标签 — LLM 研究知识库

rl 65

training 64

Agent-Native Research Artifact (Ara)Agentic Harness Engineering (AHE)AgenticQwen — Dual Data Flywheels for Small Agentic Models Are Complicated Loss Functions Necessary for Teaching LLMs to Reason?Audio-Agent Cascade Rl CoEvolve: Training LLM Agents via Agent-Data Mutual Evolution Composer 2 COS-PLAY — Co-Evolving LLM Decision and Skill Bank Agents Dual Policy Distillation Dynin-Omni: Diffusion-Based Omnimodal Foundation Model EmoOmni: Emotional Understanding in Omni-LLMs Entropy Aware On Policy Distillation Even Less Is Even Better: Agentic/Reasoning/Coding LLM Exposure Bias Free Process Rewards Generalized Knowledge Distillation (GKD)Generalized On Policy Distillation GRPO RL Training Horizon Length Training Study — arXiv 2605.02572 Hybrid Distillation Policy Optimization Imitation Learning Interactive Imitation Learning Knowledge Distillation Learning to Orchestrate Agents with the RL Conductor LLM-PRISM: GPU Faults & Silent Data Corruption LongCat-Flash LongCat-Flash-Omni LongCat-Flash-Thinking LongCat-Flash-Thinking-2601 LongCat-Image LongCat-Next LongCat-Video M+: MemoryLLM 的可扩展长期记忆扩展 MemoryLLM: 自更新大语言模型的潜在空间记忆 MiniCPM-o 4.5 Model Distillation Multi-Domain On-Policy Distillation Multi-Task Learning Negative Data Mining for Dense Retrieval at IKEA NIM4-ASR: Efficient, Robust, Customizable Real-Time LLM-Based ASR On Policy Distillation On Policy Prefix Distillation On-Policy Self-Distillation for Reasoning Compression OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories OPSDL: On-Policy Self-Distillation for Long-Context Language Models Ppo Proximal Policy Distillation Qwen3 语音家族深度解析 Qwen3.5-Omni ReasonXL: Shifting LLM Reasoning Language Without Sacrificing Performance Recursive Multi-Agent Systems (RecursiveMAS)Reinforcement-Aware Knowledge Distillation Relax: Async RL Engine for Omni-Modal Post-Training RL Post-Training Scaling Laws for LLMs Sample Routed Policy Optimization Self-Improving Pretraining: Using Post-Trained Models to Pretrain Better Models Silent Thought / FLAIR（潜在推理全双工对话）Stochastic KV Routing: Random Cross-Layer Attention for Depth-Wise Cache Sharing Strong to Weak Distillation TurnGuide U-Mind: Real-Time Multimodal Interaction V2A-DPO: Video-to-Audio Preference Optimization Why Mean Pooling Works: Quantifying Second-Order Collapse in Text Embeddings

architecture 57

A-Mem: Zettelkasten 式自组织 Agent 记忆 AriadneMem: 线团式终身记忆导航 Audio-Agent AudioKV: KV Cache Eviction for Large Audio-Language Models AVID: Audio-Visual Inconsistency Benchmark Chain of Modality: Dynamic Orchestration in Omni-MLLMs Composer 2 D-Mem: 双过程记忆系统 DuplexCascade Dynin-Omni: Diffusion-Based Omnimodal Foundation Model EAGLE (Speculative Decoding)EmoOmni: Emotional Understanding in Omni-LLMs EMOVA LongCat-Flash LongCat-Flash-Omni LongCat-Flash-Thinking LongCat-Flash-Thinking-2601 LongCat-Image LongCat-Next LongCat-Video M+: MemoryLLM 的可扩展长期记忆扩展 MemGPT: LLM 即操作系统 Memori: LLM 无关的持久记忆层 MemoryLLM: 自更新大语言模型的潜在空间记忆 MGM-Omni MiniCPM-o 4.5 Mixture of Experts Modality Preference of Omni-Modal LLMs Multi-Token Prediction (MTP)Nemori: 认知启发的自适应记忆蒸馏框架 NIM4-ASR: Efficient, Robust, Customizable Real-Time LLM-Based ASR Omni-Modal LLM Paraformer — Non-Autoregressive Parallel Transformer for ASR Paraformer-v2 — Noise-Robust Multilingual NAR ASR Qwen3 Qwen3 语音家族深度解析 Qwen3-ASR Qwen3-Omni Qwen3-TTS Qwen3.5 Qwen3.5-Omni Recursive Multi-Agent Systems (RecursiveMAS)Relax: Async RL Engine for Omni-Modal Post-Training SA-Paraformer — Speaker-Attributed Non-Autoregressive ASR SALMONN-omni SeACo-Paraformer — Non-Autoregressive ASR with Hotword Customization SenseVoice — Multi-Task Speech Understanding Foundation Model shortcut-connected-moe SimpleMem: 高效语义压缩终身记忆 Stochastic KV Routing: Random Cross-Layer Attention for Depth-Wise Cache Sharing Stream-Omni Tachiom: Efficient Multivector Retrieval with Token-Aware Clustering Text2Mem: 统一记忆操作语言 U-Mind: Real-Time Multimodal Interaction Why Mean Pooling Works: Quantifying Second-Order Collapse in Text Embeddings zero-computation-experts 全双工 Agent 融合架构（Duplex-Agent Integration）

distillation 52

model 50

Agentic World Modeling — Foundations, Capabilities, Laws Composer 2 CosyVoice CosyVoice 2 CosyVoice 3 Diffusion Language Models Dynin-Omni: Diffusion-Based Omnimodal Foundation Model EmoOmni: Emotional Understanding in Omni-LLMs EMOVA Freeze-Omni FunASR — Fundamental End-to-End Speech Recognition Toolkit FunAudioLLM — Voice Understanding & Generation Foundation Models FunCodec — Neural Speech Codec Toolkit Kimi K2 LLaMA-Omni LongCat-Flash LongCat-Flash-Omni LongCat-Flash-Thinking LongCat-Flash-Thinking-2601 LongCat-Image LongCat-Next LongCat-Video M+: MemoryLLM 的可扩展长期记忆扩展 MemoryLLM: 自更新大语言模型的潜在空间记忆 MGM-Omni MiniCPM-o 4.5 MinMo Mixture of Experts Moshi NIM4-ASR: Efficient, Robust, Customizable Real-Time LLM-Based ASR Omni-Modal LLM Paraformer — Non-Autoregressive Parallel Transformer for ASR Paraformer-v2 — Noise-Robust Multilingual NAR ASR Qwen3 Qwen3 语音家族深度解析 Qwen3-ASR Qwen3-Omni Qwen3-TTS Qwen3.5 Qwen3.5-Omni SA-Paraformer — Speaker-Attributed Non-Autoregressive ASR SALMONN-omni SeACo-Paraformer — Non-Autoregressive ASR with Hotword Customization Seed Realtime Voice Seed-TTS Seeduplex SenseVoice — Multi-Task Speech Understanding Foundation Model Speech-Omni-Lite Stream-Omni U-Mind: Real-Time Multimodal Interaction

agent 50

A-Mem: Zettelkasten 式自组织 Agent 记忆 Agent Memory System: LLM Agent 长期记忆系统总览 Agent Safety Agent Safety via RL Agent 供应链攻击 Agent-Native Research Artifact (Ara)Agent-R1: End-to-End RL for LLM Agents AgentAlign Agentic Code Reasoning Agentic Harness Engineering (AHE)Agentic Research Ideas Agentic Workflows AgenticRecTune — Multi-Agent Recommendation System Optimization AgentPoison Agent² RL-Bench: Agentic RL Post-Training AriadneMem: 线团式终身记忆导航 Audio-Agent Beyond Human-Readable: Rethinking SE Conventions for Agents ClawGuard: Runtime Security for Tool-Augmented Agents CoEvolve: Training LLM Agents via Agent-Data Mutual Evolution D-Mem: 双过程记忆系统 GASP: Guided Asymmetric Self-Play For Coding LLMs Generative Agents HyperMem: 超图记忆架构 LightMem: 轻量级 Agent 记忆系统 Mem0: AI Agent 的可扩展长期记忆层 MemEvoBench: Memory Safety Benchmark MemGPT: LLM 即操作系统 Memori: LLM 无关的持久记忆层 MemReader: 从被动到主动的记忆提取 Mixture of Agents ML-Agent: Autonomous ML via Reinforcement Learning Multi-Agent Coordination Nemori: 认知启发的自适应记忆蒸馏框架 PASK: 意图感知的主动 Agent Prompt Injection RAGEN: Multi-Turn RL for LLM Agents ReaLM-Retrieve — Adaptive Retrieval for Large Reasoning Models Reflexion Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis SAVeR: Self-Audited Verified Reasoning Self-Evolving LLM Memory Extraction SimpleMem: 高效语义压缩终身记忆 Skill Orchestration Skill-SD: Skill-Conditioned Self-Distillation Text2Mem: 统一记忆操作语言 Thought-Aligner VerificAgent Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw 全双工 Agent 融合架构（Duplex-Agent Integration）

speech-model 47

ASR Benchmarks Audio-Agent AudioKV: KV Cache Eviction for Large Audio-Language Models CosyVoice CosyVoice 2 CosyVoice 3 DiffRO DuplexCascade EMOVA Freeze-Omni FunASR — Fundamental End-to-End Speech Recognition Toolkit FunAudioLLM — Voice Understanding & Generation Foundation Models FunCodec — Neural Speech Codec Toolkit LLaMA-Omni LongCat-Flash-Omni MGM-Omni MiniCPM-o 4.5 MinMo Moshi MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models MTR-DuplexBench NIM4-ASR: Efficient, Robust, Customizable Real-Time LLM-Based ASR Omni Model Evolution Overview Omni-Modal LLM Paraformer — Non-Autoregressive Parallel Transformer for ASR Paraformer-v2 — Noise-Robust Multilingual NAR ASR Privacy-Preserving Full-Duplex Speech Dialogue Qwen3 语音家族深度解析 Qwen3-ASR Qwen3-Omni Qwen3-TTS Qwen3.5-Omni SA-Paraformer — Speaker-Attributed Non-Autoregressive ASR SALMONN-omni SeACo-Paraformer — Non-Autoregressive ASR with Hotword Customization Seed Realtime Voice Seed-TTS Seeduplex SenseVoice — Multi-Task Speech Understanding Foundation Model Silent Thought / FLAIR（潜在推理全双工对话）Speech Llm Speech-Omni-Lite Stream-Omni TurnGuide X Opd 全双工语音模型（Full-Duplex Speech Model）声学前端与后端 ASR 的平衡（Audio Frontend-Backend Balance）

reasoning 43

A Decomposition Perspective to Long-Context Reasoning for LLMs Agentic Code Reasoning Agentic World Modeling — Foundations, Capabilities, Laws Are Complicated Loss Functions Necessary for Teaching LLMs to Reason?Chain of Thought Contextual Agentic Memory is a Memo, Not True Memory DeepSeek R1 Distillation Do Vision-Language Models Truly Perform Vision Reasoning?Even Less Is Even Better: Agentic/Reasoning/Coding LLM Latent Agents — Internalized Multi-Agent Debate (IMAD)Learning to Orchestrate Agents with the RL Conductor LongCat-Flash LongCat-Flash-Thinking LongCat-Flash-Thinking-2601 LongCat-Next Mixture-of-Layers Distillation with Stepwise Attention on Key Information Multimodal Latent Reasoning via Predictive Embeddings Omni-R1: 统一生成式多模态推理 On Policy Self Distillation OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories Out of Context Reasoning Outcome Rewards Do Not Guarantee Verifiable or Causally Important Reasoning Process Supervision of Confidence Margin for Calibrated LLM Reasoning ReaLM-Retrieve — Adaptive Retrieval for Large Reasoning Models Reasoning Distillation ReasonXL: Shifting LLM Reasoning Language Without Sacrificing Performance Recursive Multi-Agent Systems (RecursiveMAS)Reflexion Reopold Rethinking OPD: Phenomenology, Mechanism, and Recipe SAVeR: Self-Audited Verified Reasoning SCOPE: Signal-Calibrated On-Policy Distillation Enhancement Silent Thought / FLAIR（潜在推理全双工对话）STaR SUPERNOVA: 自然指令 RL 激发通用推理 SVSR: 多模态推理的自验证自修正 Test-Time Scaling The Expense of Seeing: Trustworthy Multimodal Reasoning Within Monolithic Paradigm Thinking Budget Thought-Aligner TTVS: Test-Time Variational Synthesis Visually-Guided Policy Optimization for Multimodal Reasoning π²: Structure-Originated Reasoning Data Improves Long-Context Reasoning

open-source 37

Agentic Coding Composer 2 CosyVoice CosyVoice 2 CosyVoice 3 EMOVA FunASR — Fundamental End-to-End Speech Recognition Toolkit FunAudioLLM — Voice Understanding & Generation Foundation Models FunCodec — Neural Speech Codec Toolkit Kimi K2 LLaMA-Omni LongCat-Flash LongCat-Flash-Omni LongCat-Flash-Thinking LongCat-Flash-Thinking-2601 LongCat-Image LongCat-Next LongCat-Video Mem0: AI Agent 的可扩展长期记忆层 MiniCPM-o 4.5 Moshi OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories Paraformer — Non-Autoregressive Parallel Transformer for ASR Paraformer-v2 — Noise-Robust Multilingual NAR ASR Qwen3 Qwen3-ASR Qwen3-Omni Qwen3-TTS Qwen3.5 SA-Paraformer — Speaker-Attributed Non-Autoregressive ASR SALMONN-omni SeACo-Paraformer — Non-Autoregressive ASR with Hotword Customization SenseVoice — Multi-Task Speech Understanding Foundation Model shortcut-connected-moe Stable Baselines3 Stream-Omni zero-computation-experts

multimodal 36

AVID: Audio-Visual Inconsistency Benchmark Chain of Modality: Dynamic Orchestration in Omni-MLLMs Do Vision-Language Models Truly Perform Vision Reasoning?Dynin-Omni: Diffusion-Based Omnimodal Foundation Model EmoOmni: Emotional Understanding in Omni-LLMs EMOVA Gaussian GRPO (G2RPO)LongCat-Flash-Omni LongCat-Image LongCat-Next LongCat-Video MGM-Omni MinMo Modality Preference of Omni-Modal LLMs Multimodal Latent Reasoning via Predictive Embeddings Omni Model Evolution Overview Omni-Modal LLM Omni-R1: 统一生成式多模态推理 OmniTrace: Generation-Time Attribution in Omni-Modal LLMs Pseudo-Unification: 熵探测揭示多模态模型信息分歧 Qwen3 语音家族深度解析 Qwen3-Omni Qwen3.5-Omni Relax: Async RL Engine for Omni-Modal Post-Training Speech Llm Speech-Omni-Lite Stream-Omni SVSR: 多模态推理的自验证自修正 The Expense of Seeing: Trustworthy Multimodal Reasoning Within Monolithic Paradigm U-Mind: Real-Time Multimodal Interaction V2A-DPO: Video-to-Audio Preference Optimization Video Opd Visual Enhanced Depth Scaling for Multimodal Latent Reasoning Visually-Guided Policy Optimization for Multimodal Reasoning Vold X Opd

alignment 30

Agent Safety Agent Safety via RL AgentAlign Agentic World Modeling — Foundations, Capabilities, Laws AI Self-Awareness via Game Theory AudioKV: KV Cache Eviction for Large Audio-Language Models Behavioral Self-Awareness Cascade Rl Chain of Modality: Dynamic Orchestration in Omni-MLLMs Contextual Agentic Memory is a Memo, Not True Memory DP-OPD: Differentially Private On-Policy Distillation Emergent Misalignment EmoOmni: Emotional Understanding in Omni-LLMs Epistemologically-Informed LLM Self-Learning From Skill Text to Skill Structure: The SSL Representation Latent Agents — Internalized Multi-Agent Debate (IMAD)Learning to Orchestrate Agents with the RL Conductor LongCat-Flash LongCat-Flash-Omni MemEvoBench: Memory Safety Benchmark Mental Self-Modeling MiniCPM-o 4.5 Modality Preference of Omni-Modal LLMs OmniTrace: Generation-Time Attribution in Omni-Modal LLMs PARM: Pipeline-Adapted Reward Model Reinforcement Learning from Human Feedback Self-Improving Pretraining: Using Post-Trained Models to Pretrain Better Models The Illusion of Certainty: Decoupling Capability and Calibration in OPD V2A-DPO: Video-to-Audio Preference Optimization WaltzRL

on-policy 30

Co-Evolving Policy Distillation (CoPD)DP-OPD: Differentially Private On-Policy Distillation Entropy Aware On Policy Distillation Ex-OPD Generalized Knowledge Distillation (GKD)Generalized On Policy Distillation Hybrid Distillation Policy Optimization Hybrid Policy Distillation for LLMs Lightning OPD: Offline On-Policy Distillation LLM Post-Training: A Unified View of Off-Policy and On-Policy Learning Multi-Domain On-Policy Distillation On Policy Distillation On Policy Distillation Survey On Policy Prefix Distillation On Policy Self Distillation On-Policy Self-Distillation for Reasoning Compression OPSDL: On-Policy Self-Distillation for Long-Context Language Models Qwen3.5-Omni Reinforcement-Aware Knowledge Distillation Reopold Rethinking OPD: Phenomenology, Mechanism, and Recipe Sample Routed Policy Optimization Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision Self-Distilled RLVR: On-Policy Distillation for RL with Verifiable Rewards Self-Improving Pretraining: Using Post-Trained Models to Pretrain Better Models The Illusion of Certainty: Decoupling Capability and Calibration in OPD TIP: Token Importance in On-Policy Distillation Video Opd Vold X Opd

memory 27

A-Mem: Zettelkasten 式自组织 Agent 记忆 Agent Memory System: LLM Agent 长期记忆系统总览 AgenticRecTune — Multi-Agent Recommendation System Optimization AgentPoison AriadneMem: 线团式终身记忆导航 Complementary Learning Systems D-Mem: 双过程记忆系统 Generative Agents HyperMem: 超图记忆架构 LightMem: 轻量级 Agent 记忆系统 LoCoMo: 长期对话记忆评测基准 LongMemEval: 长期交互记忆评测基准 M+: MemoryLLM 的可扩展长期记忆扩展 Mem0: AI Agent 的可扩展长期记忆层 MemEvoBench: Memory Safety Benchmark MemGPT: LLM 即操作系统 Memori: LLM 无关的持久记忆层 MemoryLLM: 自更新大语言模型的潜在空间记忆 MemReader: 从被动到主动的记忆提取 Nemori: 认知启发的自适应记忆蒸馏框架 PASK: 意图感知的主动 Agent ROME / MEMIT Self-Evolving LLM Memory Extraction SimpleMem: 高效语义压缩终身记忆 Stochastic KV Routing: Random Cross-Layer Attention for Depth-Wise Cache Sharing Text2Mem: 统一记忆操作语言 VerificAgent

inference 24

benchmark 22

Agent-Native Research Artifact (Ara)Agentic Harness Engineering (AHE)Agent² RL-Bench: Agentic RL Post-Training ASR Benchmarks Atari 2600 AVID: Audio-Visual Inconsistency Benchmark Contextual Agentic Memory is a Memo, Not True Memory CursorBench From Skill Text to Skill Structure: The SSL Representation Horizon Length Training Study — arXiv 2605.02572 Learning to Orchestrate Agents with the RL Conductor LoCoMo: 长期对话记忆评测基准 LongMemEval: 长期交互记忆评测基准 MemEvoBench: Memory Safety Benchmark MiniCPM-o 4.5 Modality Preference of Omni-Modal LLMs MTR-DuplexBench Negative Data Mining for Dense Retrieval at IKEA OmniTrace: Generation-Time Attribution in Omni-Modal LLMs OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories Qwen3-ASR ReaLM-Retrieve — Adaptive Retrieval for Large Reasoning Models

safety 20

optimization 20

stub 19

Agent Safety Agentic Research Ideas Agentic Workflows AgentPoison ASR Benchmarks Complementary Learning Systems Diffusion Language Models EAGLE-3 Generative Agents LLM Inference Optimization Mixture of Agents Multi-Agent Coordination Omni Model Evolution Overview Prompt Injection Qdrant Vector Search Reflexion ROME / MEMIT Skill Orchestration Test-Time Scaling

streaming 17

CosyVoice 2 DuplexCascade FunASR — Fundamental End-to-End Speech Recognition Toolkit MGM-Omni MiniCPM-o 4.5 NIM4-ASR: Efficient, Robust, Customizable Real-Time LLM-Based ASR PASK: 意图感知的主动 Agent Privacy-Preserving Full-Duplex Speech Dialogue Qwen3-Omni Qwen3-TTS SALMONN-omni SeACo-Paraformer — Non-Autoregressive ASR with Hotword Customization Silent Thought / FLAIR（潜在推理全双工对话）Stream-Omni TurnGuide 全双工 Agent 融合架构（Duplex-Agent Integration）全双工语音模型（Full-Duplex Speech Model）

end-to-end 16

Agent-R1: End-to-End RL for LLM Agents Freeze-Omni FunASR — Fundamental End-to-End Speech Recognition Toolkit FunAudioLLM — Voice Understanding & Generation Foundation Models LLaMA-Omni Moshi Paraformer — Non-Autoregressive Parallel Transformer for ASR Paraformer-v2 — Noise-Robust Multilingual NAR ASR PerlAD: Pseudo-simulation-based RL for E2E Autonomous Driving Privacy-Preserving Full-Duplex Speech Dialogue Seed Realtime Voice Seeduplex SenseVoice — Multi-Task Speech Understanding Foundation Model Silent Thought / FLAIR（潜在推理全双工对话）TurnGuide 全双工语音模型（Full-Duplex Speech Model）

long-term 16

A-Mem: Zettelkasten 式自组织 Agent 记忆 Agent Memory System: LLM Agent 长期记忆系统总览 AriadneMem: 线团式终身记忆导航 D-Mem: 双过程记忆系统 HyperMem: 超图记忆架构 LightMem: 轻量级 Agent 记忆系统 LoCoMo: 长期对话记忆评测基准 LongMemEval: 长期交互记忆评测基准 M+: MemoryLLM 的可扩展长期记忆扩展 Mem0: AI Agent 的可扩展长期记忆层 MemGPT: LLM 即操作系统 Memori: LLM 无关的持久记忆层 MemoryLLM: 自更新大语言模型的潜在空间记忆 MemReader: 从被动到主动的记忆提取 Nemori: 认知启发的自适应记忆蒸馏框架 SimpleMem: 高效语义压缩终身记忆

grpo 13

A Decomposition Perspective to Long-Context Reasoning for LLMs Agent² RL-Bench: Agentic RL Post-Training Are Complicated Loss Functions Necessary for Teaching LLMs to Reason?Co-Evolving Policy Distillation (CoPD)Faithful GRPO Free Process Rewards Gaussian GRPO (G2RPO)GRPO GRPO RL Training MemReader: 从被动到主动的记忆提取 Sample Routed Policy Optimization Self-Distilled RLVR: On-Policy Distillation for RL with Verifiable Rewards TTVS: Test-Time Variational Synthesis

tool-use 12

Agent 供应链攻击 Agentic Coding AgenticQwen — Dual Data Flywheels for Small Agentic Models Audio-Agent ClawGuard: Runtime Security for Tool-Augmented Agents From Skill Text to Skill Structure: The SSL Representation LongCat-Flash-Thinking LongCat-Flash-Thinking-2601 OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories Skill Orchestration Text2Mem: 统一记忆操作语言全双工 Agent 融合架构（Duplex-Agent Integration）

full-duplex 12

DuplexCascade MiniCPM-o 4.5 Moshi MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models MTR-DuplexBench Privacy-Preserving Full-Duplex Speech Dialogue SALMONN-omni Seeduplex Silent Thought / FLAIR（潜在推理全双工对话）TurnGuide 全双工 Agent 融合架构（Duplex-Agent Integration）全双工语音模型（Full-Duplex Speech Model）

company 10

Composer 2 CursorBench Doubao App Freeze-Omni Mem0: AI Agent 的可扩展长期记忆层 MinMo Seed Realtime Voice Seed-TTS Seeduplex 字节跳动 Seed

agentic-coding 9

Agent-Native Research Artifact (Ara)Agentic Harness Engineering (AHE)Beyond Human-Readable: Rethinking SE Conventions for Agents Even Less Is Even Better: Agentic/Reasoning/Coding LLM ORACLE-SWE Rethinking Software Engineering for Agentic AI SWE-AGILE: Dynamic Reasoning Context SWE-HERO: 从无执行到有执行的 SWE Agent 训练 SWE-Shepherd: PRMs for Code Agents

behavioral-awareness 9

AI Self-Awareness via Game Theory Backdoor Awareness Behavioral Self-Awareness Emergent Misalignment Introspection Mental Self-Modeling Out of Context Reasoning Situational Awareness Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw

concept 9

Audio-Agent DuplexCascade MTR-DuplexBench Privacy-Preserving Full-Duplex Speech Dialogue Silent Thought / FLAIR（潜在推理全双工对话）TurnGuide 全双工 Agent 融合架构（Duplex-Agent Integration）全双工语音模型（Full-Duplex Speech Model）声学前端与后端 ASR 的平衡（Audio Frontend-Backend Balance）

evaluation 9

Contextual Agentic Memory is a Memo, Not True Memory CursorBench From Skill Text to Skill Structure: The SSL Representation Learning to Orchestrate Agents with the RL Conductor LoCoMo: 长期对话记忆评测基准 LongMemEval: 长期交互记忆评测基准 MTR-DuplexBench ORACLE-SWE Polyglot Teachers: Multilingual Synthetic Data

retrieval 7

MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models Negative Data Mining for Dense Retrieval at IKEA Qdrant Vector Search RAG-Considerate Pretraining ReaLM-Retrieve — Adaptive Retrieval for Large Reasoning Models Tachiom: Efficient Multivector Retrieval with Token-Aware Clustering Why Mean Pooling Works: Quantifying Second-Order Collapse in Text Embeddings

codec 6

EMOVA FunCodec — Neural Speech Codec Toolkit MGM-Omni Qwen3 语音家族深度解析 Qwen3-Omni Qwen3-TTS

survey 6

Agent Memory System: LLM Agent 长期记忆系统总览 Agentic World Modeling — Foundations, Capabilities, Laws LLM Post-Training: A Unified View of Off-Policy and On-Policy Learning Omni-Modal LLM On Policy Distillation Survey Qwen3 语音家族深度解析

multi-agent 6

AgenticRecTune — Multi-Agent Recommendation System Optimization Mixture of Agents Multi-Agent Coordination Recursive Multi-Agent Systems (RecursiveMAS)Rethinking Software Engineering for Agentic AI WaltzRL

fine-tuning 6

AgenticQwen — Dual Data Flywheels for Small Agentic Models COS-PLAY — Co-Evolving LLM Decision and Skill Bank Agents DP-OPD: Differentially Private On-Policy Distillation OPSDL: On-Policy Self-Distillation for Long-Context Language Models Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision SWE-HERO: 从无执行到有执行的 SWE Agent 训练

on-policy-distillation 6

On-Policy Distillation for Autonomous Vehicle Motion Planning On-Policy Distillation vs SFT 的优势与指标差距 OPD 技术中 Student/Teacher 是否需要同一 Tokenizer Qwen3 中的 On-Policy Distillation 做法 Qwen3 技术要点总览 SCOPE: Signal-Calibrated On-Policy Distillation Enhancement

speculative-decoding 5

DDTree DFlash EAGLE (Speculative Decoding)EAGLE-3 Multi-Token Prediction (MTP)

diffusion 5

Audio-Omni: 统一音频理解+生成+编辑 BlendFusion: 扩散模型合成数据生成 DDTree DFlash Diffusion Language Models

security 5

Agent 供应链攻击 AgentPoison ClawGuard: Runtime Security for Tool-Augmented Agents Combating Data Laundering in LLM Training Prompt Injection

synthetic-data 5

BlendFusion: 扩散模型合成数据生成 Optimsyn: Influence-Guided Rubrics for Synthetic Data Polyglot Teachers: Multilingual Synthetic Data RL-Guided Synthetic Data Generation WRAP++: Web Discovery Amplified Pretraining

rlhf 5

Co-Evolving Policy Distillation (CoPD)Free Process Rewards LLM Post-Training: A Unified View of Off-Policy and On-Policy Learning ReflectRM Reinforcement Learning from Human Feedback

vision 5

Do Vision-Language Models Truly Perform Vision Reasoning?Faithful GRPO Perception-Grounded Policy Optimization for VLMs Visual Enhanced Depth Scaling for Multimodal Latent Reasoning Visually-Guided Policy Optimization for Multimodal Reasoning

reward-model 5

An Imperfect Verifier is Good Enough: Learning with Noisy Rewards Outcome Rewards Do Not Guarantee Verifiable or Causally Important Reasoning PARM: Pipeline-Adapted Reward Model ReflectRM Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis

pretraining 5

LLM Training as Lossy Compression RAG-Considerate Pretraining Scalable MoE 预训练 Self-Improving Pretraining: Using Post-Trained Models to Pretrain Better Models WRAP++: Web Discovery Amplified Pretraining

dataset 5

LoCoMo: 长期对话记忆评测基准 LongMemEval: 长期交互记忆评测基准 MTR-DuplexBench OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories π²: Structure-Originated Reasoning Data Improves Long-Context Reasoning

embedding 5

Multimodal Latent Reasoning via Predictive Embeddings Negative Data Mining for Dense Retrieval at IKEA Qdrant Vector Search Tachiom: Efficient Multivector Retrieval with Token-Aware Clustering Why Mean Pooling Works: Quantifying Second-Order Collapse in Text Embeddings

inference-optimization 4

DDTree DFlash LLM Inference Optimization Speculative Decoding

acceleration 4

DFlash EAGLE-3 LLM Inference Optimization Speculative Decoding

rag 4

Agent Memory System: LLM Agent 长期记忆系统总览 MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models NIM4-ASR: Efficient, Robust, Customizable Real-Time LLM-Based ASR RAG-Considerate Pretraining

reinforcement-learning 4

Agent-R1: End-to-End RL for LLM Agents ML-Agent: Autonomous ML via Reinforcement Learning PerlAD: Pseudo-simulation-based RL for E2E Autonomous Driving RAGEN: Multi-Turn RL for LLM Agents

post-training 4

Agent² RL-Bench: Agentic RL Post-Training Gaussian GRPO (G2RPO)LLM Post-Training: A Unified View of Off-Policy and On-Policy Learning OmniJigsaw

audio 4

Audio-Omni: 统一音频理解+生成+编辑 AudioKV: 音频大模型 KV Cache 驱逐 OmniJigsaw Whisper-AuT: Domain-Adapted Audio Encoder

chain-of-thought 3

Chain of Thought LongCat-Flash-Thinking STaR

graph 3

Agent Memory System: LLM Agent 长期记忆系统总览 AriadneMem: 线团式终身记忆导航 Mem0: AI Agent 的可扩展长期记忆层

coding 3

Agentic Code Reasoning Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning GASP: Guided Asymmetric Self-Play For Coding LLMs

software-engineering 3

Agentic Code Reasoning Beyond Human-Readable: Rethinking SE Conventions for Agents Rethinking Software Engineering for Agentic AI

asr 3

ASR Benchmarks NIM4-ASR: Efficient, Robust, Customizable Real-Time LLM-Based ASR 声学前端与后端 ASR 的平衡（Audio Frontend-Backend Balance）

long-context 3

A Decomposition Perspective to Long-Context Reasoning for LLMs AudioKV: 音频大模型 KV Cache 驱逐 π²: Structure-Originated Reasoning Data Improves Long-Context Reasoning

self-play 3

Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning GASP: Guided Asymmetric Self-Play For Coding LLMs Self-Play Limitations

process-reward 3

Process Supervision of Confidence Margin for Calibrated LLM Reasoning Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis SWE-Shepherd: PRMs for Code Agents

omni-modal 3

MiniCPM-o 4.5 Omni Model Evolution Overview OmniJigsaw

swe-agent 3

ORACLE-SWE SWE-AGILE: Dynamic Reasoning Context SWE-HERO: 从无执行到有执行的 SWE Agent 训练

policy-optimization 3

Perception-Grounded Policy Optimization for VLMs Policy Split: Dual-Mode Exploration Visually-Guided Policy Optimization for Multimodal Reasoning

voice-assistant 2

Doubao App MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models

recommendation 2

AgenticRecTune — Multi-Agent Recommendation System Optimization Negative Data Mining for Dense Retrieval at IKEA

orchestration 2

Agentic Workflows Skill Orchestration

generation 2

Audio-Omni: 统一音频理解+生成+编辑 Polyglot Teachers: Multilingual Synthetic Data

efficiency 2

AudioKV: 音频大模型 KV Cache 驱逐 Even Less Is Even Better: Agentic/Reasoning/Coding LLM

data-generation 2

BlendFusion: 扩散模型合成数据生成 RL-Guided Synthetic Data Generation

data 2

CoEvolve: Training LLM Agents via Agent-Data Mutual Evolution π²: Structure-Originated Reasoning Data Improves Long-Context Reasoning

data-quality 2

Combating Data Laundering in LLM Training LLM-PRISM: GPU Faults & Silent Data Corruption

faithfulness 2

Faithful GRPO SAVeR: Self-Audited Verified Reasoning

conversation 2

HyperMem: 超图记忆架构 LoCoMo: 长期对话记忆评测基准

rlvr 2

An Imperfect Verifier is Good Enough: Learning with Noisy Rewards Outcome Rewards Do Not Guarantee Verifiable or Causally Important Reasoning

information-theory 2

LLM Training as Lossy Compression Pseudo-Unification: 熵探测揭示多模态模型信息分歧

qwen 2

ML-Agent: Autonomous ML via Reinforcement Learning RL Post-Training Scaling Laws for LLMs

comparison 2

Omni-Modal LLM Qwen3 语音家族深度解析

generative 2

Omni-R1: 统一生成式多模态推理 ReflectRM

self-distillation 2

On-Policy Self-Distillation for Reasoning Compression Skill-SD: Skill-Conditioned Self-Distillation

autonomous-driving 2

On-Policy Distillation for Autonomous Vehicle Motion Planning PerlAD: Pseudo-simulation-based RL for E2E Autonomous Driving

multilingual 2

Polyglot Teachers: Multilingual Synthetic Data ReasonXL: Shifting LLM Reasoning Language Without Sacrificing Performance

scaling-laws 2

RAG-Considerate Pretraining RL Post-Training Scaling Laws for LLMs

multi-turn 2

RAGEN: Multi-Turn RL for LLM Agents Skill-SD: Skill-Conditioned Self-Distillation

self-reflection 2

ReflectRM Reflexion

knowledge 2

ROME / MEMIT WRAP++: Web Discovery Amplified Pretraining

scaling 2

Scalable MoE 预训练 Test-Time Scaling

self-evolving 2

Self-Evolving LLM Memory Extraction TTVS: Test-Time Variational Synthesis

lab 1

字节跳动 Seed

tree-based 1

DDTree

chatbot 1

Doubao App

ycombinator 1

Mem0: AI Agent 的可扩展长期记忆层

markov-decision-process 1

Agent-R1: End-to-End RL for LLM Agents

supply-chain 1

Agent 供应链攻击

api 1

Agent 供应链攻击

automated 1

Agent² RL-Bench: Agentic RL Post-Training

research 1

Agentic Research Ideas

workflow 1

Agentic Workflows

game-theory 1

AI Self-Awareness via Game Theory

self-awareness 1

AI Self-Awareness via Game Theory

frontend 1

声学前端与后端 ASR 的平衡（Audio Frontend-Backend Balance）

signal-processing 1

声学前端与后端 ASR 的平衡（Audio Frontend-Backend Balance）

robustness 1

声学前端与后端 ASR 的平衡（Audio Frontend-Backend Balance）

omni 1

Audio-Omni: 统一音频理解+生成+编辑

editing 1

Audio-Omni: 统一音频理解+生成+编辑

understanding 1

Audio-Omni: 统一音频理解+生成+编辑

rectified-flow 1

Audio-Omni: 统一音频理解+生成+编辑

kv-cache 1

AudioKV: 音频大模型 KV Cache 驱逐

backdoor 1

Backdoor Awareness

scalable 1

BlendFusion: 扩散模型合成数据生成

runtime 1

ClawGuard: Runtime Security for Tool-Augmented Agents

indirect-prompt 1

ClawGuard: Runtime Security for Tool-Augmented Agents

adversarial 1

Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning

learning 1

Complementary Learning Systems

calibration 1

Process Supervision of Confidence Margin for Calibrated LLM Reasoning

data-laundering 1

Combating Data Laundering in LLM Training

llm-training 1

Combating Data Laundering in LLM Training

decomposition 1

A Decomposition Perspective to Long-Context Reasoning for LLMs

cascaded 1

DuplexCascade

misalignment 1

Emergent Misalignment

self-learning 1

Epistemologically-Informed LLM Self-Learning

epistemology 1

Epistemologically-Informed LLM Self-Learning

dialogue 1

Epistemologically-Informed LLM Self-Learning

spatial-reasoning 1

Faithful GRPO

reward-shaping 1

Gaussian GRPO (G2RPO)

foundation 1

Generalized Knowledge Distillation (GKD)

deepmind 1

Generalized Knowledge Distillation (GKD)

simulation 1

Generative Agents

hypergraph 1

HyperMem: 超图记忆架构

noise 1

An Imperfect Verifier is Good Enough: Learning with Noisy Rewards

slm 1

LightMem: 轻量级 Agent 记忆系统

lightweight 1

LightMem: 轻量级 Agent 记忆系统

off-policy 1

LLM Post-Training: A Unified View of Off-Policy and On-Policy Learning

gpu 1

LLM-PRISM: GPU Faults & Silent Data Corruption

hardware 1

LLM-PRISM: GPU Faults & Silent Data Corruption

silent-corruption 1

LLM-PRISM: GPU Faults & Silent Data Corruption

compression 1

LLM Training as Lossy Compression

information-bottleneck 1

LLM Training as Lossy Compression

react 1

MemReader: 从被动到主动的记忆提取

extraction 1

MemReader: 从被动到主动的记忆提取

theory-of-mind 1

Mental Self-Modeling

metacognition 1

Mental Self-Modeling

real-time 1

MiniCPM-o 4.5

edge-deployment 1

MiniCPM-o 4.5

mixture-of-experts 1

Mixture-of-Layers Distillation with Stepwise Attention on Key Information

autonomous-ml 1

ML-Agent: Autonomous ML via Reinforcement Learning

kyutai 1

MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models

moshi 1

MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models

coordination 1

Multi-Agent Coordination

latent 1

Multimodal Latent Reasoning via Predictive Embeddings

continuous 1

Multimodal Latent Reasoning via Predictive Embeddings

interleaved 1

Omni-R1: 统一生成式多模态推理

visual-reasoning 1

Omni-R1: 统一生成式多模态推理

self-supervised 1

OmniJigsaw

video 1

OmniJigsaw

gkd 1

On-Policy Distillation for Autonomous Vehicle Motion Planning

analysis 1

Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw

rubrics 1

Optimsyn: Influence-Guided Rubrics for Synthetic Data

influence-function 1

Optimsyn: Influence-Guided Rubrics for Synthetic Data

oracle 1

ORACLE-SWE

pipeline 1

PARM: Pipeline-Adapted Reward Model

proactive 1

PASK: 意图感知的主动 Agent

intent 1

PASK: 意图感知的主动 Agent

vlm 1

Perception-Grounded Policy Optimization for VLMs

perception 1

Perception-Grounded Policy Optimization for VLMs

world-model 1

PerlAD: Pseudo-simulation-based RL for E2E Autonomous Driving

planning 1

PerlAD: Pseudo-simulation-based RL for E2E Autonomous Driving

robotics 1

PerlAD: Pseudo-simulation-based RL for E2E Autonomous Driving

data-augmentation 1

PieceHint: Question Augmentation Framework for RL

reward 1

PieceHint: Question Augmentation Framework for RL

exploration 1

Policy Split: Dual-Mode Exploration

dual-mode 1

Policy Split: Dual-Mode Exploration

ppo 1

Proximal Policy Distillation

entropy 1

Pseudo-Unification: 熵探测揭示多模态模型信息分歧

unification 1

Pseudo-Unification: 熵探测揭示多模态模型信息分歧

renyi-entropy 1

Pseudo-Unification: 熵探测揭示多模态模型信息分歧

rkhs 1

Pseudo-Unification: 熵探测揭示多模态模型信息分歧

vector-search 1

Qdrant Vector Search

starpo 1

RAGEN: Multi-Turn RL for LLM Agents

paradigm 1

Rethinking Software Engineering for Agentic AI

privacy 1

RL-Guided Synthetic Data Generation

math-reasoning 1

RL Post-Training Scaling Laws for LLMs

model-editing 1

ROME / MEMIT

self-auditing 1

SAVeR: Self-Audited Verified Reasoning

belief 1

SAVeR: Self-Audited Verified Reasoning

moe 1

Scalable MoE 预训练

distributed 1

Scalable MoE 预训练

supercomputer 1

Scalable MoE 预训练

adaptive-weighting 1

SCOPE: Signal-Calibrated On-Policy Distillation Enhancement

dual-path 1

SCOPE: Signal-Calibrated On-Policy Distillation Enhancement

heterogeneous 1

Self-Evolving LLM Memory Extraction

skill-conditioned 1

Skill-SD: Skill-Conditioned Self-Distillation

decoding 1

Speculative Decoding

natural-instructions 1

SUPERNOVA: 自然指令 RL 激发通用推理

generalization 1

SUPERNOVA: 自然指令 RL 激发通用推理

self-verification 1

SVSR: 多模态推理的自验证自修正

self-rectification 1

SVSR: 多模态推理的自验证自修正

context-management 1

SWE-AGILE: Dynamic Reasoning Context

dynamic 1

SWE-AGILE: Dynamic Reasoning Context

execution 1

SWE-HERO: 从无执行到有执行的 SWE Agent 训练

prm 1

SWE-Shepherd: PRMs for Code Agents

code-agent 1

SWE-Shepherd: PRMs for Code Agents

trustworthiness 1

The Expense of Seeing: Trustworthy Multimodal Reasoning Within Monolithic Paradigm

test-time 1

TTVS: Test-Time Variational Synthesis

verification 1

VerificAgent

latent-reasoning 1

Visual Enhanced Depth Scaling for Multimodal Latent Reasoning

depth-scaling 1

Visual Enhanced Depth Scaling for Multimodal Latent Reasoning

encoder 1

Whisper-AuT: Domain-Adapted Audio Encoder

domain-adaptation 1

Whisper-AuT: Domain-Adapted Audio Encoder

efficient-training 1

Whisper-AuT: Domain-Adapted Audio Encoder

web 1

WRAP++: Web Discovery Amplified Pretraining

cross-document 1

WRAP++: Web Discovery Amplified Pretraining