Ask or search anything...
Pangram 4 is an AI text detection model that provides token-level granularity for human, AI-assisted, and AI-generated content, significantly reducing the false positive rate on human text to 0.0041% and the false negative rate on AI text to 0.3396% while demonstrating robustness against adversarial 'humanizer' tools.
Researchers from Horizon Research, Frontis.AI, Tsinghua University, and others developed OpenMLE, a full-stack system enabling AI agents to recursively improve in machine learning engineering. The system, including the Frontis-MA1-35B model, achieved a 71.21% Medal Average and 0.8126 Human Rank on MLE-Bench Lite, improving by over 30 percentage points in Medal Average compared to its base language model, and demonstrated transferability to scientific AutoResearch tasks.
Researchers from MemTensor, Renmin University, National University of Singapore, Shanghai Jiao Tong University, and Tongji University propose Metis, a memory foundation model that integrates dynamic memory states and procedures directly into the model's architecture. This approach yielded improved average performance on memory operation and memory-based question-answering tasks compared to external memory baselines, particularly for complex queries, while maintaining computational efficiency by avoiding linearly growing context windows.
Explorative Modeling introduces a paradigm where generative models explore multiple candidate generations during training, optimizing against the best match to address multimodality and enable end-to-end generation. This method establishes "generative expressivity" as a new pretraining axis, leading to notable efficiency gains—for instance, achieving comparable image generation performance with 6.2 times less data—and enabling end-to-end robotics control with up to 256 times fewer inference steps than existing multi-step diffusion models.
Perception-Correction Distillation (PCD) introduces a novel methodology to enhance knowledge distillation in multimodal large language models by addressing credit assignment ambiguity for perceptual failures. The method combines downstream task failure signals with student-teacher disagreement on perception spans, yielding up to 4.28 points macro average accuracy improvement over uniform on-policy distillation in teacher-student transfer tasks.
TURBOVLA introduces an efficient vision-language-action model that achieves real-time robotic manipulation at 32 Hz with under 1 GB VRAM on an RTX 4090, directly mapping visual and linguistic features to actions. This approach matches or surpasses the manipulation performance of larger, LLM-centric systems, significantly lowering hardware requirements for deployment.
Adobe Research developed Chimera, a hybrid visual diffusion transformer architecture coupled with a systematic scaling framework, to efficiently process token-intensive visual inputs and enable long-context visual generation. The framework achieves up to 7.3x compute efficiency gains and allows for zero-shot video length extrapolation from 5-second training clips to 30-second videos with minimal quality degradation.
PHIZERO introduces a physical world model centered on a 'physical language' that represents state-transition patterns, explicitly separating the inference of world dynamics from pixel-level synthesis. This approach enables the generation of videos with improved physical coherence, scoring 41.2 on Physics-IQ Verified, and demonstrates strong performance in physical understanding tasks and zero-shot transfer across embodiments.
Kimi K3, a 2.8-trillion-parameter multimodal Mixture-of-Experts (MoE) model developed by Moonshot AI, establishes an "open frontier" for large-scale AI by unifying extreme pre-training scale with advanced agentic and long-horizon capabilities up to a 1-million-token context length. The model achieves an approximate 2.5x improvement in scaling efficiency over its predecessor and demonstrates frontier-level performance across coding, agentic, knowledge, reasoning, and vision tasks, often outperforming other open models and competing with leading proprietary systems at a fraction of the cost.
Researchers from the University of Maryland, College Park, Microsoft Research, and MBZUAI developed Weak-to-Strong On-Policy Distillation (W2S-OPD), a framework enabling strong large language models to improve by learning from individually weaker models. This method notably achieved 11.4% relative improvement in math reasoning and 3.7% in code generation over traditional on-policy distillation, even allowing the student model to surpass its weaker teacher.
Researchers from Babson College, University of Missouri, and University of Maryland, Baltimore County, constructed a counterexample to the Maxwell conjecture by demonstrating a configuration of five point charges that yields at least 24 non-degenerate critical points in its electrostatic potential. This directly refutes the conjecture's predicted maximum of (n-1)^2 = 16 critical points for n=5 charges.
MemHarness introduces a framework for large language model agents that moves beyond verbatim memory replay, enabling them to reconstruct and adapt retrieved past experiences based on the current interaction state. This approach achieved an 85.2% success rate on ALFWorld and a 75.6% task score on WebShop, outperforming leading closed-source models and static memory baselines.
BEACON introduces a training framework for agentic visual reasoning models that leverages Necessity-Aware Adaptive Reward (NAAR) and Hint-Guided Capability Expansion (HCE) to teach multimodal large language models (MLLMs) when and how to adaptively use external tools. The model achieves the highest average performance among open-source models on 13 diverse benchmarks, demonstrating significantly improved mode adaptiveness and a positive net tool effect.
Qwen-UI-Agent introduces a real-world-centric GUI agent, achieving a 92.2% success rate on real-device mobile tasks and integrating a hybrid GUI+CLI action space, an agent-driven data flywheel, and proactive service initiation for improved utility across digital environments. It demonstrated state-of-the-art performance across mobile, desktop, and web benchmarks, often outperforming frontier models and specialized GUI agents.
This study evaluated frontier AI agents on open-ended research problems from unpublished papers, graded by the original human authors. It revealed that while agents are proficient in research engineering, they consistently failed to produce publishable work due to poor research judgment, uncreative problem-solving, and ineffective backtracking.
Researchers from Zhejiang University and collaborators developed Living-Harness, a self-evolving agent framework that enables Large Language Models to learn persistent procedural repairs from interaction failures. This system utilizes an external harness state, integrating episodic memory and a state graph, to systematically accumulate workflow knowledge, leading to improved task success rates across interactive benchmarks like "τ²-Bench" and "MultiWOZ-2.4".
Echoverse, a framework from Microsoft Research, provides deep, evolving synthetic environments for training computer-use agents on complex, stateful tasks, enabling robust learning beyond limitations of live-web interaction. A 9-billion-parameter model trained on Echoverse achieved 67.1% average success across 14 evaluation splits and demonstrated the feasibility of reinforcement learning for computer-use agents.
ReToken enhances vision-language models for visual retrieval by introducing a lightweight, learnable token that explicitly identifies relevant visual information within long contexts by leveraging the VLM's value space. This approach achieves substantial improvements in recall and accuracy across multi-image QA benchmarks and demonstrates robust zero-shot transferability to long video understanding tasks.
Flux-OPD is an On-Policy Distillation (OPD) paradigm that integrates continuously evolving contexts as in-training supervision for large language models in open-ended domains. It leverages a theoretical analysis of the reverse KL objective to introduce strategies for stable and effective distillation, achieving superior performance across prompt optimization and medical question answering tasks.