What's moving in AI.

Search-R1 uses outcome-based RL to teach an LLM to issue its own search queries mid-reasoning, reporting 41 per cent and 20 per cent gains over RAG baselines on seven QA datasets.

Agent-RLVR adds agent guidance (strategic plans and error feedback) to RL from verifiable rewards, lifting SWE-Bench Verified pass@1 from 9.4 to 22.4 per cent, and to 27.8 per cent with an added reward model.

Group-in-Group Policy Optimization adds a step-level grouping mechanism on top of episode-level groups, reporting over 12 per cent (ALFWorld) and over 9 per cent (WebShop) gains over GRPO with no extra memory or rollout cost.

AgentGym-RL trains LLM agents end-to-end with multi-turn RL across diverse environments and introduces ScalingInter-RL, a curriculum that grows the interaction horizon; trained agents reportedly match or surpass commercial models on 27 tasks.

RLinf is an open-source RL training backbone spanning embodied VLA models and agentic reasoning, with FSDP and Megatron backends, PPO/GRPO/DAPO support, and a reported 2.4x throughput over existing frameworks.

NVIDIA walks through building RL-trained scientific agents using the open-source NeMo Gym and NeMo RL libraries, with REST-API training environments, GRPO, and verifiable rewards from code execution.

A April 2026 synthesis argues that the binding constraint on agentic RL is no longer model capacity but credit assignment across long horizons, and surveys the emerging fixes: process rewards, turn-level advantages, and hierarchical grouping that push the reward signal onto the actions that actually mattered.

NVIDIA is packaging the pieces of agentic RL, verifiable reward design, scalable environment-based evaluation, and synthetic-task generation, into its NeMo RL stack (NeMo Gym, NeMo Data Designer) so teams can post-train open models like Nemotron for domain-specific agent tasks without building the rollout infrastructure themselves.

τ-bench's pass^k metric, the probability that all k independent trials of a task succeed, is becoming the default lens on agent quality, because it exposes that agents which solve a task once often fail to solve it eight times in a row; even strong function-calling agents score under 25% at pass^8 on retail tasks.

VerlTool offers a unified framework for agentic RL with tool use, formalizing agent interactions as multi-turn trajectories with multimodal observations and reporting roughly a 2x rollout speedup by executing tool calls asynchronously so one trajectory's tool latency overlaps with another's computation, evaluated across six domains from math to software engineering.

Microsoft Research introduced Argos, a multimodal reinforcement learning approach that trains an agentic verifier to check whether an agent's reasoning actually aligns with what it observed over time, reducing visual hallucinations and producing more data-efficient agents for real-world tasks.

|

|

|

Microsoft released SkillOpt, a text-space optimizer that treats a plain markdown file as the trainable parameter of a frozen LLM agent, applying optimization discipline from weight training: learning rates, validation gates, batch sizes, and epoch schedules. It wins or ties on all 52 evaluated (model, benchmark, harness) cells.

LEAP wraps a general-purpose LLM in an agentic scaffold that grounds every step in the Lean compiler and iterates against verifier feedback. It solves all 12 problems from the 2025 Putnam Competition and increases LLM solve rates on IMO-style problems from under 10% to 70%.

Nvidia announced DLSS 5, a generative AI-driven neural rendering system that fuses structured 3D graphics data with probabilistic AI generation to produce photorealistic scenes in real time at 4K. CEO Jensen Huang calls it 'the GPT moment for graphics.'

At GTC Taipei in June 2026, Nvidia unveiled FOX (Factory Operations Blueprint) for agentic factories, and Foxconn is already deploying MoMClaw, a manufacturing multi-agent system that links machine signals and sensors to hundreds of specialized agents under a central orchestrator coordinating quality, logistics, and safety in natural language.

A new paper argues that compaction, the now-standard trick of summarizing a long agent trace and restarting the window, can quietly drop safety constraints established early in a run, because a summary optimized for task progress does not know to preserve a guardrail stated hundreds of steps ago.

Anthropic's engineering guidance frames context engineering as the successor to prompt engineering: curating the entire context state (instructions, tools, retrieved data, and message history) across a long agent run, with compaction, structured note-taking, sub-agent isolation, and just-in-time retrieval as the core techniques.