell from 7 to 6.
StarHarness: Evolving Harnesses with Stratified Search for Enterprise Environments (ServiceNow). Citations rose from 2 to 3.
SciPredict: Can LLMs Predict the Outcomes of Scientific Experiments in Natural Sciences? (Scale AI). Citations rose from 3 to 4.
ResearchRubrics: A Benchmark of Prompts and Rubrics For Evaluating Deep Research Agents (Scale AI). Citations rose from 95 to 96.
Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents (Scale AI). Citations rose from 81 to 82.
Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR (Scale AI). Citations rose from 12 to 13.
Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures (Scale AI). Citations rose from 10 to 11.
Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory (Accenture). Citations rose from 16 to 17.
MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers (Scale AI). Citations rose from 64 to 65.
Least-Loaded Expert Parallelism: Load Balancing An Imbalanced Mixture-of-Experts (Salesforce). Citations rose from 10 to 11.
How to Train Your LLM Web Agent: A Statistical Diagnosis (ServiceNow). Citations rose from 22 to 23.
HarnessOpt-Bench: Evaluating LLMs at Harness Optimization (Scale AI). Influential citations fell from 2 to 1.
Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math (Salesforce). Citations rose from 22 to 23.
Grounding Computer Use Agents on Human Demonstrations (ServiceNow). Citations rose from 22 to 23.
EVA-Bench: A New End-to-end Framework for Evaluating Voice Agents (ServiceNow). Citations rose from 19 to 20.

NVIDIA 39 Alibaba (Qwen) 32 Microsoft 18 Meta 15 Google 15 Scale AI 13 ServiceNow 11 Salesforce 9 UC Berkeley 8 OpenAI 7 Google DeepMind 5 DeepSeek 5+19 labs

As JSON: /api/changes?since=