Ask questions across all of research
alphaXiv connects papers, researchers, and organizations, grounding the answer in the underlying work.
Prime Agent is an open-source, self-improving harness that extends language models' capabilities for complex, long-horizon tasks by integrating persistent execution environments, recursive subagents, and dynamic information management. It significantly improved performance on interactive reasoning benchmarks, such as increasing ARC-AGI-3 scores for Opus 5 from 30% to 95.5%, while enabling programmatic problem-solving and continual adaptation.
Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations
Stanford University researchers analyzed nearly 250,000 real-world human-AI conversations from Claude.ai to understand collaborative dynamics, finding that over half of interactions involve consequential work and are human-led, with users actively engaging in adaptation and learning despite frequent friction.
The Code World Model integrates large language models as coding agents with video models to construct open-ended virtual worlds. This approach functionally separates logical world evolution from high-fidelity visual realization, enabling consistent, causally complex environments and precise spatiotemporal visual control via an intermediate "proxy" representation.
Researchers to follow
View allAre you a researcher? Find your profile
Zero-WAM introduces an in-context world-action model that leverages human videos for open-ended robotic task generalization, achieving a 46.95% average success rate on seven unseen simulation tasks and outperforming leading baselines by up to 35.97 percentage points. This framework enables robots to perform novel manipulation tasks by interpreting human demonstrations without requiring task-specific robot data or model parameter updates.
The DiffusionOPSD Team introduces an on-policy self-distillation framework that explicitly converts image-level reward into bounded, continually refreshed intermediate targets for diffusion models. This approach achieved the best held-out scores in 19 out of 20 reward-matched settings, showing up to 44.0% performance gains over competitors while reducing training GPU-hours by 40-63%.
LeFlow introduces a generative latent flow planning framework that amortizes action planning within world models, replacing costly iterative optimization with a learned trajectory prior. The framework achieves higher success rates and demonstrates an order-of-magnitude speedup (up to 14.4x faster) compared to traditional iterative optimization methods across various pixel-control benchmarks.
Visual General Intelligence (VGI) is proposed as a pathway to AGI, defining it as the capacity to acquire usable world knowledge directly from visual experience and apply it to tasks like prediction, imagination, and action. This work synthesizes expert perspectives, advocating for large-scale generative models that recover the underlying physical structure of the world, learn continually, and engage in embodied interaction.
Z.ai introduces GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, with 320B total and 18B active parameters. It combines linear and sparse attention in a hybrid architecture, adds Manifold-Constrained Hyper-Connections, and trains on a 30T-token multimodal corpus. The result scores 57 on the Artificial Analysis Intelligence Index v4.1.1 at $0.045 per task, approaching Claude Opus 4.8 on coding and agentic benchmarks. The entire release is served on a large-scale cluster of Chinese AI chips.
Native visual reasoning treats visual generation as the medium of reasoning itself: visual states (i.e. images and videos) are not merely inputs to be understood or outputs to be rendered, but first-class substrates for problem solving beyond language. Yet progress remains bottlenecked by the lack of scalable training tasks, reliable feedback, and controlled comparisons across generative substrates. In this work, we introduce VBVR-Pro, a closed-loop testbed that makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable. 1) Task scaling. VBVR-Pro turns visual reasoning into a controlled task space of 300 procedurally generated tasks. Models trained on VBVR-Pro show strong transfer beyond the proposed suite across seven external visual reasoning benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision. 2) Verifiable rewards. VBVR-Pro provides verifiable reward scorers for task-grounded evaluation. Through a systematic study of leading MLLMs as judges, we identify recurring failure modes of the prevalent VLM-as-a-judge paradigm. In contrast, the proposed scorers are grounded in deterministic, task-specific rules, achieve fine-grained alignment with human judgments. Importantly, they serve as reliable reward signals for large-scale multi-task reinforcement learning and demonstrate stronger post-RL performance across visual reasoning tasks. 3) Mechanism study. VBVR-Pro enables controlled modality studies across more than 30 image, video, and interleaved generators. Our analysis shows that video generation remains strongest for tasks requiring persistent spatiotemporal state tracking, while interleaved generation provides a compute-efficient alternative. Critically, ablations and probing suggest the presence of vision-native trajectories that are crucial to visual reasoning. We release all data, models, scorers, and code.
AdaptiveEmbed, a framework for multimodal retrieval, introduces Sample-Adaptive Multi-Vector Representation (SAMVR) to dynamically allocate representation capacity on a per-sample basis. The system enhances retrieval performance across diverse modalities while reducing the average number of active tokens per query, leading to improved efficiency.
UCAG-P introduces a unified camera-centric geometric action space and a decoupled policy architecture to align heterogeneous robot and human manipulation data. This framework enables a single policy to learn transferable manipulation geometry, achieving competitive success rates on various tasks for single-arm, bimanual, and humanoid robots, while also facilitating direct human-to-robot skill transfer.
Researchers to follow
View allZiwei Liu
Associate Professor (Provost’s Chair in AI) @ Nanyang Technological University, Previously Research Fellow @ The Chinese University of Hong Kong
Linxi "Jim" Fan
Director & Distinguished Research Scientist @ NVIDIA, Previously CS PhD Student @ Stanford University
Recuris, an architecture for LLM agents from NUS, Princeton, Stanford, and Oxford, enables recursive self-improvement by evolving an external memory-control layer while keeping the underlying LLM fixed. This approach achieved task success improvements of up to +17.8 points on GPT-5.6 Sol for long-horizon tasks, demonstrating that the learned memory components are transferable across models and tasks.
JIT-Agent, developed at LV-NUS Lab, introduces a system that dynamically generates, repairs, and evolves agent harnesses, transforming harness engineering from a static process to a trainable capability. This system improves LLM agent performance by an average of 7.7-8.8 points across various benchmarks and reduces token consumption by an average of 36%, often enabling open-backbone models to compete with or surpass frontier models.
Researchers from LeapLab, Tsinghua University, introduced On-policy Distillation with Verifiable Reward (OPDVR), a method that modifies sampled-token On-policy Distillation (OPD) by applying a simple ReLU gating mechanism to align its implicit reward signals with trajectory correctness. This approach enabled student models to achieve up to 49.1% average accuracy on six reasoning benchmarks, surpassing sampled-token OPD (47.8%) and sometimes even the teacher model, by effectively combining dense guidance with task-level correctness.
This work establishes that an 'Effective Learning Rate' (ELR), defined as the ratio of learning rate to parameter norm, acts as a primary governor of loss dynamics during language model pretraining. It demonstrates that when ELR schedules are matched, loss trajectories closely coincide, quantitatively explaining the effects of various norm-control methods and enabling more accurate functional scaling law transfer.
Carnegie Mellon University researchers developed ℛ3, a two-stage post-training method to enable Vision-Language Models (VLMs) to generate free-form natural language reasoning for guiding robotic manipulation policies. This approach significantly enhanced robotic task generalization, particularly for out-of-distribution scenarios, and demonstrated that explicit inference-time reasoning acts as a dynamic computational mechanism.
LAION-BVD introduces a 10-million-hour open video dataset, providing an unprecedented scale of video, audio, and image data for multimodal pre-training. It enables competitive performance across video-text, audio-text, and image-text models, also offering a distinct visual distribution compared to standard web-image corpora.
Simile's model simulates how a population will act in a given situation, returning a distribution over actions, but it cannot be accurate on every individual question. This post trains a second model that predicts the error of each simulation — the total variation distance between the predicted and observed distributions — without access to ground truth. Five approaches are compared on ~8600 held-out questions, and the resulting score is mapped to product-facing confidence buckets calibrated against a human-rated notion of a decision-grade simulation.
Meta_n introduces a recursive self-improvement framework for LLM agents, achieving emergent meta-depth by applying a fixed meta-operation to its own prior outputs, including traces and generated code. This approach enables deeper, more abstract strategic reasoning and yields substantial performance gains across diverse complex benchmarks compared to existing self-improving agents.
Video deep research answers complex questions by jointly understanding video content and retrieving external knowledge from the open Web. However, diverse questions and videos require different tool-use strategies, and inappropriate tool calls can produce incorrect results. Uncertain grounding and retrieval also make unnecessary interactions costly and error-prone, increasing latency and reasoning errors. To address these challenges, we propose AdaVDR, an adaptive video deep research agent with adaptive tool invocation and reflection. AdaVDR selects tools according to the task and its capabilities, and backtracks only when unreliable intermediate results require correction. To enable these capabilities, we develop a video deep research data construction pipeline. We first discover retrieval-relevant events and entities in diverse videos and acquire detailed information through grounding and external retrieval to construct high-quality QA pairs. For each QA, task-specific prompts organize the information acquisition process into a tool-use trajectory, allowing different question and video types to follow different grounding and retrieval strategies. We further introduce model-conditioned tool necessity filtering, which evaluates tool calls against the target model's video understanding and internal knowledge, removing tools or tool chains the model can bypass. This yields trajectories tailored to the target model's video understanding capability and knowledge. Using this pipeline, we construct training data and VDR-EE, a benchmark covering entity-centric and event-centric questions. We perform supervised fine-tuning followed by reinforcement learning with a redundancy-aware reward to strengthen adaptive tool invocation and reflection. Experiments show that our method performs best among the evaluated open-source models on VDR-EE and substantially improves over its base models on VideoDR.