SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery
SwarmResearch tackles a core failure mode of long-running coding agents: they often get stuck on one high-level strategy and then churn through low-level edits. A “Shepherd Agent” steers a population of subagents across separate git branches using global context, leading to better or comparable optimization results on 13/15 open-ended tasks by adapting exploration and parallelism to the right moments.
Claude Cowork expands to mobile and web
Anthropic is rolling out Claude Cowork beyond the desktop app, making the coding-agent workflow available on mobile and web with staged access for subscribers first. Sessions default to running in the cloud and support cross-device continuation, while the “full experience” still emphasizes desktop features like local file access.
MedCalc-Pro: Solving Complex Medical Calculations with LLM Agents
MedCalc-Pro introduces a medical-calculation benchmark that reflects real clinical complexity rather than single-tool, explicitly specified prompts. It tests multi-calculator selection and nested calculation flows with 2,268 real cases spanning 77 calculators, and proposes an agent framework that reduces cascading parameter errors via structured validation and evidence review.
Object-Centric Environment Modeling for Agentic Tasks
This work argues that free-form memories don’t scale for agents because they’re hard to validate and reuse as interactions grow. Object-Centric Environment Modeling (OCM) keeps two connected knowledge bases – object knowledge (as executable Python classes) and procedure knowledge (reusable interaction patterns) – and progressively discloses compact signatures before pulling code only when needed.
iFLYTEK-Embodied-Omni
iFLYTEK-Embodied-Omni presents a unified multimodal foundation model that jointly models vision, language, and action for embodied instruction following. Instead of cascaded pipelines that compound errors, the system shares multimodal self-attention across components and uses a “brain-cerebellum” split where high-level prediction and planning feed into direct action chunk generation.
ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability
Guiding reinforcement learning agents with an LLM under partial observability often fails because the “uncertainty gate” ends up measuring action uncertainty, not state uncertainty – and the prompt provides too little context to be useful. ASK+ fixes this with trajectory-aware prompting and structured reasoning, turning the SLM from a redundant checker into an occasional corrective consultant.
Automated Data Readiness for Scientific AI
REDI targets the unglamorous bottleneck behind scientific AI: turning massive datasets into AI-ready training material with provenance, validation, and deployable workflows. The open-source framework orchestrates ingest-to-output pipeline stages with instrumentation and agent-callable deployment, scaling to near-ideal parallel performance on real domains and highlighting file I/O and format choice as first-order optimization levers.
VERITAS: Towards a General-Purpose Replication Tool for Scientific Research
VERITAS builds a domain-agnostic replication engine that uses CLI coding agents to extract claims, run methodology with troubleshooting, and verify results against experimental evidence. Across 65 papers and multiple replication benchmarks, it produces patched codebases plus an importance-weighted Replication Score, showing state-of-the-art performance over strong Claude Code baselines.
Oyster-II: Reinforcement Learning for Constructive Safety Alignment in Large Language Models
Oyster-II extends constructive safety alignment by addressing two key limitations of earlier SFT-based approaches: weak generalization out-of-distribution and an over-application of “safety chain-of-thought” that harms helpfulness. The proposed RL-based framework uses a Zero-RL scheme with multi-stage strategy to improve safety performance while preserving user utility.
Solos AirGo A6 debuts a camera-less smart glasses design
Solos announced the AirGo A6 smart glasses with a major privacy-and-weight goal: no cameras and a voice-first AI assistant. The new glasses drop to around 19 grams, weighing far less than many competing smart glasses, with battery and speaker hardware packed into thinner temple arms.
LeRobot v0.6.0: Imagine, Evaluate, Improve
LeRobot v0.6.0 pushes forward open robot-learning by emphasizing a loop of imagination, evaluation, and improvement. The release focuses on better ways to generate, test, and refine robot behaviors, aiming to make embodied AI progress more systematic and reproducible.
Hugging Face Models on Foundry Managed Compute
This update shows how to run Hugging Face model workloads on Foundry Managed Compute, reducing the operational friction of scaling training or inference. The emphasis is on smoother deployment and compute management for teams that want production-like infrastructure without rebuilding end-to-end pipelines.
Run AI workloads on any cloud, store on Hugging Face: zero-egress storage with SkyPilot
SkyPilot’s “zero-egress storage” approach targets a practical pain point for AI teams: moving data across clouds is expensive and slow. The workflow integrates SkyPilot orchestration with Hugging Face storage to enable running jobs on other clouds while keeping data handling efficient.
The foundational elements of AI architecture that IT leaders need to scale
This guidance article distills what IT teams must get right as AI programs move from pilots to production scale. It focuses on the architecture decisions behind reliable model operations, governance, and integration – the stuff that determines whether AI deployments stay stable under real organizational load.
Australian Payments Plus moves faster with ChatGPT and Codex
Australian Payments Plus describes how it used ChatGPT Enterprise and Codex to accelerate development through complex payments workflows while keeping humans in control. The report frames the changes as productivity and quality improvements first, with operational judgment remaining a central requirement for correctness and compliance.