Articles & Guides

FM² tackles a practical barrier in medical AI: foundation models need data from many institutions, but privacy rules often prevent central aggregation. The paper focuses on modality heterogeneity across clients and proposes a unified federated training framework.

A new ACM MM 2026 paper argues that strong benchmark scores in video LLMs do not necessarily prove visual understanding. It introduces the Visual Dependency Gap, a diagnostic that compares performance on original videos with black-screen inputs.

The paper introduces GFlowRL, a streamlined GFlowNet-style reinforcement learning method that removes the learned partition network and replaces it with an in-batch Monte Carlo estimate.

A new arXiv paper examines why on-policy distillation works in LLM post-training—and why it can fail. The study argues that OPD is most useful as an exploration catalyst whose success depends on the faithfulness of its guiding signal.

Boogu-Image-0.1 is an open-source family of unified multimodal models for image generation, editing, fast inference, and bilingual text rendering. The paper argues that careful data, training, and inference-time design can narrow the gap with closed systems under constrained compute.

A new arXiv paper introduces SPINE, an agentic framework designed to debug and deploy bimanual robots with less reliance on robotics experts. The work targets the often overlooked integration layer between AI “brains” and physical machines.

A new arXiv paper proposes an AI-native insurance framework for agentic AI deployments, linking autonomy, permissions, governance, and dependency risk to underwriting and contract design. The work frames insurance as both an operational cost and a potential governance mechanism for autonomous AI systems.

A new arXiv paper argues that reporting a single post-training FLOP number hides the real allocation choices behind RL training. It studies how compute is split across model size, rollout search, policy learning, and reward feedback.

A new arXiv paper presents BitMind Forensics, a deepfake detection system trained through an open adversarial competition that continually refreshes its data distribution. Across 19 public benchmarks, the dated BMF snapshot reports strong results on in-the-wild images, video, and AI-generated media.

A new arXiv position paper argues that autoformalization should evolve from translating isolated statements into building complete, machine-checkable theory libraries.

A new arXiv survey frames self-improving agents as adaptive systems that turn experience into accumulated capability gains. Its focus is not a single algorithm, but the full configuration of models, prompts, memory, tools, and control logic.

A new arXiv paper introduces Safe-Psych, a sequential benchmark for evaluating LLM behavior under evolving psychiatric evidence. The study finds that capable models often diagnose too early instead of asking for clarification or abstaining.

A new arXiv paper proposes OriginBlame, a provenance system designed to identify which training records belong to a given data contributor. The goal is to turn removal requests into precise forget sets for machine unlearning.

A new arXiv paper argues that attention-based KV cache eviction can over-preserve structural tokens in schema-heavy inputs such as nested JSON. The proposed retraining-free role-aware allocation method aims to keep more answer-bearing values under tight memory budgets.

A new arXiv paper uses persona vectors to audit open-weight language models beyond what prompt-based testing can reveal. The study maps which traits models express by default, which remain latent but steerable, and which resist standard extraction.

The paper introduces Mycelium, an active shared workspace designed to connect researchers and AI agents so that scientific observations and hypotheses can move to the people or systems able to act on them.

Agora explores whether large language model pretraining can move beyond tightly controlled datacenters. Its system coordinates heterogeneous, preemptible internet-connected GPUs through pipeline-parallel sharding and fault-tolerant collective operations.

Hugging Face disclosed a security incident in which an autonomous AI agent system allegedly carried out a multi-stage intrusion against part of its production infrastructure. The case highlights how data-processing pipelines and model platforms are becoming first-class security targets.

SDABench shifts the evaluation of LLMs in scientific data analysis from workflow completion to the capabilities needed to support scientific claims.

Microsoft is reportedly coaching its sales teams to position its own AI models and Copilot as more efficient, cost-effective, and enterprise-ready than rival offerings from OpenAI, Anthropic, and others.

Thinking Machines Lab has released Inkling, its first in-house open-weight model, positioning it as a customizable foundation rather than a one-size-fits-all chatbot. The move highlights a growing split between closed frontier AI services and enterprise-controlled model adaptation.