A multi-agent framework using large language models for stock trading simulates real-world trading firms, improving performance metrics like cumulative returns and Sharpe ratio.

• 
• 4 authors

A multi-agent framework using large language models for stock trading simulates real-world trading firms, improving performance metrics like cumulative returns and Sharpe ratio.

• 
• 4 authors

Edge0 enables efficient mixture-of-experts inference on consumer hardware by predicting expert routing ahead of time and using recovery adapters to maintain quality with minimal active memory.

Edge0 enables efficient mixture-of-experts inference on consumer hardware by predicting expert routing ahead of time and using recovery adapters to maintain quality with minimal active memory.

TokenEconomizer reduces recurring token costs of coding agents across sequential real-world tasks by pruning irrelevant tools, injecting micro-skills, and interrupting failure loops.

TokenEconomizer reduces recurring token costs of coding agents across sequential real-world tasks by pruning irrelevant tools, injecting micro-skills, and interrupting failure loops.

MACaQ is an MXFP4 W4A4 post-training quantization method that quantizes each linear layer independently using online activation transforms and Hessian-aware refinement, reducing cost while preserving accuracy across language and diffusion models.

MACaQ is an MXFP4 W4A4 post-training quantization method that quantizes each linear layer independently using online activation transforms and Hessian-aware refinement, reducing cost while preserving accuracy across language and diffusion models.

LingBot-Touch is a tactile foundation model that learns a unified, extensible representation across heterogeneous tactile sensors, yielding plug-and-play tactile tokens that consistently enhance manipulation policies such as diffusion policy and VLA models.

LingBot-Touch is a tactile foundation model that learns a unified, extensible representation across heterogeneous tactile sensors, yielding plug-and-play tactile tokens that consistently enhance manipulation policies such as diffusion policy and VLA models.

Kronos, a specialized pre-training framework for financial K-line data, outperforms existing models in forecasting and synthetic data generation through a unique tokenizer and autoregressive pre-training on a large dataset.

• 7 authors

Kronos, a specialized pre-training framework for financial K-line data, outperforms existing models in forecasting and synthetic data generation through a unique tokenizer and autoregressive pre-training on a large dataset.

• 7 authors

VibeVoice synthesizes long-form multi-speaker speech using next-token diffusion and a highly efficient continuous speech tokenizer, achieving superior performance and fidelity.

VibeVoice synthesizes long-form multi-speaker speech using next-token diffusion and a highly efficient continuous speech tokenizer, achieving superior performance and fidelity.

Harness² improves agent execution harnesses through recursive task-level refinement and continual harness-level learning, boosting performance across benchmarks without ground-truth verifiers.

Harness² improves agent execution harnesses through recursive task-level refinement and continual harness-level learning, boosting performance across benchmarks without ground-truth verifiers.

Lloyd's k-means algorithm is a long-running standard for clustering, with applications in data analysis, retrieval, quantization and data curation. In this paper, we introduce an exponential moving average optimization procedure for k-means with mini-batch updates, powered by modern deep learning machinery and new insights. We also establish scaling laws on how decay rate and batch size impact final performance. This gives us a formulaic rule to fix them for multiple problem sizes and embedding datasets, without the need for any hyperparameter sweep. The resulting clustering algorithm, coined e-means, is much faster than Lloyd k-means and converges to lower losses, frequently even beating k-means++. On a wide array of embeddings datasets, it beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs, which translates to 10-45x GPU wall-clock speedups compared to state-of-the-art k-means implementations. As an example, e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd's at 25 epochs.

Lloyd's k-means algorithm is a long-running standard for clustering, with applications in data analysis, retrieval, quantization and data curation. In this paper, we introduce an exponential moving average optimization procedure for k-means with mini-batch updates, powered by modern deep learning machinery and new insights. We also establish scaling laws on how decay rate and batch size impact final performance. This gives us a formulaic rule to fix them for multiple problem sizes and embedding datasets, without the need for any hyperparameter sweep. The resulting clustering algorithm, coined e-means, is much faster than Lloyd k-means and converges to lower losses, frequently even beating k-means++. On a wide array of embeddings datasets, it beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs, which translates to 10-45x GPU wall-clock speedups compared to state-of-the-art k-means implementations. As an example, e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd's at 25 epochs.

OpenDevin is a platform for developing AI agents that interact with the world by writing code, using command lines, and browsing the web, with support for multiple agents and evaluation benchmarks.

• 
• 
• 
• 
• 
• 24 authors

OpenDevin is a platform for developing AI agents that interact with the world by writing code, using command lines, and browsing the web, with support for multiple agents and evaluation benchmarks.

• 
• 
• 
• 
• 
• 24 authors

PagedAttention algorithm and vLLM system enhance the throughput of large language models by efficiently managing memory and reducing waste in the key-value cache.

• 
• 
• 
• 
• 
• 9 authors

PagedAttention algorithm and vLLM system enhance the throughput of large language models by efficiently managing memory and reducing waste in the key-value cache.

• 
• 
• 
• 
• 
• 9 authors

Flide is an exact inference engine that accelerates decision-model LLM agents by decoupling logical request trees from physical execution and using ForkAttention to preserve exact branch evaluations.

Flide is an exact inference engine that accelerates decision-model LLM agents by decoupling logical request trees from physical execution and using ForkAttention to preserve exact branch evaluations.

FreeToken is an edge-native Mixture-of-Experts serving system that dynamically maps computation and model state onto heterogeneous local hardware to run large open-weight models on personal machines.

FreeToken is an edge-native Mixture-of-Experts serving system that dynamically maps computation and model state onto heterogeneous local hardware to run large open-weight models on personal machines.

SplatAnything applies Gaussian Spacing Normalization to canonicalize scene-level Gaussian spacing, yielding more consistent supervision and improved feed-forward 3D Gaussian Splatting under varying scene scales and viewpoints.

SplatAnything applies Gaussian Spacing Normalization to canonicalize scene-level Gaussian spacing, yielding more consistent supervision and improved feed-forward 3D Gaussian Splatting under varying scene scales and viewpoints.

Unlimited OCR introduces Reference Sliding Window Attention to eliminate growing memory consumption during long-sequence OCR tasks, enabling efficient transcription of multiple pages in a single forward pass.

Unlimited OCR introduces Reference Sliding Window Attention to eliminate growing memory consumption during long-sequence OCR tasks, enabling efficient transcription of multiple pages in a single forward pass.

Enhancements to the AgentScope platform improve scalability, efficiency, and ease of use for large-scale multi-agent simulations through distributed mechanisms, flexible environments, and user-friendly tools.

• 
• 
• 
• 
• 8 authors

Enhancements to the AgentScope platform improve scalability, efficiency, and ease of use for large-scale multi-agent simulations through distributed mechanisms, flexible environments, and user-friendly tools.

• 
• 
• 
• 
• 8 authors

AgentScope enhances agentic applications by providing flexible tool-based interactions, unified interfaces, and advanced infrastructure based on the ReAct paradigm, supporting efficient and safe development and deployment.

• 
• 
• 
• 23 authors

AgentScope enhances agentic applications by providing flexible tool-based interactions, unified interfaces, and advanced infrastructure based on the ReAct paradigm, supporting efficient and safe development and deployment.

• 
• 
• 
• 23 authors

OmniEdu is a family of instruction-tuned educational foundation models that improve K-12 problem solving, curriculum understanding, and tutoring through capability-balanced training data.

OmniEdu is a family of instruction-tuned educational foundation models that improve K-12 problem solving, curriculum understanding, and tutoring through capability-balanced training data.

Transformers library provides state-of-the-art Transformer architectures and pretrained models for natural language processing tasks with a unified API and emphasis on extensibility and robust deployment.

Transformers library provides state-of-the-art Transformer architectures and pretrained models for natural language processing tasks with a unified API and emphasis on extensibility and robust deployment.

Mem0, a memory-centric architecture with graph-based memory, enhances long-term conversational coherence in LLMs by efficiently extracting, consolidating, and retrieving information, outperforming existing memory systems in terms of accuracy and computational efficiency.

• 
• 
• 5 authors

Mem0, a memory-centric architecture with graph-based memory, enhances long-term conversational coherence in LLMs by efficiently extracting, consolidating, and retrieving information, outperforming existing memory systems in terms of accuracy and computational efficiency.

• 
• 
• 5 authors

MinerU2.5, a 1.2B-parameter document parsing vision-language model, achieves state-of-the-art recognition accuracy with computational efficiency through a coarse-to-fine parsing strategy.

• 
• 
• 
• 
• 
• 61 authors

MinerU2.5, a 1.2B-parameter document parsing vision-language model, achieves state-of-the-art recognition accuracy with computational efficiency through a coarse-to-fine parsing strategy.

• 
• 
• 
• 
• 
• 61 authors

ARIS is an open-source research harness that uses cross-model adversarial collaboration to ensure reliable long-term research outcomes through coordinated execution, orchestration, and assurance layers.

ARIS is an open-source research harness that uses cross-model adversarial collaboration to ensure reliable long-term research outcomes through coordinated execution, orchestration, and assurance layers.

Next-Concept Modeling uses hierarchical discrete codes and parallel level-wise generation to accelerate diffusion language models while improving quality and reasoning performance.

Next-Concept Modeling uses hierarchical discrete codes and parallel level-wise generation to accelerate diffusion language models while improving quality and reasoning performance.

YuE2 unifies symbolic and audio music generation via symbolic planning, producing readable scores and full-song audio with state-of-the-art quality and agentic editing capabilities.

YuE2 unifies symbolic and audio music generation via symbolic planning, producing readable scores and full-song audio with state-of-the-art quality and agentic editing capabilities.

SmolDocling is a compact vision-language model that performs end-to-end document conversion with robust performance across various document types using 256M parameters and a new markup format.

SmolDocling is a compact vision-language model that performs end-to-end document conversion with robust performance across various document types using 256M parameters and a new markup format.

PaddleOCR-VL-1.6 enhances document parsing performance through targeted data optimization and progressive post-training techniques, achieving state-of-the-art results on OmniDocBench v1.6.

PaddleOCR-VL-1.6 enhances document parsing performance through targeted data optimization and progressive post-training techniques, achieving state-of-the-art results on OmniDocBench v1.6.

The PyTorch distributed data parallel module optimizes large-scale model training using techniques like gradient bucketing, computation-communication overlap, and selective synchronization to achieve near-linear scalability.

• 
• 
• 
• 11 authors

The PyTorch distributed data parallel module optimizes large-scale model training using techniques like gradient bucketing, computation-communication overlap, and selective synchronization to achieve near-linear scalability.

• 
• 
• 
• 11 authors

World action models jointly predict video and actions conditioned on language. Existing studies focus primarily on architectures and training mechanisms, typically treating language-video alignment acquired through video pretraining as an established foundation. However, introducing action learning may impair this alignment. Principled guidance on data and supervision design for maintaining language-video alignment during action learning remains limited. To investigate this, we curate a pool of approximately 47,000 hours of general videos, egocentric videos, UMI demonstrations, and robot demonstrations. Building on this, we systematically study how language-video supervision, language-video-action supervision, and the joint use of these two forms affect WAM performance. In particular, we examine how language annotation quality affects pretrained language-video capabilities, and how action supervision from UMI and robot demonstrations extends these capabilities to language-video-action modeling. To better scale the model, we further conduct detailed studies on joint training recipe and model architecture. Under a fixed training budget, mixing egocentric videos and UMI demonstrations improves control generalization. Model architecture affects how effectively these experiences are used, while scaling pretraining under the selected configuration brings further gains. The resulting recipe, post-trained on RoboTwin Clean demonstrations only, achieves success rates of 88.10% and 61.56% under Easy (Clean-to-Clean) and Hard (Clean-to-Random) conditions, respectively. Our model achieves 40.00% SR across three real-world tasks, compared with 23.33% for the strongest baseline, indicating better generalization under unseen conditions. We release the curated data, training code, and model weights.

World action models jointly predict video and actions conditioned on language. Existing studies focus primarily on architectures and training mechanisms, typically treating language-video alignment acquired through video pretraining as an established foundation. However, introducing action learning may impair this alignment. Principled guidance on data and supervision design for maintaining language-video alignment during action learning remains limited. To investigate this, we curate a pool of approximately 47,000 hours of general videos, egocentric videos, UMI demonstrations, and robot demonstrations. Building on this, we systematically study how language-video supervision, language-video-action supervision, and the joint use of these two forms affect WAM performance. In particular, we examine how language annotation quality affects pretrained language-video capabilities, and how action supervision from UMI and robot demonstrations extends these capabilities to language-video-action modeling. To better scale the model, we further conduct detailed studies on joint training recipe and model architecture. Under a fixed training budget, mixing egocentric videos and UMI demonstrations improves control generalization. Model architecture affects how effectively these experiences are used, while scaling pretraining under the selected configuration brings further gains. The resulting recipe, post-trained on RoboTwin Clean demonstrations only, achieves success rates of 88.10% and 61.56% under Easy (Clean-to-Clean) and Hard (Clean-to-Random) conditions, respectively. Our model achieves 40.00% SR across three real-world tasks, compared with 23.33% for the strongest baseline, indicating better generalization under unseen conditions. We release the curated data, training code, and model weights.

HairDeck generates stylized hair cards and spring-bone animations as unordered token sets using a diffusion transformer, improving accuracy and speed over prior methods.

HairDeck generates stylized hair cards and spring-bone animations as unordered token sets using a diffusion transformer, improving accuracy and speed over prior methods.

Genesis organizes long-horizon software development around a persistent project rather than persistent agents, enabling multi-day compiler construction and numerical module reimplementation with low cost and high performance.

Genesis organizes long-horizon software development around a persistent project rather than persistent agents, enabling multi-day compiler construction and numerical module reimplementation with low cost and high performance.

Generating high-fidelity voxels while keeping their lightweight nature remains an open problem in computer vision. Current generators are either high-fidelity but computationally heavy, or lightweight but of lower fidelity. Hierarchical methods keep the lightweight nature of voxels. However, their coarse levels, which decide where all finer cells can appear, are generated with the least information about the shape and are never revisited, so they are often plausible but wrong. Our key insight is to let finer cells take part in deciding coarse cells. To this end, we propose JoVox, a discrete flow matching framework that learns the joint distribution of all levels of the voxel hierarchy, instead of a separate distribution for each level, so that coarse cells can be predicted with finer cells already generated nearby. Within this framework, we further introduce a two-level prediction, which outputs the cells two levels finer at once instead of one level finer, halving the number of levels that the network processes and yielding voxels of higher fidelity. We also introduce an unmasking strategy that revisits early decisions once finer cells are generated and corrects inconsistencies in masked diffusion models. Our framework outperforms the baselines on both unconditional and image-conditioned voxel generation. Compared with the current state-of-the-art TRELLIS.2, we achieve 8x higher training throughput, 1/15 of the FLOPs in a forward pass, train on 4 vs. 32 GPUs, and obtain higher quality with an 11% lower Chamfer distance.

Generating high-fidelity voxels while keeping their lightweight nature remains an open problem in computer vision. Current generators are either high-fidelity but computationally heavy, or lightweight but of lower fidelity. Hierarchical methods keep the lightweight nature of voxels. However, their coarse levels, which decide where all finer cells can appear, are generated with the least information about the shape and are never revisited, so they are often plausible but wrong. Our key insight is to let finer cells take part in deciding coarse cells. To this end, we propose JoVox, a discrete flow matching framework that learns the joint distribution of all levels of the voxel hierarchy, instead of a separate distribution for each level, so that coarse cells can be predicted with finer cells already generated nearby. Within this framework, we further introduce a two-level prediction, which outputs the cells two levels finer at once instead of one level finer, halving the number of levels that the network processes and yielding voxels of higher fidelity. We also introduce an unmasking strategy that revisits early decisions once finer cells are generated and corrects inconsistencies in masked diffusion models. Our framework outperforms the baselines on both unconditional and image-conditioned voxel generation. Compared with the current state-of-the-art TRELLIS.2, we achieve 8x higher training throughput, 1/15 of the FLOPs in a forward pass, train on 4 vs. 32 GPUs, and obtain higher quality with an 11% lower Chamfer distance.

Uni-robot-agent enables open-weight robot-use agents to surpass proprietary models in zero-shot manipulation and suggests joint development of four core capabilities for future training.

• 
• 
• 
• 
• 
• 28 authors

Uni-robot-agent enables open-weight robot-use agents to surpass proprietary models in zero-shot manipulation and suggests joint development of four core capabilities for future training.

• 
• 
• 
• 
• 
• 28 authors

Trained action policies write manipulation skill into network weights, tying what is learned to the distribution it is learned from. In-context learning (ICL) avoids this coupling but has not matched trained policies, because a demonstration recorded in another scene carries no feedback from the world the robot currently faces. We observe that what an agent learns in context is semantic, recording task steps and object roles rather than visual appearance, so experience gathered in one world transfers to another. Based on this insight, we propose World In-Context Learning (WICL), a paradigm in which a frozen model starts from a demonstration, rehearses the task in a world such as a simulator, and writes the feedback into its context rather than its weights. We build RoboScript, a modular infrastructure that implements WICL for agentic manipulation. From a single demonstration, RoboScript learns each task through rehearsal in simulation, freezes the learned knowledge as a Task Memory, and deploys the frozen memory on new scenes with data isolation that ensures every deployment input is available on a real robot. On all 50 tasks of RoboTwin 2.0, RoboScript reaches 83.4% mean success across clean and randomized settings with one demonstration per task and no gradient updates, exceeding the strongest trained action policy by over 20 points. The method generalizes across visual distributions, improves through reflection, scales with the foundation model, and transfers to a physical dual-arm robot through cross-embodiment memory adaptation. With a world in the loop, a robot can learn a manipulation task in its context as well as it learns one in its weights. All code is available at https://github.com/ArcheOne-AI/RoboScript

Trained action policies write manipulation skill into network weights, tying what is learned to the distribution it is learned from. In-context learning (ICL) avoids this coupling but has not matched trained policies, because a demonstration recorded in another scene carries no feedback from the world the robot currently faces. We observe that what an agent learns in context is semantic, recording task steps and object roles rather than visual appearance, so experience gathered in one world transfers to another. Based on this insight, we propose World In-Context Learning (WICL), a paradigm in which a frozen model starts from a demonstration, rehearses the task in a world such as a simulator, and writes the feedback into its context rather than its weights. We build RoboScript, a modular infrastructure that implements WICL for agentic manipulation. From a single demonstration, RoboScript learns each task through rehearsal in simulation, freezes the learned knowledge as a Task Memory, and deploys the frozen memory on new scenes with data isolation that ensures every deployment input is available on a real robot. On all 50 tasks of RoboTwin 2.0, RoboScript reaches 83.4% mean success across clean and randomized settings with one demonstration per task and no gradient updates, exceeding the strongest trained action policy by over 20 points. The method generalizes across visual distributions, improves through reflection, scales with the foundation model, and transfers to a physical dual-arm robot through cross-embodiment memory adaptation. With a world in the loop, a robot can learn a manipulation task in its context as well as it learns one in its weights. All code is available at https://github.com/ArcheOne-AI/RoboScript

WaveGen is a fully end-to-end diffusion framework for text-to-waveform generation that jointly learns spectral and waveform trajectories via Spectral Forcing without external audio codecs, vocoders, latent models, alignment modules, or self-supervised representations, achieving state-of-the-art text-to-speech results.

WaveGen is a fully end-to-end diffusion framework for text-to-waveform generation that jointly learns spectral and waveform trajectories via Spectral Forcing without external audio codecs, vocoders, latent models, alignment modules, or self-supervised representations, achieving state-of-the-art text-to-speech results.

A novel GPT-based model, OmniFlatten, enables real-time natural full-duplex spoken dialogue through a multi-stage post-training technique that integrates speech and text without altering the original model's architecture.

• 
• 9 authors

A novel GPT-based model, OmniFlatten, enables real-time natural full-duplex spoken dialogue through a multi-stage post-training technique that integrates speech and text without altering the original model's architecture.

• 
• 9 authors

Assistants from 2011 answered and waited, and agents from 2023 did a task and stopped. In September 2026 Meta's Muse showed an agent for one person, with accounts, devices, memory and a conversation that lasts, closed, in a vendor's cloud, in one country. Such an agent is expected to act on a person's accounts and devices, remember them across weeks, speak first when it is worth it, and answer for what it did. It is a kind of software, not a model, and until now had no open counterpart. This report defines the personal agent in five questions and three horizons. It reads how Muse is built from Meta's public record and a copy of its production prompt, each statement marked by its source. It then presents nanoMuse, the open-source counterpart under the GPL-3.0, one agent on every device a person owns, with hands on the phone's screen and the computer's. They share one conversation over a relay anyone can run; every action goes through a Sentinel, memory is files the person can read, and the model is their choice. Its size and cost are given as estimates. What is open, memory with provenance, an evaluation suite for the hands and an open model for them, is set out as a roadmap.

Assistants from 2011 answered and waited, and agents from 2023 did a task and stopped. In September 2026 Meta's Muse showed an agent for one person, with accounts, devices, memory and a conversation that lasts, closed, in a vendor's cloud, in one country. Such an agent is expected to act on a person's accounts and devices, remember them across weeks, speak first when it is worth it, and answer for what it did. It is a kind of software, not a model, and until now had no open counterpart. This report defines the personal agent in five questions and three horizons. It reads how Muse is built from Meta's public record and a copy of its production prompt, each statement marked by its source. It then presents nanoMuse, the open-source counterpart under the GPL-3.0, one agent on every device a person owns, with hands on the phone's screen and the computer's. They share one conversation over a relay anyone can run; every action goes through a Sentinel, memory is files the person can read, and the model is their choice. Its size and cost are given as estimates. What is open, memory with provenance, an evaluation suite for the hands and an open model for them, is set out as a roadmap.

A large language model adapted for time-series forecasting achieves near-optimal zero-shot performance on diverse datasets across different time scales and granularities.

• 4 authors

A large language model adapted for time-series forecasting achieves near-optimal zero-shot performance on diverse datasets across different time scales and granularities.

• 4 authors

WeVisOne-VL unifies object-level and pixel-level visual perception via structured token prediction, supporting detection, localization, counting, OCR, segmentation, and matting with a lightweight pixel decoder and multi-stage training.

WeVisOne-VL unifies object-level and pixel-level visual perception via structured token prediction, supporting detection, localization, counting, OCR, segmentation, and matting with a lightweight pixel decoder and multi-stage training.

MOSS-Image is an open diffusion transformer that combines high-fidelity synthesis with intent-aware prompt enhancement to generate text-rich, layout-coherent images, achieving state-of-the-art open-weight results on long-form text and layout benchmarks.

MOSS-Image is an open diffusion transformer that combines high-fidelity synthesis with intent-aware prompt enhancement to generate text-rich, layout-coherent images, achieving state-of-the-art open-weight results on long-form text and layout benchmarks.

Zep, a memory layer service, outperforms MemGPT in the DMR benchmark and LongMemEval by excelling in dynamic knowledge integration and temporal reasoning, critical for enterprise use cases.

• 
• 
• 5 authors

Zep, a memory layer service, outperforms MemGPT in the DMR benchmark and LongMemEval by excelling in dynamic knowledge integration and temporal reasoning, critical for enterprise use cases.

• 
• 
• 5 authors

GPT-6-Astra achieves high performance across diverse robotic tasks via in-context learning, yet real-world deployment reveals gaps in demonstration cost, execution time, and practical embodied intelligence.

GPT-6-Astra achieves high performance across diverse robotic tasks via in-context learning, yet real-world deployment reveals gaps in demonstration cost, execution time, and practical embodied intelligence.

A new sparse voxel representation called O-Voxel enables high-quality 3D generative modeling with efficient inference and robust topology handling.

A new sparse voxel representation called O-Voxel enables high-quality 3D generative modeling with efficient inference and robust topology handling.

A benchmark of 204 visually intensive search tasks shows that even the best multimodal agents with search tools solve only 26.47% of tasks, underscoring the difficulty of linking visual evidence across search steps.

A benchmark of 204 visually intensive search tasks shows that even the best multimodal agents with search tools solve only 26.47% of tasks, underscoring the difficulty of linking visual evidence across search steps.

YOLO26 addresses real-time vision challenges through a unified model family with NMS-free inference, improved training strategies, and multi-task capabilities spanning detection, segmentation, and pose estimation.

YOLO26 addresses real-time vision challenges through a unified model family with NMS-free inference, improved training strategies, and multi-task capabilities spanning detection, segmentation, and pose estimation.

RoboHarness enables LLM agents to perform embodied manipulation via a visual-geometric control interface, surpassing specialized vision-language-action models on robotics benchmarks.

RoboHarness enables LLM agents to perform embodied manipulation via a visual-geometric control interface, surpassing specialized vision-language-action models on robotics benchmarks.

While Recursive Self-Improvement (RSI) accelerates isolated algorithm and kernel synthesis, distributed LLM training depends on tightly coupled launch settings, operator kernels, and communication schedules, where single edits often hit out-of-memory (OOM) boundaries, compiler conflicts, measurement jitter, or silent numerical drift. We present InfraRSI, the agentic harness that turns open-ended human commands into verified training optimizations via a five-stage workflow: init locks editable scope and a fixed task contract over workload, loss, and timing invariants to prevent reward hacking, doctor profiles baseline phases and calibrates cluster noise, run co-optimizes launch knobs and code patches before re-verifying top candidates against the baseline and logging findings to a cross-hardware experience bank, and report and apply audit source paths, track failed-trial costs, and commit promoted changes. Rather than stepping greedily from a single incumbent, InfraRSI initializes from diverse experience-seeded configurations and fits global execution-time and censored peak-memory models across all trials, including OOMs. Using these models, it prices per-setting memory costs, revises miscalibrated LLM priors from live runs, and pairs memory-hungry optimizations with memory savers and verified code patches. Across four single- and multi-node dense and Mixture-of-Experts (MoE) LLM workloads, InfraRSI achieves a speedup of up to 2.75times and lifts mean Model FLOPs Utilization (MFU) from 17.4% to 30.0%, outperforming existing tuners and coding agents while transferring experience across GPUs. As it improves training infrastructure efficiency, InfraRSI forms a two-level Recursive Self-Improvement loop---revising priors online within runs and distilling experience across clusters---to push foundation model scaling.

While Recursive Self-Improvement (RSI) accelerates isolated algorithm and kernel synthesis, distributed LLM training depends on tightly coupled launch settings, operator kernels, and communication schedules, where single edits often hit out-of-memory (OOM) boundaries, compiler conflicts, measurement jitter, or silent numerical drift. We present InfraRSI, the agentic harness that turns open-ended human commands into verified training optimizations via a five-stage workflow: init locks editable scope and a fixed task contract over workload, loss, and timing invariants to prevent reward hacking, doctor profiles baseline phases and calibrates cluster noise, run co-optimizes launch knobs and code patches before re-verifying top candidates against the baseline and logging findings to a cross-hardware experience bank, and report and apply audit source paths, track failed-trial costs, and commit promoted changes. Rather than stepping greedily from a single incumbent, InfraRSI initializes from diverse experience-seeded configurations and fits global execution-time and censored peak-memory models across all trials, including OOMs. Using these models, it prices per-setting memory costs, revises miscalibrated LLM priors from live runs, and pairs memory-hungry optimizations with memory savers and verified code patches. Across four single- and multi-node dense and Mixture-of-Experts (MoE) LLM workloads, InfraRSI achieves a speedup of up to 2.75times and lifts mean Model FLOPs Utilization (MFU) from 17.4% to 30.0%, outperforming existing tuners and coding agents while transferring experience across GPUs. As it improves training infrastructure efficiency, InfraRSI forms a two-level Recursive Self-Improvement loop---revising priors online within runs and distilling experience across clusters---to push foundation model scaling.

PDFMathTranslate enables layout-preserving scientific document translation using large language models and precise layout detection, offering improved precision, flexibility, and efficiency.

• 4 authors

PDFMathTranslate enables layout-preserving scientific document translation using large language models and precise layout detection, offering improved precision, flexibility, and efficiency.

• 4 authors

Xiaomi-OCR-0 is a compact vision-language model that unifies document parsing and OCR understanding through automated data synthesis, progressive training with text anchoring and reinforcement learning, and achieves state-of-the-art results on multiple benchmarks.

• 
• 
• 10 authors

Xiaomi-OCR-0 is a compact vision-language model that unifies document parsing and OCR understanding through automated data synthesis, progressive training with text anchoring and reinforcement learning, and achieves state-of-the-art results on multiple benchmarks.

• 
• 
• 10 authors

UniMate is a unified diffusion transformer that generates articulated motion for arbitrary skeletons from text and rigged 3D assets without per-skeleton retraining, using topology-aware attention and a large curated motion dataset.

UniMate is a unified diffusion transformer that generates articulated motion for arbitrary skeletons from text and rigged 3D assets without per-skeleton retraining, using topology-aware attention and a large curated motion dataset.

SUPRA-Tex decouples visible-material estimation, surface alignment, and 3D completion to generate relightable 3D materials that preserve reference image details and complete unobserved surfaces.

• 
• 
• 
• 
• 8 authors

SUPRA-Tex decouples visible-material estimation, surface alignment, and 3D completion to generate relightable 3D materials that preserve reference image details and complete unobserved surfaces.

• 
• 
• 
• 
• 8 authors