Latest AI updates

Simulation-based reinforcement learning improves AI psychiatric intake in role-played encounters

An arXiv preprint reports that multi-turn reinforcement learning in simulation improved a language model’s structured psychiatric intake performance compared with its identically prompted base model. In 120 encounters with psychiatrists role-playing standardized patients, independent psychiatrist evaluators preferred the trained agent in 76.7% of matched scenario pairs, with improvements in communication and differential diagnosis. In adversarial simulated presentations, under-triage of acute safety risks fell from 30.5% to 12.9%; the findings do not establish effectiveness with real patients, and residual gaps require safeguards and clinician oversight.

Why it matters: The study provides concrete evidence that simulation-based training can improve psychiatric conversational agents while highlighting safety limitations and the gap between role-played evaluation and real-world care.

TeamGym benchmarks LLM agents on team decisions with distributed information

Researchers introduced TeamGym, an environment for evaluating and training language-model agents on team decision-making tasks built from legal, finance, medical, and software engineering corpora. TeamGym uses a method to identify "pivotal information"—a minimal set of facts that is individually necessary and collectively sufficient for a correct decision—so it can control how information is distributed and measure whether agents integrate it. Across 8 language models and 12 source corpora, distributing pivotal information lowered success by 22.9 percentage points versus giving it to a single model, and requiring agents to extract that information themselves lowered success by a further 19.9 points. Reinforcement learning on final-decision loss improved decision quality and produced emergent coordination in participation, speaker selection, and communication.

Why it matters: It provides a concrete benchmark and training method for a key multi-agent weakness: sharing and integrating distributed information.

Amazon SageMaker adds training session APIs for custom LLM training loops

Amazon SageMaker AI introduced training session APIs that let developers write their own LLM training loops in Python while SageMaker handles GPU allocation, distributed execution, fault recovery and checkpointing. The APIs expose operations such as sample(), forward_backward() and optim_step() through ServiceClient, TrainingClient and SamplingClient, with billing per token processed. A companion cookbook provides SFT and RL recipes, and AWS reports a GRPO example on GSM8K raised held-out accura

Why it matters: It gives ML engineers and alignment researchers a concrete way to run custom SFT, RL and preference-optimization loops without provisioning or managing GPU infrastructure.

Perplexity and NVIDIA Researchers Introduce JOLT: Joint Teacher-Student RL for Post-Training

Perplexity and NVIDIA researchers propose JOLT (Joint On-policy Learning and Teaching), a post-training method where a single model is trained as both teacher and student, with the teacher receiving privileged information and also being optimized. JOLT+ adds the student's own outcome reward to the distillation signal. On MATH, JOLT matched GRPO accuracy while using 6.5x fewer processed training tokens according to the authors' fitted curves, and JOLT+ led GRPO across the five tested settings. Th

Why it matters: It offers a concrete post-training technique with claimed compute savings and benchmark comparisons that RL practitioners can evaluate.

AWS shows SFT-then-GRPO recipe to make GPT-OSS-20B reason in non-English languages

AWS details a fine-tuning workflow on Amazon SageMaker HyperPod that teaches GPT-OSS-20B to reason and answer in French, German, Spanish and Italian instead of defaulting to English. Supervised fine-tuning with LoRA raised reasoning-language accuracy from 22% to 100% on a 50-case evaluation, while adding GRPO on top of the SFT checkpoint lifted answer-language accuracy from 62% to 70%. A controlled Qwen2.5-7B experiment across multilingual reasoning and GSM8K tasks supports the argument that GRP

Why it matters: It offers a concrete, reproducible fine-tuning methodology and decision framework for choosing between SFT and GRPO when adapting open-weight reasoning models to non-English languages.

Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning for XLong-Horizon Agents

Alibaba's Qwen team introduced QwenGyre, an online reinforcement learning framework for extreme-long-horizon agent training using black-box harnesses. It uses elastic GPU cells that switch between rollout and training, and a trajectory processor that bounds branching histories to at most five samples per execution while scoring timed-out runs on preserved workspaces. Tested on Qwen 3.6 122B and Qwen 3.8 2.4T, it reports up to 1.85x end-to-end speedup over Colocate and NL2RepoBench improvement fr

Why it matters: It offers a detailed look at infrastructure and sampling techniques for training long-horizon coding agents, with reported speed and accuracy gains.

OpenJiuwen Unveils MANGO: A Data-Driven Framework for Optimizing Multi-Agent Flow Networks

Unlocking the Potential of Agent Swarms: OpenJiuwen&#39;s New Masterpiece in Multi-Agent Flow NetworksMay 13, 2026&nbsp;— While multi-agent systems hold immense promise for solving complex problems, their architectural nature is inherently prone to e<!--autointro-->

Alibaba PAI Open-Sources AgenticQwen: &quot;Dual Data Flywheels&quot; Propel 8B Model to Rival 235B Performanc

Alibaba’s Platform for AI (PAI) team has officially released and open-sourced&nbsp;AgenticQwen, a new series of small language models specifically engineered for industrial-grade tool usage. Available in&nbsp;8B&nbsp;and&nbsp;30B-A3B&nbsp;parameter s<!--autointro-->

Alibaba Unveils Qwen3-Max-Thinking: A Trillion-Parameter AI Model That Outperforms GPT-5.2, Claude O

On the evening of January 26, Alibaba officially launched Qwen3-Max-Thinking, its flagship reasoning model in the Qwen series. The new model has set new global records on multiple key benchmarks, outperforming leading models such as GPT-5.2, Claude O<!--autointro-->

Moonshot AI Quietly Launches Next-Gen Open Model: Kimi K2.5

Moonshot AI has officially rolled out its latest open model, Kimi K2.5, via a silent update to its web-based chat interface. The upgrade—deployed server-side—has seamlessly replaced the previous K2 model without requiring any manual intervention from<!--autointro-->

TRL: Divide-and-Conquer Value Learning Without TD Bootstrapping

Berkeley AI Research presents Transitive RL (TRL), an off-policy value-learning algorithm based on divide and conquer rather than temporal difference (TD) learning. TRL updates values via a transitive Bellman rule using candidate midpoints from dataset trajectories and expectile regression, avoiding explicit search over the state space. On OGBench offline goal-conditioned benchmarks including humanoidmaze and puzzle with 1B-sized datasets, TRL matches the best individually tuned n-step TD baseli

Why it matters: It offers a concrete new RL algorithm with benchmark evidence, which is useful for researchers working on scalable off-policy RL and long-horizon decision making.