Ask or search anything...

This study empirically investigates the safety and security of autonomous language model agents deployed in interactive, real-world environments. Researchers uncovered emergent vulnerabilities such as unauthorized compliance, sensitive data leakage, resource exhaustion, identity spoofing, and the propagation of unsafe practices, arising from the complex interplay of LLM capabilities with autonomy and tool use.

This study systematically investigates how prompt templates, reward functions, and policy optimization algorithms impact the performance and stability of Deep Research agents. It identifies that a "Fast Thinking" prompt, an augmented F1+ reward, and the REINFORCE algorithm lead to superior accuracy and training robustness, culminating in Search-R1++ which improves average accuracy from 0.403 to 0.442 for Qwen2.5-7B.

DualPath introduces a dual-path KV-Cache loading architecture to overcome the storage bandwidth bottleneck in agentic Large Language Model inference. This system achieves up to 1.87x higher throughput in offline inference and 1.96x greater capacity in online serving compared to a baseline, while maintaining low latency.

VIDEO-REASON.COM introduces a very big video reasoning (VBVR) suite, a community-driven, large-scale dataset, and a verifiable evaluation benchmark designed to advance video model reasoning capabilities. A scaling study demonstrates that increased training data improves both in-domain and out-of-domain generalization, boosting a base model's performance by 84.6%, yet performance plateaus, indicating architectural limitations for human-level reasoning.

SkillOrchestra introduces a framework that orchestrates AI agents by explicitly learning and leveraging a "Skill Handbook" of capabilities and costs, moving beyond implicit policy optimization. This approach yields higher accuracy at lower costs, achieving up to a 10.0 exact match point improvement on QA tasks and an 8.0 percentage point increase in accuracy on complex agent orchestration benchmarks, while significantly reducing training and inference expenses by hundreds of times.

Researchers at New York University developed Solaris, the first action-conditioned video world model capable of simulating consistent multi-view observations from coordinated multi-agent interactions in Minecraft. The model generates stable, long-horizon gameplay and outperforms baselines, achieving a 20.8% VLM accuracy on building tasks and a 71.4% VLM accuracy on view consistency.

The "World Guidance" (WoG) framework from ByteDance Seed and The University of Hong Kong enhances Vision-Language-Action models by leveraging a compact "condition space" derived from future observations for precise action generation. This approach overcomes the trade-off between efficient future representation and fine-grained control, achieving an average 69.4% success rate on Google Robot tasks in simulation and demonstrating robust generalization across real-world out-of-distribution scenarios and effective learning from diverse human and UMI datasets.

NVIDIA researchers developed a data engineering framework called Terminal-Task-Gen to create high-quality synthetic training data for large language models interacting with terminal environments. The resulting Nemotron-Terminal models, despite their smaller scale (8B-32B parameters), achieve performance comparable to or exceeding significantly larger frontier models on the Terminal-Bench 2.0 benchmark by leveraging carefully curated trajectory data.

Researchers at NVIDIA, University of Toronto, Vector Institute, and Technion analytically demonstrate that Test-Time Training (TTT) with Key-Value (KV) binding is fundamentally a form of learned linear attention, resolving empirical contradictions with the prior "memorization" interpretation. This re-conceptualization leads to architectural simplifications and enables a fully parallel implementation, yielding up to a 4.0x inference throughput speedup for the TTT layer.

PyVision-RL introduces an open-weight reinforcement learning framework for agentic multimodal models, enabling dynamic Python-based tool use for images and videos. The framework effectively counters interaction collapse and achieves state-of-the-art performance on image benchmarks, while significantly improving visual token efficiency for video tasks (e.g., 7.3% accuracy increase on VSI-Bench with an order of magnitude fewer visual tokens).

Researchers from MIT CSAIL and Tsinghua University developed GeoPT, a pre-training framework that scales neural physics simulators by leveraging vast unlabeled 3D geometries augmented with synthetic dynamics. The approach reduced labeled data requirements by 20-60% and accelerated model convergence by up to 2x on industrial-scale benchmarks, including aerodynamics, hydrodynamics, and solid mechanics.

Ryan O'Dowd's dissertation introduces "Learning Without Training," a paradigm for machine learning that constructs models directly from data using mathematically derived kernels and functional analysis instead of iterative optimization, providing explicit error bounds and improving efficiency for supervised learning on manifolds, local transfer learning, and active classification tasks.

Google DeepMind's Aletheia agent, powered by Gemini 3 Deep Think, autonomously addressed the FirstProof challenge, a collection of ten research-level mathematical problems. The agent successfully solved 6 out of 10 problems, generating publishable-quality proofs without human intervention during the solution generation process.

SkyReels-V4 introduces a unified multi-modal foundation model that simultaneously performs video-audio generation, inpainting, and editing from diverse inputs, achieving high-resolution, long-duration content with synchronized audio. Developed by Skywork AI, the model ranked third on a public video arena leaderboard and achieved the highest overall human evaluation scores on a comprehensive benchmark against leading commercial systems.

Intuit AI Research developed Trace-Free+, a curriculum learning framework that generates high-quality tool descriptions for LLM-agents without needing execution traces at deployment, achieving a 74.9% query-level success rate on the TMDB RestBench dataset and maintaining robust performance with over 100 candidate tools.

An economic framework is introduced to analyze Artificial General Intelligence (AGI), positing that human verification bandwidth, rather than computational power, becomes the ultimate binding constraint on economic growth. The framework formalizes a 'Measurability Gap' between AI's execution capabilities and human verification capacity, outlining how this gap can lead to a 'Hollow Economy' without strategic interventions.

tttLRM introduces a novel large 3D reconstruction model that leverages Test-Time Training (TTT) and Large Chunk Test-Time Training (LaCT) blocks, enabling high-resolution, long-context (up to 128 views), and autoregressive 3D reconstruction from streaming visual inputs, while producing explicit 3D representations like 3D Gaussian Splatting with linear computational complexity.

Researchers from Cornell, Databricks, and Harvard developed OAPL, an off-policy reinforcement learning algorithm for Large Language Models designed to handle asynchronous training and policy lag. OAPL improves reasoning capabilities on math and code generation benchmarks while achieving up to 3x greater sample efficiency compared to on-policy methods.

Researchers from The University of Hong Kong and Tencent systematically investigated the interplay of data, reward, and optimization in training efficient large language models, uncovering a two-stage learning paradigm and distilling practical guidelines. Their methods enabled significant reductions in average response length, such as from 14.9k to 8.9k for Qwen3-0.6B, while maintaining or improving reasoning performance across various LLMs and tasks.

Labelbox researchers introduced the "Implicit Intelligence" framework and "Agent-as-a-World" methodology to evaluate AI agents' capacity for inferring unstated human intentions. Evaluations demonstrated that even frontier models satisfied all implicit requirements in less than half of the assessed scenarios, highlighting a significant gap in contextual understanding.