📄 Trending AI Research

A 150M-parameter reasoning model using recurrent latent reasoning and in-context learning achieves a new cost-accuracy frontier on ARC-AGI-1.

StateM is a runtime system that improves long-horizon agent execution through durable states, recoverable runbooks, and enforceable procedural controls without altering model weights.

Kimi K3 is a 2.8-trillion-parameter open-source AI model that uses a Mixture-of-Experts design to activate only 104 billion parameters per token, making it efficient despite its size. It introduces new attention mechanisms and a stable expert-routing method to improve performance across long sequences and deep networks, achieving roughly 2.5x better scaling efficiency than its predecessor. The model excels at long-horizon coding, agentic tasks, reasoning, and vision, outperforming all other open models and most proprietary ones, though it still trails the top closed models like Claude Fable 5 and GPT-5.6 Sol. Its release as open weights aims to accelerate research and deployment of frontier-level AI.

Macaron-V1 is an open agent-model family that uses a Mixture-of-LoRA architecture and recursive self-improvement to enable continual learning and collaboration across specialized tasks.

This paper introduces Qwen-UI-Agent, a foundation model designed to act as a general-purpose executor across mobile, desktop, web, and search interfaces. Unlike typical GUI agents that operate in isolated sandboxes, it runs on real devices and unifies GUI clicks with command-line actions, generating multiple steps per turn to handle long, complex workflows. A key innovation is an automated data flywheel where the agent itself builds tasks, diagnoses its own failures, and plans improvements, plus online reinforcement learning on trajectories over 100 steps using 10,000+ parallel environments. The headline results show state-of-the-art mobile performance (82.1% on MobileWorld, 92.2% on MobileWorld-Real) and competitive computer/browser scores (79.5% on OSWorld-Verified, 73.6% on WebArena) against frontier models like GPT-5.6 and Gemini 3.1 Pro. The work matters because it pushes agents from scripted demos toward reliable, self-improving operation on actual devices, a prerequisite for practical digital assistants.

AskChem changes chemistry literature search from returning ranked papers to returning atomic, provenance-carrying claims—each tied to a source DOI and a verbatim quote or evidence locator. It indexes 2.4M claims from 147K papers and offers multiple access paths: a faceted taxonomy for browsing, an evidence graph linking claims, and a living taxonomy organized by scientific principles, plus REST, SDK, and MCP interfaces for AI agents. In their benchmark, grounding a GPT-5.5 reader in AskChem achieved 100% resolvable DOIs versus 88.3% without retrieval, and the highest citation density among five systems tested. This matters because it directly addresses the manual assembly and verification burden in cross-paper synthesis, for both humans and AI agents.

A new benchmark for AI-generated video detection reveals that current detectors fail to generalize across realistic crisis-related videos and become less reliable as content spreads socially.

This paper introduces the concept of a 'memory foundation model,' arguing that current AI agents rely on external memory systems (like vector databases) rather than building memory directly into the model's weights and computation. The authors propose Metis, a prototype that maintains a persistent, evolving memory state inside the model's backbone, updated via a simple forward pass without gradient updates. This design allows historical information to be compressed into the model and accessed through a special 'memory attention' mechanism, keeping all learned weights frozen during inference. The key result is that Metis demonstrates native memory capabilities—storing and retrieving information autonomously—while offering architectural and efficiency advantages over external memory approaches, though the paper also analyzes its current limitations.

EnvHarness and EnvRigger dynamically reshape static environments via programmable plugins to target agent weaknesses and improve reinforcement learning co-evolution.

OpenART introduces a scalable red-teaming arena with evolving stateful environments to evaluate long-horizon AI agent safety, using the EMHA attack policy to expose increasing failure rates as task complexity grows.

Spark-to-Paper is a lightweight, composable workflow inside coding assistants that generates research papers by separating planning from reporting, enforcing evidence-based claim revision, and using integrity checks to reduce fabrication.

This paper introduces a method to automatically generate large volumes of training data for AI agents that perform long, multi-step tasks in terminal environments (like running commands in a shell). The key problem is that such data is normally hand-crafted and costs hundreds to thousands of dollars per task because the instruction, environment, and solution must all be consistent. The new approach, called Recursive Synthetic Terminal Tasks (RST), starts with a few verified tasks and repeatedly extends their solutions, then automatically updates the instructions and checks to match, validating each new task in a sandbox. Over 15 rounds, it produced 37,484 tasks at about $0.05 each, with difficulty rising sharply—median solution length grew from 67 to 374 lines, and the best model's success rate dropped from 90% to 2.5%. Training models on these synthetic tasks improved their performance on standard terminal-agent benchmarks by up to 41% relative, and the process shows no signs of plateauing, suggesting it could scale further.

ABot-World-0 is a video world model that runs real-time, interactive simulations on a single desktop GPU. It learns to generate coherent video sequences in response to keyboard actions by training on a diverse mix of AAA games, simulations, and internet videos. The key innovation is a training method called LongForcing that aligns the model's long self-generated sequences with a more accurate teacher model, preventing the drift that typically plagues long video generation. On a single RTX 5090, it streams 720P video at up to 16 FPS with only 1.2 seconds of latency, enabling interactive scene roaming and character control.

Orca establishes a unified world latent space through next-state-prediction modeling using multimodal data and demonstrates superior performance in downstream tasks compared to specialized baselines.

Modern AI agents rely on a 'harness'—the code that builds prompts, manages state, calls tools, and coordinates execution—which must be constantly updated as models and APIs change. The main bottleneck is finding exactly which lines of code implement a desired behavior, because harness code is large, tangled, and spread across many files. This paper introduces the Harness Handbook, an automatically generated document that maps each high-level behavior to its source code locations, and a method called Behavior-Guided Progressive Disclosure (BGPD) that helps both human developers and coding agents navigate from a behavior description down to the relevant code. In tests on two real open-source harnesses, the handbook approach improved the accuracy of locating code and planning edits while using fewer tokens, especially for behaviors scattered across modules or on rarely executed paths.

Unsupervised on-policy self-distillation improves large language models by using internal consistency and majority-vote pseudo-solutions to correct confident errors without external supervision.

RynnBrain 1.1 is a family of embodied AI models (2B, 9B, and a 122B-A10B mixture-of-experts variant) trained to handle perception, spatial reasoning, localization, and planning for robots. The key upgrades over version 1.0 are contact-point prediction (where a robot should touch an object) and native 3D grounding for the smaller models, which makes their outputs more directly usable for manipulation. The authors also built a separate VLA (vision-language-action) model with a shared action space across different robot bodies, using masking to handle their differing morphologies, and tested it on three real robots. On benchmarks like VSI-Bench and RefSpatial-Bench, the largest model beats all proprietary and open-source competitors, and real-robot tests show that policies initialized with RynnBrain outperform those based on Qwen or other generalist VLAs. Notably, training on multiple tasks and robot types simultaneously improved both process scores and success rates compared to training each task separately.

LongStraw is a system that enables reinforcement learning (RL) post-training on contexts exceeding 2 million tokens using a fixed GPU budget, addressing the gap where inference handles million-token contexts but RL training typically tops out at 256K tokens. It works by evaluating the shared prompt without autograd, keeping only essential model state for later tokens, and replaying short response branches one at a time—trading extra replay time for a smaller live training graph. Implemented on Qwen3.6-27B and GLM-5.2 models, LongStraw completed grouped scoring and backward passes at 2.1 million positions on eight H20 GPUs, with a stress test reaching 4.46 million positions. This matters for AI agents that accumulate long trajectories of observations and tool outputs, though the paper notes these experiments establish execution capacity rather than full training correctness.

Combodied Agents integrate digital and embodied tools into a closed-loop framework that models individual human-state trajectories over time to provide proportionate, consent-aware support.

Latent-to-4D enables reusable direct 4D generation from video diffusion latents via alignment with a pretrained decoder and spatiotemporal attention, transferring across generators without retraining.

This paper introduces OpenMLE, an open-source framework for studying recursive self-improvement (RSI) in machine learning engineering, where AI systems help build better AI. The framework includes a verifiable task environment, reinforcement learning for operator learning, and long-horizon evolutionary search. Using this stack, the authors post-train a 35B-parameter model called Frontis-MA1 to act as a meta-evolution agent, composing four atomic program-editing operations—Draft, Improve, Debug, and Crossover—into an iterative loop. On the MLE-Bench Lite benchmark, the model improves its Medal Average from 39.39% to 60.61% over its base model, and reaches 71.21% with an enhanced search variant, outperforming GPT-5.5 + Codex and approaching much larger models like GPT-5.6 Sol and Kimi K3. The results demonstrate that both the trained model and the search framework independently contribute to performance gains, and the authors release the model weights and code to enable reproducible RSI research.

PhiZero introduces a world model that predicts video by first translating visual changes into a compact, discrete 'physical language' — essentially a sequence of symbolic descriptions of state transitions — and then rendering that sequence back into pixels. This reason-then-render approach contrasts with prior models that directly generate future frames, which leaves dynamics implicit and hard to control. By learning this language self-supervised from raw videos, the model can explicitly reason about physical evolution, enabling more coherent generation and offering practical benefits like fine-grained action-conditioned simulation and zero-shot motion transfer across different scenes. The paper reports that this approach performs strongly on both video generation and understanding benchmarks, suggesting that explicit symbolic abstraction can improve physical consistency in learned world models.

Self-supervised visual on-policy distillation improves small vision-language models by distilling from original images into strongly augmented student views without privileged annotations or larger teachers.

Skills enhance LLM agents primarily by stabilizing execution through procedural anchoring rather than injecting missing knowledge, though retrieval bottlenecks and brittle assumptions limit their effectiveness.

Current video AI models can describe what happens in a video but struggle to pinpoint exactly when events occur. TimeLens2 solves this by treating temporal evidence as a set of intervals rather than individual moments, using a new training dataset (TimeLens2-93K) with reliable multi-span labels and a novel reward function (temporal Wasserstein reward) that provides dense feedback without requiring exact segment matching. This approach allows a single model to handle videos of varying lengths, domains, and query types. On seven benchmarks, TimeLens2-2B outperforms all similarly-sized models, while the 4B and 8B versions achieve state-of-the-art results, surpassing models up to 397B parameters and improving over their base Qwen3-VL backbones by 13-18 mean IoU points.

VideoChat3 is a fully open-source video AI model that balances efficiency and broad capability with only 4 billion parameters. It introduces two key innovations: an Inflated 3D Vision Transformer (I3D-ViT) and adaptive frame resolution to cut computational costs, plus a scalable data pipeline that generates diverse training datasets for general, long-form, and streaming video. Unlike many open models that are only partially released, VideoChat3 provides all code, training strategies, and datasets. The headline result: it outperforms prior open-source models with equal or larger parameter counts across multiple video benchmarks, demonstrating that smaller, efficient models can achieve strong generalization.

Semantic task completion video generation evaluates whether generated videos achieve intended outcomes with semantic grounding, supported by a curated dataset and vision-language model-based benchmark.

Training-inference mismatch in reinforcement learning for large language models leads to instability, which is addressed through a new policy optimization objective and framework that ensures consistent policy improvements between training and inference phases.

SA-MRPO independently standardizes multi-objective rewards and adaptively discounts saturated objectives to redirect optimization toward under-optimized goals.

Zetta is a closed-loop embodied harness that evolves runtime critics and recovery skills online to govern physical execution at action frequency, achieving high success on robot benchmarks with faster inference and scaling self-exploration.

Deep research often involves finding answers that satisfy multiple constraints, where verifying a candidate answer is much easier than discovering it from scratch. AREX exploits this asymmetry by using a recursive self-improvement loop: an inner loop gathers evidence and builds a provisional answer, while an outer loop audits that answer constraint-by-constraint, identifies unresolved claims, and launches targeted follow-up research. To sustain this process over long horizons without an external model, AREX learns a context-update tool that compresses the interaction history into a compact state of verified evidence and unresolved constraints. Trained via agentic mid-training and long-horizon reinforcement learning, AREX models (4B and 122B-A10B MoE) substantially outperform comparable baselines on benchmarks like BrowseComp, WideSearch, and Humanity's Last Exam, and remain competitive with models using far more activated parameters.

This paper introduces HiFi-UMI, a portable data-collection system that captures robot demonstrations without a physical robot, using head-mounted stereo-inertial SLAM, native inter-gripper pose estimation, microsecond-synchronized triggers, and ultra-wide-angle cameras to achieve 3 mm end-effector accuracy. The key finding is that this high-fidelity data alone can train policies that deploy directly on real robots, eliminating the usual need for a small real-robot fine-tuning 'anchor' dataset. Across three different model backbones, policies trained only on HiFi-UMI data matched or slightly exceeded teleoperation baselines (success-rate differences of -2.5, +3.1, and -0.6 percentage points), with the best policy hitting 85% on a precision insertion task. Additionally, pre-training on 4,000 hours of this data reduced action error on ten unseen tasks by 41% and boosted real-robot success by 18.1 percentage points on one model. The authors open-source 2,000 hours of this synchronized, wide-field-of-view data, offering a scalable alternative to expensive robot teleoperation for robot learning.

SwanTale is a unified model that generates both speech and audio for tasks like dubbing or podcasts, handling two modes: following a text caption that describes speakers, environment, and effects, or cloning a voice from a reference clip. The authors built a new dataset, SwanData-Caption, with cleaned data and detailed captions to train the model, and introduced components like SwanVAE for high-quality multi-modal output and a mixture-of-experts architecture to handle multiple tasks simultaneously. They also used reinforcement learning-style post-training to improve quality. Results show SwanTale outperforms existing systems on key metrics for both zero-shot voice cloning and instruction-following generation, and it can produce complex scenes with multiple speakers and audio effects from a single caption.

Long-horizon LLM agents often fail because they keep all task state and self-assessments in a growing context, letting early mistakes corrupt later decisions. This paper reframes the problem as task-state management and introduces LongHorizon-Harness, which stores state externally and only updates it with facts independently verified from the environment. Its Manage-Execute-Audit loop splits work into a manager that plans subtasks, an executor that runs each subtask with a fresh context, and a read-only auditor that checks the environment before proceeding. The harness is model-agnostic and works with existing agent frameworks. On benchmarks, it lifts Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench, 69.7% to 77.2% on Terminal-Bench 2.1, and 2.8% to 8.3% on OSWorld 2.0, with similar gains for Claude Opus 4.7, showing that separating state from execution is a broadly effective fix.

Vidu S1 is a real-time interactive video generation model that supports voice-controlled digital character animation with infinite-length output and high frame rate on consumer hardware.

This paper introduces RESOURCE2SKILL, a framework that automatically converts human-created resources like tutorial videos, code repositories, and articles into reusable, executable skills for software agents. Instead of relying on hand-written or text-only skill libraries, it builds a hierarchical multimodal Skill Wiki that combines text, code, and visual examples to capture complementary information from different sources. At inference time, agents retrieve and compose relevant skills, and can even acquire new ones online if needed. Across seven authoring domains, RESOURCE2SKILL improved task performance by an average of 11.9 percentage points over agents with no skills, and outperformed strong baselines in 26 out of 28 test scenarios.

RAGU is a new open-source GraphRAG engine that improves how AI systems retrieve structured knowledge by separating the extraction of entities and relations from their consolidation. Instead of a single noisy pass, it uses two-stage typed extraction, deduplication, and community detection to build cleaner knowledge graphs. The key insight is that the language skills needed for extraction—comprehension and reasoning—don't require huge models, so the authors trained a compact 7B model (Meno-Lite-0.1) that outperforms a 32B model on graph construction and matches it on English tasks. On medical benchmarks, RAGU achieves higher evidence recall (up to 0.84 vs. ≤0.76) and beats prior systems on synthesis tasks, while running on a single GPU.

Agentic abstention involves determining when an AI agent should cease interaction under uncertainty, requiring sequential decision-making across multiple environments and task types.

Boogu-Image-0.1 is an open-source family of models that can both understand and generate images, with variants for fast inference and instruction-based editing. While closed-source systems like GPT-Image-2 achieve strong results through undisclosed system-level tricks, this work shows that careful improvements in data quality, training pipelines, and inference-time scaling can match or approach their performance on a tight budget—using only 208 million unique images and an estimated $400K in training cost. The models match or beat other open-source alternatives on standard benchmarks and support bilingual text rendering in Chinese and English. The authors release all weights, code, and recipes under Apache 2.0 to help advance open multimodal AI.

Agentic systems can achieve open-ended improvement through multi-component co-evolution that progressively removes fixed human constraints across agents, environments, and evolution mechanisms.