Ask or search anything...

Meta AI researchers introduced Tuna-2, a unified multimodal model that performs visual understanding and generation directly from pixel embeddings, eliminating the need for pretrained vision encoders. This encoder-free architecture achieved competitive performance across nine VQA benchmarks and state-of-the-art results among native UMMs for image generation and editing tasks.

A collaborative group of researchers presents a synthetic argument for the emergence of 'learning mechanics,' a scientific theory of deep learning that aims to explain and predict neural network behavior through mathematical, first-principles calculations. It consolidates diverse theoretical and empirical evidence, suggesting a path toward a unified understanding of phenomena like scaling laws, training dynamics, and universal representations.

Researchers from Zhejiang University and Microsoft Research developed World-R1, a reinforcement learning framework that imbues existing text-to-video foundation models with robust 3D geometric consistency without architectural modifications. This approach led to a PSNR improvement of up to 10.23dB over baseline models and a 92% user preference for geometric consistency in generated videos.

DeepSeek-V4 introduces models that efficiently process contexts up to one million tokens through a hybrid attention architecture and optimized infrastructure, reducing single-token inference FLOPs by up to 73% and KV cache usage by up to 90% compared to its predecessor. The models achieve competitive performance across reasoning, coding, and long-context tasks, establishing new open-source benchmarks.

Researchers at IBM Research AI developed Abstract Chain-of-Thought (Abstract-CoT), a post-training framework that replaces lengthy verbalized rationales with a short sequence of discrete, abstract tokens. This method achieves comparable or improved performance over traditional Chain-of-Thought while reducing reasoning token usage by up to 12 times across various benchmarks and language models.

Kuaishou's OneRec Team developed Kwai Summary Attention (KSA), a hybrid attention mechanism that compresses historical context into learnable summary tokens to enable efficient processing of long input sequences in Large Language Models. This approach reduces KV cache cost to O(N/k) and enhances long-context retrieval, achieving a 5.81-point gain over full attention on RULER-128K and decreasing KV cache memory by 2.5 times at 128K context length.

This research introduces a "levels × laws" taxonomy for agentic world modeling, categorizing capabilities into Predictor, Simulator, and Evolver levels and world types into physical, digital, social, and scientific regimes. The framework unifies disparate research efforts, providing criteria for decision-usable simulation and autonomous model revision, and highlights the shift from passive prediction to active, adaptable environmental understanding for AI agents.

This research re-evaluates mixed-policy optimization methods for large language model (LLM) reasoning by identifying and correcting bugs in widely used supervised fine-tuning (SFT) training frameworks. It demonstrates that a correctly implemented SFT-then-reinforcement learning (RL) pipeline consistently outperforms current mixed-policy approaches on mathematical reasoning benchmarks, often with greater computational efficiency.

Joy Future Academy introduced EgoLive, an extensive open-source egocentric dataset featuring 1,680 hours of high-resolution stereo video and multi-modal annotations from 65,866 real-world human task episodes. This dataset provides 6-DoF motion tracking, semantic segmentation, 3D scene reconstruction, and hierarchical language descriptions to foster generalizable robot manipulation models, demonstrating broader semantic coverage and more accurate depth estimation compared to prior work.

Researchers from Alibaba Group and The Chinese University of Hong Kong introduced Temporal Curriculum On-Policy Distillation (TCOD), a framework designed to stabilize on-policy distillation for multi-turn autonomous agents by controlling trajectory depth. TCOD improved success rates by up to 15.71 points over vanilla OPD on ALFWorld and reduced total training time by nearly 32%.

GS-Playground is a high-throughput simulator combining a parallel physics engine with a memory-efficient 3D Gaussian Splatting renderer to create scalable photorealistic environments for vision-informed robot learning. It achieved 10,000 FPS rendering at 640x480 resolution and enabled successful zero-shot sim-to-real transfer for various tasks, including manipulation with a 90% success rate.

The Massachusetts Institute of Technology developed Hyperloop Transformers, an architecture integrating looped Transformers with strategic hyper-connections to achieve parameter efficiency. This approach yielded approximately 50% fewer parameters than depth-matched baselines while achieving lower perplexity, improved downstream task performance, and strong robustness to INT4 quantization.

Google's Vision Banana model, created by instruction-tuning a pretrained image generator, demonstrates that generative models can achieve state-of-the-art performance in both visual understanding and generation. It surpasses existing specialized models on tasks like semantic segmentation and metric depth estimation while maintaining high-quality image generation capabilities.

ClawMark is presented as a benchmark designed to evaluate language model agents functioning as persistent coworkers across multi-day, multi-turn workflows in dynamic environments requiring raw multimodal evidence. Evaluation of frontier models on ClawMark revealed that while partial progress is achievable, strict task success rates remain low, with models struggling particularly with detecting unannounced environmental changes and committing backend writebacks.

Researchers at Simon Fraser University and collaborators developed ReVSI, a meticulously rebuilt benchmark for evaluating Visual Spatial Intelligence in Vision-Language Models (VLMs), addressing critical validity flaws in previous evaluation datasets. ReVSI revealed that proprietary VLMs generally outperform open-source models on numerical tasks, while specialized fine-tuned 3D models often exhibit reduced or even negative performance gains, frequently hallucinating due to noisy annotations and data biases in prior benchmarks.

RecursiveMAS introduces a framework that integrates recursive computation into multi-agent systems, enabling agents to refine collaborative reasoning through iterative latent-space interactions rather than explicit text. This approach leads to average accuracy improvements of up to 20.2% and inference speedups of up to 2.4x compared to text-based recursive multi-agent systems, significantly reducing token usage.

This research introduces MMEB-V3, a comprehensive benchmark spanning text, image, video, and audio modalities, and OmniSET, a diagnostic framework designed to evaluate omni-modality embedding models' ability to interpret and enforce explicit modality constraints. Experiments demonstrate that current models frequently fail to reliably retrieve content in the instructed target modality, often exhibiting strong modality biases and insufficient, misaligned instruction-induced embedding shifts.

A protocol called Agent-Native Research Artifact (ARA) is introduced, reframing research contributions as machine-executable knowledge packages. This framework quantifiably improves knowledge extraction by 21.3%, boosts reproduction success by 7.0%, and accelerates early-phase research extensions for AI agents.

A new paradigm, Skill Retrieval Augmentation (SRA), is proposed to address the scalability of external skill utilization in agentic Large Language Models. Researchers at Tsinghua University developed SRA-Bench, a benchmark featuring 26,262 skills, and demonstrated that while external skills enhance agent performance, current LLMs face significant challenges in effectively incorporating and applying these skills under noisy conditions, lacking both relevance and need-awareness.