The Year in AI Papers
The year/December 2025

Monthly collection

December 2025

The papers that drew the most attention during this month of AI research.

Papers
66
Citations
2.1K
Research labs
6

Explore the year

View all months →
Aug2025, 63 papersSep2025, 79 papersOct2025, 98 papersNov2025, 73 papersDec2025, current month, 66 papersJan2026, 86 papersFeb2026, 86 papersMar2026, 100 papersApr2026, 77 papersMay2026, 115 papersJun2026, 86 papersJul2026, 84 papersAug2026, 3 papers
1,016papers
curated and indexed

66 papers published in December 2025

01

DeepSeek

mHC: Manifold-Constrained Hyper-Connections

The paper introduces Manifold-Constrained Hyper-Connections (mHC), a framework that addresses training instability and scalability issues in Hyper-Connections (HC) by projecting residual connection matrices onto the Birkhoff polytope (doubly stochastic matrices) using the Sinkhorn-Knopp algorithm. This restores the identity mapping property, ensuring norm…

Zhenda Xie, Yixuan Wei, Huanqi Cao, Chenggang Zhao, et al.
Published
Dec 2025
Upvotes
331
Citations
65
Read summary Original
02

DeepSeek

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

DeepSeek-V3.2 introduces three key innovations to close the gap with closed-source models: DeepSeek Sparse Attention (DSA) reduces computational complexity from O(L^2) to O(Lk) while preserving long-context performance; a scalable reinforcement learning (RL) framework with a post-training compute budget exceeding 10% of pre-training cost enables…

DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, et al.
Published
Dec 2025
Upvotes
271
Citations
671
Read summary Original
03

arXiv.org

DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI

DataFlow is a unified, LLM-driven framework for data preparation in the era of data-centric AI, addressing the fragmentation of ad-hoc scripts and workflows. It provides a PyTorch-style API with nearly 200 reusable operators and six domain-general pipelines (text, math, code, Text-to-SQL, agentic RAG, knowledge extraction). The framework includes a global…

Hao Liang, Xiaochen Ma, Zhou Liu, Zhen Hao Wong, et al.
Published
Dec 2025
Upvotes
226
Citations
19
Read summary Original
04

arXiv.org

Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

Live Avatar is an algorithm-system co-designed framework enabling real-time, streaming, and infinite-length audio-driven avatar generation using a 14-billion-parameter diffusion model. It addresses two key challenges: long-horizon consistency and the real-time-fidelity trade-off. The algorithm side uses a two-stage pipeline (Diffusion Forcing pretraining…

Yubo Huang, Hailong Guo, Fangtai Wu, Weiqiang Wang, et al.
Published
Dec 2025
Upvotes
179
Citations
0
Read summary Original
05

arXiv.org

Kling-Omni Technical Report

Kling-Omni is a generalist generative framework from Kuaishou Technology that unifies video generation, editing, and reasoning into a single end-to-end system. It introduces Multi-modal Visual Language (MVL) as an interaction paradigm, combining text, images, and videos into a unified representation. The architecture includes a Prompt Enhancer (PE) based…

Kling Team, Jialu Chen, Yuanzheng Ci, Xiangyu Du, et al.
Published
Dec 2025
Upvotes
174
Citations
41
Read summary Original
06

arXiv.org

DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle

DAComp is a benchmark of 210 tasks evaluating LLM-based agents on enterprise data intelligence, covering repository-level data engineering (DE) and open-ended data analysis (DA). DE tasks include architecture, implementation, and evolution, requiring multi-stage SQL pipelines on schemas averaging 412 columns and over 2,000 lines of code. DA tasks involve…

Fangyu Lei, Jinxiang Meng, Yiming Huang, Junjie Zhao, et al.
Published
Dec 2025
Upvotes
160
Citations
11
Read summary Original
07

arXiv.org

Memory in the Age of AI Agents

This survey provides a comprehensive overview of memory systems for foundation model-based AI agents, addressing the fragmentation in the field by proposing a unified taxonomy based on forms, functions, and dynamics. It defines agent memory, distinguishing it from LLM memory, RAG, and context engineering. The survey categorizes memory forms into…

Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, et al.
Published
Dec 2025
Upvotes
160
Citations
232
Read summary Original
08

Research paper

Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models

Youtu-LLM is a 1.96B-parameter language model pre-trained from scratch to achieve native agentic intelligence, balancing efficiency with strong reasoning and planning. It uses a dense Multi-Latent Attention (MLA) architecture with a 128k context window and a STEM-oriented tokenizer. Training follows a 'Commonsense-STEM-Agent' curriculum over 10.84T tokens,…

Junru Lu, Jiarui Qin, Lingfeng Qiao, Yinghui Li, et al.
Published
Dec 2025
Upvotes
155
Citations
5
Read summary Original
09

arXiv.org

Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance

Wan-Move is a framework for motion-controllable video generation that enhances existing image-to-video (I2V) models without adding auxiliary modules. It represents object motion using dense point trajectories, which are transferred into latent space and used to replicate first-frame features along each trajectory, creating a motion-aware condition feature…

Ruihang Chu, Yefei He, Zhekai Chen, Shiwei Zhang, et al.
Published
Dec 2025
Upvotes
134
Citations
37
Read summary Original
10

arXiv.org

Step-GUI Technical Report

This technical report introduces Step-GUI, a family of multimodal GUI agent models (4B/8B) built on Qwen3-VL, achieving state-of-the-art performance across benchmarks (8B: 80.2% AndroidWorld, 48.5% OSWorld, 62.6% ScreenShot-Pro). The core innovation is the Calibrated Step Reward System (CSRS), a self-evolving training pipeline that converts model-generated…

Haolong Yan, Jia Wang, Xin Huang, Yeqing Shen, et al.
Published
Dec 2025
Upvotes
134
Citations
27
Read summary Original
11

arXiv.org

EgoX: Egocentric Video Generation from a Single Exocentric Video

EgoX is a novel framework that generates egocentric (first-person) videos from a single exocentric (third-person) video input. It leverages a pretrained video diffusion model (Wan 2.1 14B) with lightweight LoRA adaptation, avoiding the need for additional inputs like multiple views or initial frames. The method lifts the exocentric video into a 3D point…

Taewoong Kang, Kinam Kim, Dohyeon Kim, Minho Park, et al.
Published
Dec 2025
Upvotes
124
Citations
6
Read summary Original
12

Research paper

Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows

This paper introduces SGI-Bench, a benchmark for evaluating Scientific General Intelligence (SGI) in large language models (LLMs). SGI is defined as an AI's ability to autonomously navigate the complete, iterative cycle of scientific inquiry, grounded in the Practical Inquiry Model (PIM) with four quadrants: Deliberation, Conception, Action, and…

Wanghan Xu, Yuhao Zhou, Yifan Zhou, Qinglong Cao, et al.
Published
Dec 2025
Upvotes
122
Citations
21
Read summary Original
13

arXiv.org

T-pro 2.0: An Efficient Russian Hybrid-Reasoning Model and Playground

T-pro 2.0 is an open-weight Russian LLM for hybrid reasoning and efficient inference, supporting direct answering and reasoning-trace generation. It uses a Cyrillic-dense tokenizer and an adapted EAGLE speculative-decoding pipeline to reduce latency. The authors release the model weights, the T-Wix 500k instruction corpus, the T-Math reasoning benchmark,…

Dmitrii Stoianov, Danil Taranets, Olga Tsymboi, Ramil Latypov, et al.
Published
Dec 2025
Upvotes
121
Citations
1
Read summary Original
14

arXiv.org

MMGR: Multi-Modal Generative Reasoning

The paper introduces MMGR (Multi-Modal Generative Reasoning), a benchmark suite to evaluate the reasoning capabilities of video and image generation models across five core abilities: Physical, Logical, 3D Spatial, 2D Spatial, and Temporal reasoning. It comprises three domains: Abstract Reasoning (Maze, Sudoku, ARC-AGI, Math), Embodied Navigation (four…

Zefan Cai, Haoyi Qiu, Tianyi Ma, Haozhe Zhao, et al.
Published
Dec 2025
Upvotes
121
Citations
9
Read summary Original
15

arXiv.org

Youtu-Agent: Scaling Agent Productivity with Automated Generation and Hybrid Policy Optimization

Youtu-Agent is a modular framework addressing high configuration costs and static capabilities in LLM agents. It decouples environments, toolkits, and agents via a YAML-based system, enabling automated generation through a deterministic Workflow mode and a flexible Meta-Agent mode that synthesizes tools, prompts, and configurations. For continuous…

Yuchen Shi, Yuzheng Cai, Siqi Cai, Zihan Xu, et al.
Published
Dec 2025
Upvotes
119
Citations
5
Read summary Original
16

arXiv.org

Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding

The paper introduces Mindscape-Aware RAG (MiA-RAG), a framework that equips LLM-based RAG systems with a global semantic representation, called a mindscape, to improve long-context understanding. The mindscape is built via hierarchical summarization of a document. MiA-RAG conditions both retrieval and generation on this mindscape: a retriever (MiA-Emb)…

Yuqing Li, Jiangnan Li, Zheng Lin, Ziyan Zhou, et al.
Published
Dec 2025
Upvotes
115
Citations
6
Read summary Original
17

arXiv.org

QwenLong-L1.5: Post-Training Recipe for Long-Context Reasoning and Memory Management

QwenLong-L1.5 is a long-context reasoning model built on Qwen3-30B-A3B-Thinking, introduced by Alibaba's Tongyi Lab. It achieves performance comparable to GPT-5 and Gemini-2.5-Pro on long-context benchmarks, surpassing its baseline by 9.90 points on average. The model's post-training recipe includes three key innovations: (1) a data synthesis pipeline that…

Weizhou Shen, Ziyi Yang, Chenliang Li, Zhiyuan Lu, et al.
Published
Dec 2025
Upvotes
113
Citations
23
Read summary Original
18

Research paper

HGMEM: Hypergraph-based Working Memory to Improve Multi-step RAG for Long-Context Complex Relational Modeling

HGMEM is a hypergraph-based working memory system for multi-step retrieval-augmented generation (RAG) that improves complex relational modeling in long contexts. Unlike existing memory mechanisms that passively store isolated facts, HGMEM represents memory as a hypergraph where hyperedges serve as memory points, enabling the progressive formation of…

Chulun Zhou, Chunkang Zhang, Guoxin Yu, Fandong Meng, et al.
Published
Dec 2025
Upvotes
112
Citations
4
Read summary Original
19

arXiv.org

Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills

This survey examines how agentic AI systems are adapted after pretraining, organizing the field into a four-paradigm framework based on what is optimized (the agent or its tools) and the source of the adaptation signal (tool execution or agent output). The paradigms are: A1 (tool-execution-signaled agent adaptation), A2 (agent-output-signaled agent…

Pengcheng Jiang, Jiacheng Lin, Zhiyi Shi, Zifeng Wang, et al.
Published
Dec 2025
Upvotes
111
Citations
15
Read summary Original
20

Qwen

Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem

This technical report introduces the Agentic Learning Ecosystem (ALE), a full-stack infrastructure for developing agentic LLMs, and ROME, an open-source agent model trained within it. ALE comprises three components: ROLL, a scalable RL training framework; ROCK, a sandboxed environment execution engine; and iFlow CLI, an agent framework for context…

Weixun Wang, XiaoXiao Xu, Wanhe An, Fangwen Dai, et al.
Published
Dec 2025
Upvotes
110
Citations
23
Read summary Original
21

MiniMax

Towards Scalable Pre-training of Visual Tokenizers for Generation

The paper introduces VTP, a visual tokenizer pre-training framework that integrates image-text contrastive learning, self-supervised learning (MIM and self-distillation), and reconstruction losses to address the 'pre-training scaling problem' in latent diffusion models. The authors argue that reconstruction-only training biases the latent space toward…

Jingfeng Yao, Yuda Song, Yucong Zhou, Xinggang Wang
Published
Dec 2025
Upvotes
108
Citations
23
Read summary Original
22

Qwen

Stabilizing Reinforcement Learning with LLMs: Formulation and Practices

This paper proposes a formulation for reinforcement learning (RL) with large language models (LLMs), showing that optimizing sequence-level rewards via token-level objectives is a first-order approximation that holds only when training–inference discrepancy and policy staleness are minimized. The authors introduce MiniRL, a minimalist REINFORCE-based…

Chujie Zheng, Kai Dang, Bowen Yu, Mingze Li, et al.
Published
Dec 2025
Upvotes
107
Citations
35
Read summary Original
23

arXiv.org

Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary Loss

The paper introduces the expert-router coupling (ERC) loss, a lightweight auxiliary loss for Mixture-of-Experts (MoE) models that aligns router decisions with expert capabilities. The method treats each router embedding as a proxy token for the tokens routed to that expert, perturbs it with bounded noise, and feeds it through all experts to compute…

Ang Lv, Jin Ma, Yiyuan Ma, Siyuan Qiao
Published
Dec 2025
Upvotes
100
Citations
9
Read summary Original
24

Research paper

InsertAnywhere: Geometrically Grounded and Optics-Aware Video Object Insertion

InsertAnywhere is a framework for video object insertion (VOI) that addresses limitations in 4D scene understanding and optical effects. It uses a two-stage pipeline: first, a 4D-aware mask generation module reconstructs the video into a 4D scene, allowing users to anchor an object's 3D pose in one frame, then propagates it via scene flow tracking to…

Hoiyeong Jin, Hyojin Jang, Junha Hyung, Jeongho Kim, et al.
Published
Dec 2025
Upvotes
99
Citations
0
Read summary Original
25

arXiv.org

Recursive Language Models

The paper introduces Recursive Language Models (RLMs), an inference-time framework that treats arbitrarily long prompts as part of an external environment, allowing an LLM to programmatically examine, decompose, and recursively call itself over prompt snippets. RLMs use a REPL environment where the prompt is stored as a variable, and the model writes code…

Alex L. Zhang, Tim Kraska, Omar Khattab
Published
Dec 2025
Upvotes
99
Citations
64
Read summary Original
26

arXiv.org

TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times

TurboDiffusion is a video generation acceleration framework that achieves 100–200× end-to-end speedup while maintaining video quality. It combines four main techniques: low-bit SageAttention for attention acceleration, Sparse-Linear Attention (SLA) for sparse attention, rCM for step distillation, and W8A8 quantization for linear layers. Training involves…

Jintao Zhang, Kaiwen Zheng, Kai Jiang, Haoxu Wang, et al.
Published
Dec 2025
Upvotes
96
Citations
28
Read summary Original
27

arXiv.org

SemanticGen: Video Generation in Semantic Space

SemanticGen is a novel video generation framework that operates in a compact semantic space rather than directly in the VAE latent space. It uses a two-stage process: first, a diffusion model generates compressed semantic video features (using Qwen-2.5-VL as the semantic encoder) that define the global layout; second, another diffusion model generates VAE…

Jianhong Bai, Xiaoshi Wu, Xintao Wang, Xiao Fu, et al.
Published
Dec 2025
Upvotes
95
Citations
7
Read summary Original
28

arXiv.org

Envision: Benchmarking Unified Understanding & Generation for Causal World Process Insights

The paper introduces Envision, a benchmark for evaluating text-to-image (T2I) and unified multimodal models (UMMs) on causal, multi-image event generation. It addresses the limitation of static single-image benchmarks by proposing chained text-to-multi-image generation with 1,000 four-stage prompts across six scientific and humanities domains.…

Juanxi Tian, Siyuan Li, Conghui He, Lijun Wu, et al.
Published
Dec 2025
Upvotes
94
Citations
1
Read summary Original
29

arXiv.org

ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding

ReFusion is a masked diffusion language model that integrates sequence reorganization into a causal attention framework to overcome the limitations of standard masked diffusion models (MDMs), which suffer from high computational overhead due to the lack of Key-Value (KV) caching and incoherent generation from intractable token combination learning.…

Jia-Nan Li, Jian Guan, Wei Wu, Chongxuan Li
Published
Dec 2025
Upvotes
93
Citations
14
Read summary Original
30

arXiv.org

Next-Embedding Prediction Makes Strong Vision Learners

The paper introduces Next-Embedding Predictive Autoregression (NEPA), a self-supervised visual pretraining method that trains a Vision Transformer to predict future patch embeddings from past ones, using causal masking and stop-gradient, without pixel reconstruction, discrete tokens, or contrastive losses. NEPA achieves 83.8% and 85.3% top-1 accuracy on…

Sihan Xu, Ziqiao Ma, Wenhao Chai, Xuweiyi Chen, et al.
Published
Dec 2025
Upvotes
91
Citations
13
Read summary Original
31

arXiv.org

LLaDA2.0: Scaling Up Diffusion Language Models to 100B

LLaDA2.0 introduces a family of discrete diffusion language models (dLLMs) scaled up to 100B parameters by converting pre-trained auto-regressive (AR) models, avoiding costly training from scratch. The conversion uses a three-phase Warmup-Stable-Decay (WSD) continual pre-training strategy: progressively increasing block size in block diffusion (warmup),…

Tiwei Bie, Maosong Cao, Kun Chen, Lun Du, et al.
Published
Dec 2025
Upvotes
89
Citations
123
Read summary Original
32

arXiv.org

Step-DeepResearch Technical Report

Step-DeepResearch is a 32B-parameter, end-to-end Deep Research agent model developed by StepFun. It uses a data synthesis strategy based on atomic capabilities (planning, information seeking, reflection, and report writing) and a progressive training pipeline (agentic mid-training, SFT, RL) with a Checklist-style Judger reward. The model achieves a score…

Chen Hu, Haikuo Du, Heng Wang, Lin Lin, et al.
Published
Dec 2025
Upvotes
89
Citations
11
Read summary Original
33

Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers

MACE-Dance: Motion-Appearance Cascaded Experts for Music-Driven Dance Video Generation

MACE-Dance is a music-driven dance video generation framework using a cascaded Mixture-of-Experts (MoE) design, decoupling the task into a Motion Expert and an Appearance Expert. The Motion Expert generates 3D SMPL motion from music using a Diffusion Model with a BiMamba-Transformer hybrid architecture and Guidance-Free Training (GFT), achieving…

Kaixing Yang, Jiashu Zhu, Xulong Tang, Ziqiao Peng, et al.
Published
Dec 2025
Upvotes
88
Citations
9
Read summary Original
34

arXiv.org

Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits

The paper introduces Gnosis, a lightweight self-awareness mechanism that enables frozen large language models (LLMs) to predict their own failures by decoding signals from internal hidden states and attention patterns during inference. Gnosis compresses these internal traces into fixed-budget descriptors using a dual-stream architecture (hidden-state and…

Amirhosein Ghasemabadi, Di Niu
Published
Dec 2025
Upvotes
87
Citations
10
Read summary Original
35

arXiv.org

Nex-N1: Agentic Models Trained via a Unified Ecosystem for Large-Scale Environment Construction

The paper introduces Nex-N1, a series of agentic models trained using a unified ecosystem (NexAU, NexA4A, NexGAP) for large-scale environment construction. NexAU is a modular runtime for scalable agent frameworks, NexA4A automatically generates diverse agent hierarchies from natural language, and NexGAP generates end-to-end agentic trajectories using real…

Nex-AGI Team, :, Yuxuan Cai, Lu Chen, et al.
Published
Dec 2025
Upvotes
85
Citations
15
Read summary Original
36

arXiv.org

Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning

The paper introduces Native Parallel Reasoner (NPR), a teacher-free framework that enables large language models to develop genuine parallel reasoning capabilities through self-distillation and reinforcement learning. NPR uses a three-stage progressive training paradigm: Stage 1 applies format-following RL (DAPO) to induce a structured parallel format,…

Tong Wu, Yang Liu, Jun Bai, Zixia Jia, et al.
Published
Dec 2025
Upvotes
80
Citations
11
Read summary Original
37

Meta AI

TUNA: Taming Unified Visual Representations for Native Unified Multimodal Models

Tuna is a native unified multimodal model (UMM) that creates a unified continuous visual representation by cascading a VAE encoder with a representation encoder (SigLIP 2). This design avoids the representation format mismatches of decoupled models, improving both understanding and generation. The model uses an LLM decoder (Qwen2.5) for autoregressive text…

Zhiheng Liu, Weiming Ren, Haozhe Liu, Zijian Zhou, et al.
Published
Dec 2025
Upvotes
78
Citations
31
Read summary Original
38

arXiv.org

Visionary: The World Model Carrier Built on WebGPU-Powered Gaussian Splatting Platform

Visionary is an open, web-native platform for real-time rendering of 3D Gaussian Splatting (3DGS) and meshes, built on WebGPU and ONNX. It introduces a Gaussian Generator contract, a standardized ONNX I/O interface that allows plug-and-play integration of various 3DGS algorithms (e.g., MLP-based 3DGS, 4DGS, neural avatars) without modifying the renderer.…

Yuning Gong, Yifei Liu, Yifan Zhan, Muyao Niu, et al.
Published
Dec 2025
Upvotes
77
Citations
2
Read summary Original
39

arXiv.org

TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows

TWINFLOW is a framework for training one-step generative models on large multimodal models, addressing the inefficiency of multi-step diffusion and flow matching models that require 40-100 NFEs. Existing few-step methods either rely on auxiliary trained models (e.g., GAN discriminators) or frozen teachers, causing instability and memory overhead, or…

Zhenglin Cheng, Peng Sun, Jianguo Li, Tao Lin
Published
Dec 2025
Upvotes
76
Citations
16
Read summary Original
40

arXiv.org

LongVie 2: Multimodal Controllable Ultra-Long Video World Model

LongVie 2 is an end-to-end autoregressive framework for controllable ultra-long video generation, extending pretrained diffusion backbones (Wan2.1-I2V-14B) into a video world model. It is trained in three progressive stages: (1) multi-modal guidance integrating dense (depth maps) and sparse (point maps) control signals via a ControlNet-style architecture…

Jianxiong Gao, Zhaoxi Chen, Xian Liu, Junhao Zhuang, et al.
Published
Dec 2025
Upvotes
76
Citations
6
Read summary Original
41

arXiv.org

PhysBrain: Human Egocentric Data as a Bridge from Vision Language Models to Physical Intelligence

The paper introduces PhysBrain, a model that bridges vision-language models (VLMs) and physical intelligence for robotics by using human egocentric videos. The authors propose the Egocentric2Embodiment (E2E) Translation Pipeline, which converts raw egocentric videos into multi-level, schema-driven VQA supervision with evidence grounding and temporal…

Xiaopeng Lin, Shijie Lian, Bin Yu, Ruoqi Yang, et al.
Published
Dec 2025
Upvotes
76
Citations
8
Read summary Original
42

arXiv.org

StereoWorld: Geometry-Aware Monocular-to-Stereo Video Generation

StereoWorld is an end-to-end diffusion-based framework that converts monocular videos into high-fidelity stereo videos by adapting a pretrained video generator. It conditions the model on the left-view video and uses a geometry-aware regularization combining disparity and depth supervision to ensure 3D structural fidelity. A spatio-temporal tiling scheme…

Ke Xing, Xiaojie Jin, Longfei Li, Yuyang Yin, et al.
Published
Dec 2025
Upvotes
74
Citations
2
Read summary Original
43

The Web Conference

PaperDebugger: A Plugin-Based Multi-Agent System for In-Editor Academic Writing, Review, and Editing

PaperDebugger is an in-editor, multi-agent, plugin-based academic writing assistant that integrates directly into Overleaf via a Chrome extension, eliminating copy-paste workflows. It addresses the challenge of external LLM assistants by providing bidirectional synchronization, version control, secure state management, and multi-agent scheduling. The…

Junyi Hou, Andre Lin Huikai, Nuo Chen, Yiwei Gong, et al.
Published
Dec 2025
Upvotes
73
Citations
9
Read summary Original
44

arXiv.org

WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling

WorldPlay is a real-time interactive world model that generates streaming 720p video at 24 FPS while maintaining long-term geometric consistency. It addresses the trade-off between speed and memory in existing methods. The model uses three key components: Dual Action Representation combining discrete keys and continuous camera poses for robust control;…

Wenqiang Sun, Haiyu Zhang, Haoyuan Wang, Junta Wu, et al.
Published
Dec 2025
Upvotes
72
Citations
97
Read summary Original
45

Qwen

Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition

Qwen-Image-Layered is an end-to-end diffusion model that decomposes a single RGB image into multiple semantically disentangled RGBA layers, enabling consistent image editing where each layer can be independently manipulated. The model introduces three key components: an RGBA-VAE that unifies latent representations for RGB and RGBA images, a VLD-MMDiT…

Shengming Yin, Zekai Zhang, Zecheng Tang, Kaiyuan Gao, et al.
Published
Dec 2025
Upvotes
71
Citations
24
Read summary Original
46

arXiv.org

Latent Implicit Visual Reasoning

The paper introduces Latent Implicit Visual Reasoning (LIVR), a method to enhance visual reasoning in Large Multimodal Models (LMMs) without explicit intermediate supervision. LIVR adds latent tokens to the model's vocabulary and trains them via a two-stage visual bottlenecking approach: Stage 1 forces visual information to pass through latent tokens by…

Kelvin Li, Chuyi Shang, Leonid Karlinsky, Rogerio Feris, et al.
Published
Dec 2025
Upvotes
70
Citations
8
Read summary Original
47

AAAI Conference on Artificial Intelligence

Robust-R1: Degradation-Aware Reasoning for Robust Visual Understanding

Robust-R1 is a framework designed to improve the robustness of Multimodal Large Language Models (MLLMs) against real-world visual degradations. Unlike existing methods that rely on implicit training or adaptation of visual encoders, Robust-R1 explicitly models degradations through structured reasoning chains. The approach consists of three stages:…

Jiaqi Tang, Jianmin Chen, Wei Wei, Xiaogang Xu, et al.
Published
Dec 2025
Upvotes
68
Citations
8
Read summary Original
48

arXiv.org

The Prism Hypothesis: Harmonizing Semantic and Pixel Representations via Unified Autoencoding

The paper introduces the Prism Hypothesis, which posits that multimodal data can be understood through a shared frequency spectrum: semantic encoders capture low-frequency components (abstract meaning), while pixel encoders retain high-frequency details (fine texture). This is supported by experiments showing that text-image retrieval relies on low…

Weichen Fan, Haiwen Diao, Quan Wang, Dahua Lin, et al.
Published
Dec 2025
Upvotes
68
Citations
15
Read summary Original
49

arXiv.org

Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space

The paper introduces Dynamic Large Concept Models (DLCM), a hierarchical language modeling framework that learns variable-length semantic concepts from latent representations and performs reasoning in a compressed concept space, shifting computation from tokens to concepts. DLCM uses a four-stage pipeline: encoding, dynamic segmentation via learned…

Xingwei Qu, Shaowen Wang, Zihao Huang, Kai Hua, et al.
Published
Dec 2025
Upvotes
67
Citations
12
Read summary Original
50

arXiv.org

Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies

This paper introduces Bottom-up Policy Optimization (BuPO), a novel reinforcement learning (RL) paradigm for large language models (LLMs) that optimizes internal layer policies before the full language model policy. The authors decompose the LLM policy into Internal Layer Policies and Internal Modular Policies via the Transformer's residual stream. Entropy…

Yuqiao Tan, Minzheng Wang, Shizhu He, Huanxuan Liao, et al.
Published
Dec 2025
Upvotes
66
Citations
21
Read summary Original
51

arXiv.org

LiveTalk: Real-Time Multimodal Interactive Video Diffusion via Improved On-Policy Distillation

The paper introduces LiveTalk, a real-time multimodal interactive video diffusion system. It addresses the high inference cost of diffusion models by distilling a bidirectional, many-step model into a causal, 4-step autoregressive one. The authors identify that the leading on-policy distillation method, Self Forcing, suffers from training instability and…

Ethan Chern, Zhulin Hu, Bohao Tang, Jiadi Su, et al.
Published
Dec 2025
Upvotes
66
Citations
7
Read summary Original
52

arXiv.org

MultiShotMaster: A Controllable Multi-Shot Video Generation Framework

MultiShotMaster is a framework for controllable multi-shot video generation, extending a pretrained single-shot text-to-video model. It introduces two RoPE variants: Multi-Shot Narrative RoPE, which applies phase shifts at shot boundaries for flexible shot arrangement while preserving narrative order, and Spatiotemporal Position-Aware RoPE, which enables…

Qinghe Wang, Xiaoyu Shi, Baolu Li, Weikang Bian, et al.
Published
Dec 2025
Upvotes
65
Citations
28
Read summary Original
53

Research paper

VideoASMR-Bench: Can AI-Generated ASMR Videos Fool VLMs and Humans?

VideoASMR-Bench is a new benchmark for evaluating AI-generated ASMR videos, focusing on fine-grained audio-visual perception and sensory immersion. It includes 1,500 real ASMR videos from social media and 2,235 synthetic videos from nine video generation models (VGMs) under four settings. The benchmark introduces an adversarial evaluation framework where…

Jiaqi Wang, Weijia Wu, Yi Zhan, Rui Zhao, et al.
Published
Dec 2025
Upvotes
65
Citations
2
Read summary Original
54

arXiv.org

Taming Hallucinations: Boosting MLLMs' Video Understanding via Counterfactual Video Generation

The paper addresses visual ungrounded hallucinations in Multimodal Large Language Models (MLLMs), which over-rely on language priors when processing counterfactual videos that defy common sense. To mitigate this, the authors introduce DualityForge, a framework using diffusion-based controllable video editing to transform real-world videos into…

Zhe Huang, Hao Wen, Aiming Hao, Bingze Song, et al.
Published
Dec 2025
Upvotes
63
Citations
6
Read summary Original
55

arXiv.org

When Reasoning Meets Its Laws

The paper introduces the Laws of Reasoning (LORE), a framework formalizing desired reasoning behaviors in Large Reasoning Models (LRMs). LORE posits a compute law (reasoning compute scales linearly with question complexity) and an accuracy law (accuracy decays exponentially with complexity). Since complexity is hard to measure, the laws are approximated…

Junyu Zhang, Yifan Sun, Tianang Leng, Jingyan Shen, et al.
Published
Dec 2025
Upvotes
62
Citations
2
Read summary Original
56

Google DeepMind

Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning

This paper introduces a method for hierarchical reinforcement learning (RL) in autoregressive models by discovering and using temporally-abstract actions within the model's internal representations. The authors show that autoregressive models pretrained on next-token prediction learn latent representations of subgoals in their residual stream activations.…

Seijin Kobayashi, Yanick Schimpf, Maximilian Schlegel, Angelika Steger, et al.
Published
Dec 2025
Upvotes
62
Citations
5
Read summary Original
57

arXiv.org

Yume-1.5: A Text-Controlled Interactive World Generation Model

Yume1.5 is a framework for generating interactive, continuous virtual worlds from a single image or text prompt, with keyboard-based control for person and camera movement. It addresses limitations in existing video diffusion models, such as limited generalizability, high latency, and insufficient text control. The framework introduces three core…

Xiaofeng Mao, Zhen Li, Chuanhao Li, Xiaojie Xu, et al.
Published
Dec 2025
Upvotes
61
Citations
52
Read summary Original
58

arXiv.org

Beyond Real: Imaginary Extension of Rotary Position Embeddings for Long-Context LLMs

The paper introduces RoPE++, an extension of Rotary Position Embeddings (RoPE) for large language models. Standard RoPE discards the imaginary component of the complex-valued attention score, losing phase information. RoPE++ re-incorporates this imaginary part as a new group of attention heads, computed in parallel with real attention. Two configurations…

Xiaoran Liu, Yuerong Song, Zhigeng Liu, Zengfeng Huang, et al.
Published
Dec 2025
Upvotes
60
Citations
1
Read summary Original
59

arXiv.org

How Far Are We from Genuinely Useful Deep Research Agents?

This paper introduces FINDER, a benchmark for evaluating Deep Research Agents (DRAs) on report generation, and DEFT, a failure taxonomy for diagnosing DRA errors. FINDER consists of 100 human-curated research tasks with 419 structured checklist items, refining the DeepResearch Bench by adding detailed prompts and checklists. DEFT, built via grounded theory…

Dingling Zhang, He Zhu, Jincheng Ren, Kangqi Song, et al.
Published
Dec 2025
Upvotes
58
Citations
9
Read summary Original
60

Volume 1

LongVideoAgent: Multi-Agent Reasoning with Long Videos

LongVideoAgent is a multi-agent framework for long-video question answering. A master LLM coordinates a grounding agent to localize question-relevant segments and a vision agent to extract targeted visual observations. The master agent plans with a step limit and is trained with reinforcement learning (GRPO) to encourage concise, correct, and efficient…

Runtao Liu, Ziyi Liu, Jiaqi Tang, Yue Ma, et al.
Published
Dec 2025
Upvotes
56
Citations
21
Read summary Original
61

arXiv.org

Guided Self-Evolving LLMs with Minimal Human Supervision

The paper introduces R-FEW, a guided self-evolving framework for large language models (LLMs) that uses minimal human supervision to achieve stable and controllable self-improvement. R-FEW addresses issues like concept drift and diversity collapse in unguided self-play methods (e.g., R-Zero) by integrating a few-shot grounded Challenger and an online…

Wenhao Yu, Zhenwen Liang, Chengsong Huang, Kishan Panaganti, et al.
Published
Dec 2025
Upvotes
55
Citations
31
Read summary Original
62

Annual Meeting of the Association for Computational Linguistics

Finch: Benchmarking Finance & Accounting across Spreadsheet-Centric Enterprise Workflows

FINCH (FinWorkBench) is a benchmark for evaluating AI agents on real-world, enterprise-grade finance and accounting workflows. It is built from authentic enterprise data, including Enron emails and spreadsheets, EUSES, and financial reports from institutions like the World Bank, covering 2000–2025. The dataset comprises 172 composite workflows with 384…

Haoyu Dong, Pengkun Zhang, Yan Gao, Xuanyu Dong, et al.
Published
Dec 2025
Upvotes
54
Citations
5
Read summary Original
63

arXiv.org

Seed-Prover 1.5: Mastering Undergraduate-Level Theorem Proving via Learning from Experience

Seed-Prover 1.5 is a formal theorem-proving system for Lean developed by ByteDance Seed AI4Math. It uses an agentic prover trained via large-scale reinforcement learning (RL) with tool integration (Lean verification, Mathlib search, Python execution) and a sketch model trained with rubric RL to bridge natural language proofs and Lean sketches. The system…

Jiangjie Chen, Wenxiang Chen, Jiacheng Du, Jinyi Hu, et al.
Published
Dec 2025
Upvotes
53
Citations
29
Read summary Original
64

arXiv.org

Region-Constraint In-Context Generation for Instructional Video Editing

ReCo is a novel framework for instruction-based video editing that uses in-context generation with region constraints. It concatenates source and target videos for joint denoising and introduces two regularization terms: latent-space regularization increases latent discrepancy in editing regions while reducing it in non-editing areas, and attention-space…

Zhongwei Zhang, Fuchen Long, Wei Li, Zhaofan Qiu, et al.
Published
Dec 2025
Upvotes
53
Citations
14
Read summary Original
65

arXiv.org

DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models

DiffThinker is a diffusion-based framework that introduces Generative Multimodal Reasoning, reformulating multimodal reasoning as a native image-to-image generative task rather than text-centric symbolic mapping. Built on Qwen-Image-Edit with Flow Matching and a Multimodal Diffusion Transformer, it directly produces solution images that are parsed into…

Zefeng He, Xiaoye Qu, Yafu Li, Tong Zhu, et al.
Published
Dec 2025
Upvotes
53
Citations
12
Read summary Original
66

OpenAI

Training LLMs for Honesty via Confessions

The paper proposes a method to train LLMs to produce 'confessions'—self-reports of compliance with instructions and policies—to improve honesty. Confession training adds a system message after the model's answer, requesting a structured report enumerating objectives, compliance analysis, and uncertainties. The confession reward is based solely on honesty…

Manas Joglekar, Jeremy Chen, Gabriel Wu, Jason Yosinski, et al.
Published
Dec 2025
Upvotes
0
Citations
17
Read summary Original