News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedAWS describes how Rocket Close built Supercharger to optimize title operations with agentic AI. The case shows enterprise agents being applied to document-heavy operational workflows where speed and consistency matter.
AWS published an architecture for intelligent document processing with generative AI services. The workflow focuses on moving from raw PDFs to structured insights, a persistent bottleneck for enterprise AI systems.
A new arXiv paper proposes the Theory of Mind Utility, a formal specification for mentalizing mechanisms. The work targets AI systems that need to model beliefs, goals, and perspectives in social or collaborative settings.
Google sued an alleged Chinese cybercrime operation accused of using AI to automate scam texts at large scale. The case shows how generative AI is being folded into fraud workflows that already operate across platforms and borders.
Blackwell systems led the first AgentPerf benchmark from Artificial Analysis. The result puts more attention on infrastructure built for agent workloads, where latency, throughput, and tool-heavy execution matter together.
LangChain details how Box AI built enterprise content agents with Deep Agents. The case shows how agent systems are moving from chat interfaces toward workflow layers that act across large stores of business documents.
OpenAI added new Academy courses focused on applying AI at work. The release reflects a growing push to turn AI adoption from tool access into practical training for everyday workflows.
A new arXiv paper proposes a hierarchical framework for synthetic question generation in RAG benchmarks. The work asks how fine-grained evaluation needs to be to reveal retrieval and reasoning failures.
Mistral is reportedly raising €3 billion at a €20 billion valuation. If completed, the round would give the French AI company more capital to compete in frontier models, enterprise products, and European AI infrastructure.
AWS published a guide for building a meeting prep and follow-up assistant using Amazon Quick and Cisco Webex MCP servers. The example shows MCP moving from developer demos into practical enterprise workflow automation.
AllenAI and Hugging Face introduced OLMo-Eval, an evaluation workbench for the model development loop. The project aims to make model evaluation more systematic while teams iterate on data, training, and release decisions.
OpenAI is adding flexible rate-limit resets for its Codex coding agent, according to The Decoder. The change points to a broader competition over how AI coding tools price long-running agent work, not just single prompts.
A Ukrainian test involving fully autonomous drones used in combat. The story underscores how quickly military AI questions move from policy debate to battlefield experimentation.
Google Research is studying how AI can help people understand possible skin conditions from visual and contextual signals. The work points to a careful role for consumer-facing medical AI: guidance and triage, not a replacement for clinicians.
Google launched a Colab command-line interface for developers, automation, and AI agents. The tool makes notebook environments easier to script and integrate into repeatable engineering workflows.
A new arXiv paper studies prefill awareness in large language models. The work examines how models use information available before token generation, a detail that matters for inference behavior and system design.
Opposition to US data centers cannot be reduced to foreign influence narratives. The story highlights a practical AI infrastructure constraint: communities are pushing back over power, water, land, and local control.
A new arXiv paper introduces a benchmark for evaluating AI agents on scientific challenges across scales. The work focuses on whether agents can help with complex research tasks rather than only solve narrow benchmark prompts.
A new arXiv paper introduces Pythagoras-Prover, an open-source family of Lean theorem provers built for practical compute budgets. The release includes autoregressive models and a proof-of-concept diffusion prover that refines proofs at inference time.
A new arXiv paper evaluates LLM lie detectors on model organisms with verified hidden beliefs. The work argues that many prior results are hard to interpret because the test models may not actually believe the opposite of what they say.
A new arXiv paper introduces PersonaDrive, a retrieval-augmented vision-language-action agent for closed-loop driving simulation. The system conditions non-ego traffic agents on human demonstrations collected under explicit driving styles.
A new arXiv paper presents Arbor, a multi-agent framework that uses a shared search tree as working memory. The system treats failures as useful signals while optimizing full-stack LLM inference systems.
A new arXiv paper proposes deployment-centered evaluation for clinical LLM systems. Instead of only scoring correctness on static benchmarks, the method predicts query-level risk that a user will reject the model response.
MIT researchers upgraded random utility models by using three-way preference comparisons instead of simple pairs. The work could make AI systems better at predicting human choices when trade-offs are subtle.
Avataar is pitching a distilled video generation model priced at half a cent per second. The company is positioning cultural adaptation and speed as core requirements for large consumer markets, not optional polish.
A new arXiv paper studies latent reasoning models that replace explicit chain-of-thought with continuous thoughts. The authors find that observable internal patterns can appear even when they do not causally explain model behavior.
A new arXiv paper introduces ToolSense, a diagnostic framework for testing parametric tool knowledge in LLM agents. The benchmark probes whether models understand tool semantics instead of merely retrieving valid tool names.
A new arXiv paper introduces Evoflux, an inference-time method for evolving executable tool workflows. The approach targets compact agents that struggle with schema validation, dependency tracking, and changing tool catalogs.
A new arXiv paper introduces paper-grounded figure-to-video generation and the FigTalk benchmark. The work focuses on narrated walkthrough videos that align explanations with specific regions of complex scientific figures.
A new arXiv paper introduces Shopping Reasoning Bench, an expert-authored benchmark for conversational shopping assistants. It evaluates multi-turn reasoning, subjective trade-offs, and criterion-level quality in retail conversations.
A new arXiv paper revisits decision support for agentic systems where AI agents act and humans or tools provide support. The framework focuses on when agents should seek help and how uncertainty should shape that choice.
A new paper argues that long-horizon research agents can choose bad scientific candidates when they optimize a single aggregate score. The failure appears when headline metrics improve while important slices or cohorts quietly get worse.
A new arXiv paper introduces PoQ-Judge, a reference-free framework for scoring LLM outputs in decentralized inference networks. The system uses smaller judge models and cascade evaluation to reduce cost while preserving useful quality signals.
A new arXiv paper frames AI behavior forecasting as a learnable task. Instead of relying on explanations of how a reasoning model works, the proposed forecasters learn from trajectories to predict what the system will do next.
A new arXiv paper introduces TreeSeeker, a tree-structured method for trial, error, and return in deep search. The work aims to improve how reasoning agents explore alternatives and recover from bad paths.
A new arXiv paper introduces MoCA-Agent, a code agent that verifies financial and numerical reasoning through claim-level markets. The system decomposes problems into atomic claims before synthesizing executable Python answers.
A new arXiv paper reports that architecture-aware reinforcement learning can make sliding-window attention competitive on math reasoning. The result suggests training strategy can narrow gaps that look like fixed architectural limits.
Nous Research released a Hermes Agent profile builder that combines identity, model setup, skills, and MCP servers in one dashboard flow. The launch shows how agent configuration is becoming a product surface rather than a developer-only setup task.
A new arXiv paper introduces SkillJuror, a framework for testing how the organization of agent skills affects behavior. The study compares progressive disclosure with flat skill files while holding the underlying task knowledge fixed.
A new arXiv paper proposes a structured LLM pipeline for pre-mediation in human negotiation. The system separates dialogue, preference prediction, critique, and summarization to support negotiations that often skip preparation because of cost or access.