This article reviews seven recent AI research papers, detailing how SpatialClaw enables code‑based spatial reasoning for vision‑language models, SkillWeaver introduces compositional skill routing, PreAct compiles agent actions into reusable state‑machines, and other works advance world‑model inference, self‑designing RL environments, collective skill‑tree search, and process‑aligned reinforcement learning for diffusion LLMs.
Large Language Modelsagent reasoningdiffusion models
This article surveys the latest on‑policy distillation (OPD) research, categorizing each work as either a reinterpretation of an existing problem or a modification of a different module, and highlights the experimental findings, design choices, and trade‑offs reported across the papers.
DrPO introduces a drifting‑field based, rank‑only reward mechanism for one‑step text‑to‑image models, enabling reinforcement‑learning‑after‑training without back‑propagating reward gradients; it speeds up training 3.51× versus DRaFT, works with non‑differentiable rewards, and improves generation quality on SD‑Turbo and SDXL‑Turbo.
The article recounts how Proximal Policy Optimization, initially dismissed by NeurIPS 2017 for limited novelty, later became a cornerstone of RLHF and large‑language‑model training, illustrating how academic evaluation can miss long‑term impact, with parallels to other once‑rejected breakthroughs such as LSTM, SIFT and Dropout.
OpenAI’s latest study shows that injecting just 5% of beneficial‑trait data into reinforcement‑learning training yields over 80% improvement across more than 50 alignment evaluations, revealing that a few underlying personality traits drive cross‑domain alignment and persist under adversarial pressure.
AI alignmentLarge Language Modelsadversarial robustness
The article explains how post‑training shifts AI investment from massive pre‑training compute to targeted, low‑cost fine‑tuning that lets enterprises embed proprietary business logic, outlines the ROI formula based on data exclusivity, eval formalizability, and task frequency, and presents real‑world case studies.
GLM‑5.2 replaces the Group Relative Policy Optimization (GRPO) algorithm with a critic‑based PPO approach for long‑horizon tasks, arguing that GRPO’s group comparison breaks down on variable‑length trajectories, a shift that has sparked vigorous debate across the reinforcement‑learning community.
DrPO introduces a ranking‑only reward that builds a drift field from on‑policy image samples to fine‑tune one‑step text‑to‑image models, achieving up to 3.51× faster training on large multimodal rewards, supporting non‑differentiable signals, and demonstrating superior quality across multiple benchmarks.
This article surveys the latest research on On-Policy Distillation for large language models, covering methods that improve training stability, self‑distillation frameworks, and detailed analyses of when and why OPD succeeds or fails, with concrete experimental results and practical insights.
Entropy-AwareLarge Language ModelsOn-Policy Distillation
The Kuaishou technology team reports that several of its papers were accepted at the prestigious ICML 2026 conference—including a spotlight paper on metaphor video understanding, works on causal discovery for irregular time series, image super‑resolution, large‑scale notification dispatch, full‑order ranking, phase‑aware MoE for RL, end‑to‑end e‑commerce search, spatial‑reasoning rewards, a unified SWE benchmark, video temporal grounding, and interpretable transformers—while also inviting attendees to visit their booth B101 in Seoul.
VibeThinker‑3B, a 3‑billion‑parameter language model built on Qwen2.5‑Coder‑3B, achieves scores within the range of 671 B‑parameter models on benchmarks such as LiveCodeBench, AIME26, IMO‑AnswerBench and GPQA, thanks to a two‑stage SFT, multi‑domain reinforcement learning, offline self‑distillation and a claim‑reliability (CLR) evaluator that together push its reasoning ability to the frontier.
Large Language ModelsVibeThinker-3Bbenchmark performance
The paper introduces Thinking‑Based Non‑Thinking (TNT), a reinforcement‑learning approach that sets a per‑question dynamic token ceiling for non‑thinking mode using the answer length from thinking mode, cutting reward‑hacking incidence to under 10% while boosting accuracy and cutting token usage by nearly half across several math benchmarks.
The CVPR 2026 paper introduces BeautyGRPO, a reinforcement‑learning framework that leverages the fine‑grained FRPref‑10K portrait‑retouching preference dataset and a novel Dynamic Path Guidance algorithm to simultaneously enhance skin texture, preserve identity features, and achieve superior aesthetic alignment, outperforming existing retouching models on objective metrics and user preference tests.
Despite a surge to over 140 trillion daily token calls in China, enterprises find general large models can answer but cannot execute business workflows, a gap Jiuzhang Yunji addresses with its AI Factory that combines reinforcement‑learning‑driven professional model production, a five‑capability training platform, and an Inference OS to industrialize AI at scale.
The paper reveals that outcome‑based reinforcement learning often traps LLM agents in an information self‑locking regime where weak action selection and belief tracking prevent proper credit assignment, and introduces AREW, a lightweight advantage‑reweighting method that restores active reasoning across multiple tasks and models.
HyVLA-0.5, an open‑source embodied VLA model from Tencent Robotics X, leverages over 10,000 hours of sub‑millimeter UMI demonstration data and a novel FlowPRO reinforcement pipeline to achieve more than 90% success on simulated and real‑world tasks, while supporting cross‑embodiment transfer and asynchronous deployment.
BudgetMem introduces a query‑aware budget‑tier routing mechanism for LLM agents, allowing the memory system to dynamically allocate computational resources based on query complexity and achieving a superior performance‑cost trade‑off on several benchmarks.
The article explains inference‑type large language models, how they differ from traditional models by breaking questions into reasoning steps, the shift from training‑time to test‑time compute, scaling‑law insights, validation techniques, proposal‑distribution tricks, and the detailed training pipeline of DeepSeek‑R1, while also discussing failed experiments and future directions.
The paper introduces CASTER, a multimodal AI system that uses Social‑CoT reasoning and the MEDEA framework to simulate diverse audience reactions, benchmarked on the large‑scale CASTER‑Bench dataset, and demonstrates superior performance over GPT‑5.2, Claude‑4.5‑Opus, and traditional VQA methods while already being deployed on Bilibili.
The article compiles 184 practical PINN innovations—including theory advances, new training paradigms, and integrations with Bayesian methods, reinforcement learning, Transformers, and graph neural networks—along with ready-to-use source code and starter resources for researchers seeking cutting‑edge physics‑informed neural network solutions.
Alibaba’s free Qianwen high‑school volunteer‑filling Agent combines a knowledge base of 3,000 schools, proactive calendar planning, persistent memory and reinforcement‑learning‑trained LLM to guide 12.9 million candidates, and its performance was stress‑tested with 400,000 simulated AI applicants.
OneReason introduces a systematic reasoning capability into industrial recommendation models through multi‑stage pre‑training, chain‑of‑thought fine‑tuning, and reinforcement learning, achieving significant gains in click‑through, revenue, and cross‑domain recommendation performance while preserving the underlying language abilities of the base model.
The OneReason report details how Kuaishou’s recommendation team injects reasoning into large‑scale recommender models through a four‑level pre‑training pipeline, chain‑of‑thought (CoT) fine‑tuning, and specialized reinforcement learning, achieving significant offline gains and a 10.33% exposure lift in a live A/B test.
The MANGO framework integrates reinforcement learning, text‑gradient updates, and a node‑skipping mechanism to jointly optimize workflow structure and execution policy in multi‑agent systems, achieving superior accuracy and dramatically lower compute costs across seven benchmark datasets.
The article details how 1688 transformed its platform from a traditional intent‑matching traffic hub into an Agentic AI system that understands business tasks, outlining a three‑step implementation of knowledge, trajectory and environment redesign, dual‑track evolution, novel evaluation methods, and the emerging role of product managers as evaluation engineers.
The article showcases five ICML 2026 papers from the Taotian Group that tackle core multimodal AI challenges—interactive video try‑on, high‑resolution vision, e‑commerce video reasoning, sparse‑reward reinforcement learning, and curriculum learning for large language models—detailing their problem statements, novel solutions, and strong experimental results.
This free collection gathers 127 top‑conference papers covering 17 large‑model research directions—from perception and decision to safety—providing PDFs, GitHub links, and a web interface to help AI engineers, researchers, and students stay up‑to‑date.
MetaAgent‑X integrates system design and execution within a single base model, using hierarchical rollout and stagewise co‑evolution to jointly train Designer and Executor roles, and achieves significant gains over single‑agent and prior multi‑agent baselines on math and code benchmarks.
The SE‑GA framework introduces hierarchical memory and a two‑stage self‑evolution training pipeline that enable GUI agents to retain critical context, learn from past successes, and achieve state‑of‑the‑art performance on long‑horizon benchmarks.
MetaAgent-X tackles the limitation of fixed‑executor multi‑agent systems by jointly training a Designer that creates lightweight Python‑based collaboration scripts and an Executor that runs them, using hierarchical rollouts and stagewise co‑evolution to improve both design and execution across math and code benchmarks.
A reinforcement‑learning framework called SOFisher, developed by teams from Fudan and Beijing Institute of Technology, enables intelligent, adaptive selection of field‑of‑view positions in costly spatial‑omics experiments, dramatically improving target detection efficiency and revealing disease‑relevant cellular niches with far fewer measurements.
ClawGym provides a complete open‑source framework for Claw‑style personal agents, linking a 13.5 K synthetic task dataset, black‑box rollout training, sandbox‑parallel reinforcement learning, and a rigorously verified benchmark of 200 tasks, and demonstrates that synthetic data can lift a 30 B model beyond a 235 B baseline.
The article examines IBM Research's Abstract‑CoT approach, which replaces verbose natural‑language chain‑of‑thought reasoning with a compact abstract token vocabulary, achieving up to an 11‑fold reduction in inference tokens while maintaining comparable accuracy across math, instruction‑following, and multi‑hop QA benchmarks.
Researchers tackling world‑model problems often rebuild data pipelines, environments, and baselines from scratch, but the open‑source stable‑worldmodel platform consolidates diverse dataset formats, SOTA baselines, hundreds of environments, and multiple solvers, offering a three‑step workflow with demonstrated storage and speed advantages.
The SenseNova-U1-8B-MoT-Infographic model dramatically improves AI‑generated infographics by enhancing dense‑text rendering, layout stability, and chart accuracy through targeted data, extended mid‑training, and reinforcement‑learning fine‑tuning, achieving top scores on BizGenEval and IGenBench and surpassing many commercial rivals.
Integrating diverse AI agent harnesses into reinforcement‑learning pipelines is notoriously labor‑intensive, but NVIDIA’s new Polar system inserts an API‑proxy layer that treats any harness as a black box, enabling seamless rollout recording and trajectory reconstruction, as demonstrated by dramatic performance gains on a 4B model across multiple harnesses.
AlphaCFG introduces a grammar‑based framework that defines a controllable search space for discovering syntactically valid, financially interpretable alpha factors, using syntax‑aware Monte‑Carlo tree search guided by value and policy networks, and demonstrates superior search efficiency and profitability on Chinese and US stock datasets.
Grammar Guided SearchMonte Carlo Tree SearchQuantitative Finance
DiffusionOPD introduces a unified on‑policy distillation framework for diffusion models that decouples single‑task online policy exploration from multi‑task capability integration, training expert teachers per task and distilling their skills into a single student model, achieving faster convergence and higher performance across composition, OCR, and aesthetic tasks.
Microsoft’s Lens team shows that a 3.8 B‑parameter image‑generation model can match or surpass 6 B‑plus models while consuming only about 19 % of the GPU compute, thanks to aggressive model compression, dense captioning, mixed‑resolution training, optimized VAE and language encoders, and targeted RL fine‑tuning.
The paper introduces the Trajectory Auto‑Regressive (TAR) model, which uses multi‑scale trajectory generation, a VQ‑VAE latent compression, and a state‑action fusion architecture to address granularity mismatch between fine‑grained decision steps and coarse‑grained feedback in online advertising, achieving over 12% CTR lift, smoother budget pacing, and faster inference compared to prior baselines.
The open‑source 35‑billion‑parameter Intern‑S2‑Preview model achieves scientific‑task performance comparable to trillion‑parameter models, thanks to full‑link “general‑specialized” training, reinforced‑learning scaling, and hardware‑aware optimizations, and it outperforms leading closed‑source models on benchmarks such as MolecularIQ and crystal‑structure generation.
The article explains how large‑language‑model post‑training suffers from rollout bottlenecks, introduces the Trajectory Balance with Asynchrony (TBA) framework that separates a Searcher from a Trainer, reuses off‑policy trajectories via a Trajectory Balance objective, and demonstrates up to 50× speed‑ups while preserving or improving performance on math reasoning, preference fine‑tuning, and automated red‑team tasks.
The SkyClaw‑v1.0 model from Skywork AI offers a free, soon‑to‑be open‑source large‑language model for agent applications that matches Claude Opus 4.6 in performance while cutting token costs dramatically, and the article details its benchmarks, training pipeline, and deployment recommendations.
SkyClaw‑v1.0, a domestically released Agent model, delivers benchmark scores that surpass many open‑source rivals and approach top‑tier closed models like Claude Opus 4.6, while offering a dramatically lower price and a frictionless deployment experience for developers.
The article analyzes why large language models cannot simply adopt AlphaGo’s Monte‑Carlo Tree Search, highlighting credit‑assignment difficulties, gradient‑variance explosion in multi‑step RL, and how AlphaGo’s tight integration of value and policy networks amortizes search in a way LLMs cannot replicate.
ATLAS introduces a discrete functional token that simultaneously serves as an agentic operation and a latent reasoning unit, enabling large multimodal models to perform visual tasks without external tools or intermediate image generation, and achieves competitive results through SFT‑plus‑RL training and a token‑level gradient‑anchor technique.
The paper introduces PACE, a Parameter‑Change based Unsupervised Environment Design method that evaluates training levels by the magnitude of induced policy‑parameter updates, offering a low‑variance, computationally cheap signal that consistently outperforms prior UED approaches on MiniGrid and Craftax benchmarks.
The paper introduces PACE, a Parameter Change Environment Design method that evaluates training levels by measuring induced policy parameter updates, offering a low‑variance learning‑progress signal that outperforms prior UED approaches on MiniGrid and Craftax benchmarks, achieving higher success rates and more stable generalization.
SkillOS reframes skill management for LLM agents as a long‑horizon reinforcement‑learning problem, letting a trainable Skill Curator automatically insert, update, or delete markdown‑based skills, which the frozen Agent Executor then consumes, improving memory‑free performance and cross‑task transfer.
Graphia, a new LLM‑based social simulation framework, leverages social graph data as high‑quality supervision to jointly align microscopic interaction predictions and macroscopic network structures, achieving significant gains on TDGG and IDGG benchmarks across three real‑world datasets.
OneModel 1.7 FrontoStria‑RL achieves a 99% average success rate on the LIBERO benchmark, surpassing π0.5, GR00T‑N1.5 and OpenVLA‑OFT, by introducing a Predictive Policy Latent that implicitly links world‑model understanding to action execution and is continuously refined through a reinforcement‑learning loop and a Retrieve‑then‑Steer memory mechanism.
MedScope proposes a "Think with Videos" paradigm that lets AI models actively locate and verify evidence in long clinical videos, using coarse‑to‑fine tool calling, evidence‑centric training data (ClinVideoSuite) and a grounding‑aware reinforcement learning objective, achieving superior performance on multiple video‑understanding benchmarks.
Evidence-based QAMedical Video AIlong video reasoning
The article compiles 238 cutting‑edge reinforcement‑learning ideas across 21 research directions, highlights recent breakthroughs such as Sutton’s Intentional Updates, and provides brief overviews of representative papers—including knowledge‑graph, Kalman‑filter, agentic, LLM‑driven, and world‑model approaches—along with links to the accompanying source code.
Composer 2.5, the latest AI‑coding model from Cursor, claims near‑par performance with Claude 4.7 Opus and GPT‑5.5 while delivering up to ten‑times higher efficiency and a pricing model of $0.5 per M input tokens and $2.5 per M output tokens, backed by novel reinforcement‑learning tricks, massive synthetic data, and a custom Muon optimizer with dual‑grid HSDP architecture.
The article reviews five years of LLM‑centric reinforcement learning, tracing the evolution from early Q‑learning to PPO, then to Direct Preference Optimization, Group Relative Policy Optimization, and finally multi‑agent RL, detailing each method’s mechanics, strengths, failure modes, practical considerations, and emerging open‑source toolchains.
HyperEyes introduces a unified‑location‑as‑search (UGS) action space, parallel data synthesis, and a dual‑granularity efficiency‑aware RL framework that enable multimodal agents to perform simultaneous multi‑target retrieval, dramatically reducing interaction rounds while improving accuracy and cost‑efficiency across benchmark evaluations.
A 30B‑A3B model, trained with reverse‑perplexity supervised fine‑tuning, two‑stage reinforcement learning, and a multi‑round generate‑verify‑revise inference loop, achieves gold‑medal performance on IMO, USAMO and IPhO contests using over 100 k token natural‑language reasoning without external tools.
Uni-Agent is an open‑source framework that unifies building, running, and training of general AI agents, offering extensible model, tool, and environment modules, scalable sandbox execution via veFaaS, live monitoring, and demonstrated performance gains on large‑scale coding‑agent experiments.
Cursor unveiled Composer 2.5, reporting benchmark scores comparable to Opus 4.7 and GPT‑5.5, a ten‑fold cost reduction, explicit use of Moonshot’s Kimi K2.5 as a base, new RL training techniques, and a partnership with SpaceXAI that multiplies compute power, all highlighted by Elon Musk’s retweet.
The paper introduces PreRL, which removes the input condition to directly optimize the reasoning trajectory (P(y)) of large language models, and combines it with standard RL in Dual Space RL (DSRL), achieving consistent gains on math and out‑of‑distribution benchmarks, faster training, and richer reasoning behaviors.
Composer 2.5, Cursor’s latest LLM, matches Claude Opus 4.7‑level capabilities while costing roughly one‑tenth as much, thanks to larger training scale, precise text‑feedback reinforcement learning, 25× more synthetic tasks, and a new Muon‑HSDP optimizer that boosts efficiency up to ten‑fold.
FineFT introduces a three‑stage ensemble reinforcement‑learning framework that tackles high‑leverage reward volatility and missing ability‑boundary awareness in crypto futures trading by using selective TD‑error updates, VAE‑based market‑state boundary detection, and a risk‑aware routing mechanism, ultimately outperforming twelve baselines on six financial metrics while cutting risk by over 40%.
The paper “When to Think, When to Speak” introduces Side‑by‑Side Interleaved Reasoning, a learnable disclosure policy that lets LLMs alternate between internal thinking and user‑visible answer fragments, reducing content latency while preserving or improving accuracy on math and scientific QA benchmarks.
The article traces the two decades‑long, parallel research lines that birthed video world models—dreaming agents in reinforcement learning and learning physics from human video—explains how they converged in 2024‑2025, evaluates current capabilities and limitations, and analyzes the $10 billion investment landscape and strategic moves by NVIDIA, OpenAI, and others.
The article analyzes SubQ, a new LLM architecture using Subquadratic Sparse Attention (SSA) to achieve a 12‑million‑token context window with linear compute scaling, delivering up to 52× speedup and costing just 5% of Opus while matching dense‑attention performance on long‑context benchmarks.
GIPO (Gaussian Importance Sampling Policy Optimization) replaces PPO’s hard clipping with a smooth Gaussian‑weighted trust region, achieving log‑space symmetry and bias‑variance balance that mitigates policy lag and utilization collapse, and demonstrates superior stability and sample efficiency on GridWorld, LIBERO, MetaWorld, and 7‑billion‑parameter VLA experiments.
In a deep interview, former Google TPU architect Reiner Pope explains that low‑concurrency fast‑mode services trade higher fees for faster streaming but are limited by memory‑bandwidth bottlenecks, that optimal concurrency balances compute and memory costs, and that pipeline‑parallel sparse expert models and reinforcement‑learning fine‑tuning introduce new inefficiencies and overtraining risks.
I²B‑LPO is an exploration‑enhancement framework for RLVR that branches rollouts at high‑entropy nodes, injects latent variables via pseudo self‑attention, and filters paths with an information‑bottleneck self‑reward, achieving up to 5.3% accuracy and 7.4% diversity improvements on multiple math reasoning benchmarks.
PsiBot demonstrates that, with a 100,000‑hour human‑operation dataset captured via exoskeleton gloves and ego‑vision, a world‑model (W0) and reinforcement‑learning policy (R2) can bridge the gap to robot control, offering a scalable alternative to costly teleoperation pipelines.
OneSearch‑V2 introduces a latent‑reasoning enhanced self‑distillation framework that augments query understanding with thought‑augmented CoT, aligns preferences via direct user behavior feedback, and achieves up to 4 % CTR lift and significant order growth without adding inference cost or latency.
Trajectory Balance with Asynchrony (TBA) separates sample generation (Searcher) from model updates (Trainer), uses a trajectory‑balance objective to incorporate off‑policy data, and achieves up to 50× speedup in large‑model RL post‑training while preserving or improving performance on math reasoning, preference fine‑tuning, and red‑team tasks.
LaST‑R1 presents a new embodied AI framework that inserts latent physical reasoning before action generation and jointly optimizes reasoning and control with LAPO, achieving 99.9% average success on the LIBERO benchmark after a single‑trajectory warm‑up and boosting real‑world task success from 52.5% to 93.75%, while showing superior generalization to unseen objects, backgrounds and lighting.
MathForge tackles the long‑standing question of which math problems deserve focus in reinforcement‑learning‑based training, introducing a difficulty‑aware optimizer (DGPO) and multi‑aspect question reformulation (MQR) that together prioritize harder‑but‑learnable questions, yielding consistent performance gains across model sizes and modalities.
DGPODifficulty‑Aware OptimizationLarge Language Models
This article analyses DeepMind’s three‑pronged AGI paradigm—combining neural networks, symbolic systems, and reinforcement learning—by dissecting AlphaGo, AlphaFold 2, Gemini, and the Genie‑Sima loop, mapping the biological inspiration, outlining engineering and safety challenges, and proposing research directions for large‑scale deployment in communication scenarios.
UI‑Venus‑1.5, an open‑source end‑to‑end GUI agent from Ant Group, tackles knowledge gaps, offline‑online mismatch, and high coordination costs by introducing mid‑stage training with a 10 B‑token corpus, online reinforcement learning, a “split‑then‑merge” model fusion, and a scalable DaaS layer, achieving top scores on VenusBench‑GD, ScreenSpot‑Pro and real‑world AndroidWorld tests.
The article proposes Heuristic Learning (HL) as a way to tackle continual learning’s catastrophic forgetting by using coding agents that iteratively refine rule‑based policies, showing empirical gains on Atari, MuJoCo, and VizDoom tasks and outlining HL’s benefits, challenges, and future integration with neural networks.
SkillOS addresses the one‑off nature of current LLM agents by introducing a closed‑loop system where a trainable Skill Curator continuously extracts, updates, and manages reusable skills from execution traces, leading to measurable gains in success rates, efficiency, and cross‑task generalization.
The article reviews the recent Intentional Updates framework—co‑authored by Turing laureate Richard Sutton—that redefines step‑size in streaming reinforcement learning using a 1967 NLMS‑style formula, details its algorithmic design, experimental validation, and remaining challenges.
In a detailed interview, Ted Xiao, former Google DeepMind researcher, walks through the existence‑proof, foundation‑model, and scaling eras of embodied robot learning, explaining the technical challenges, pivotal decisions, and the evolving role of large language and vision models in robotics.
The article analyzes why DeepSeek’s large‑model training struggles with Monte‑Carlo Tree Search, explains its use of Chain‑of‑Thought prompting, GRPO entropy‑boosting and rejection‑sampling fine‑tuning, compares these methods with Google’s OmegaPRM and PRM approaches, and proposes a concrete MCTS‑driven data‑generation pipeline to overcome the “chicken and liquor” trade‑off.
The HiLight approach inserts lightweight highlight tags into full-length inputs, training a small Emphasis Actor to score token importance and guide a frozen large language model, improving performance on tasks like recommendation and QA without modifying the solver, while keeping low latency and training cost.
OpenAI researcher Yong Jiayi introduces Heuristic Learning, a reinforcement paradigm that replaces gradient‑based neural network updates with code‑editing driven by GPT‑5.4, achieving the theoretical 864‑point Atari Breakout score and matching or surpassing PPO on multiple Atari and robot tasks.
Anthropic’s Natural Language Autoencoders (NLA) translate high‑dimensional LLM activation vectors into readable text, using an Activation Verbalizer and Reconstruction module trained via RL to maximize Fraction of Variance Explained, and reveal internal planning, language bias, tool‑call hallucinations, and hidden reasoning across multiple Claude models.
This tutorial walks through creating a custom 8×8 GridWorld environment in Gymnasium, implementing a Q‑Learning agent that learns to navigate from the top‑left corner to the bottom‑right goal while avoiding walls, and visualizing training curves, learned policies, and a performance comparison with a random agent.
RadixArk raised a $100 million seed round backed by top hardware and AI investors to turn the open‑source SGLang inference engine and the Miles RL framework into day‑0 standards, aiming to democratize AI infrastructure and eliminate bottlenecks from training to inference.
AI infrastructureHardware‑agnostic AIOpen-source inference
By learning a compact latent‑action space from paired image‑text and large‑scale text data, the authors reduce the RL search space from a vocabulary of over 150 k tokens to a 128‑codebook, enabling more efficient fine‑tuning of multimodal conversational agents and achieving consistent gains across several RL algorithms.
The DARA framework splits online advertising budget allocation into a few‑shot LLM reasoning stage and a fine‑grained optimizer stage, enhanced by a dynamically updated RL‑fine‑tuning algorithm (GRPO‑Adaptive), achieving significantly lower ROI variance than traditional baselines in both real and simulated environments.
The article provides a free compiled resource containing 190 essential AI Agent papers—from fundamentals to cutting‑edge topics—along with 321 Google‑released implementation cases and 500 open‑source agent applications, all with source code to help beginners and researchers quickly understand the field and reproduce results.
The paper identifies reward sparsity as the core obstacle for small language models in reinforcement‑learning‑based reasoning, proposes G²RPO‑A which injects high‑quality thinking trajectories and dynamically adjusts guidance length, and demonstrates large accuracy gains on math and code benchmarks such as Qwen3‑1.7B improving from 50.96 % to 67.21 % on MATH500 and from 46.08 % to 75.93 % on HumanEval.
PromptEcho computes a continuous reward for text‑to‑image generation by measuring how well a frozen vision‑language model can reconstruct the original prompt from the generated image, eliminating the need for annotated data or a trained reward model and outperforming prior methods across multiple benchmarks.
The LLMBeginner project from the MLNLP community offers a staged, project‑oriented learning path—covering big‑picture concepts, deep learning and reinforcement learning fundamentals, LLM theory and practice, and agent development—to guide beginners from fragmented resources to systematic mastery, with both concise and detailed versions hosted on GitHub.
This article provides a concise overview of a ten‑chapter reinforcement‑learning textbook, outlining the progression from basic concepts such as states and rewards to advanced algorithms like policy gradients and actor‑critic methods, and explains how each chapter builds on the previous ones.
The article presents the LWD (Learning While Deploying) framework, detailing its reinforcement‑learning‑driven data flywheel, the DIVL value‑evaluation and QAM policy‑optimization modules, and experimental results where a dual‑arm robot improves success rates by up to 17% and reduces cycle time by 23.75 seconds across eight real‑world tasks, surpassing strong baselines.
In May 2026, the AI sector saw a 77% rise in capital spending by the four biggest tech firms, Meta's acquisition of robot startup ARI, reinforcement‑learning advances boosting LLM inference, OpenAI's ChatGPT Images 2.0 launch, Tencent's Hy‑MT model outperforming Google, Microsoft's legal‑AI assistant, a 400B model running on iPhone, and notable research from CMU and independent scholars.
This article surveys the rapid evolution of reinforcement‑learning algorithms for large‑language‑model inference from early REINFORCE and PPO to newer approaches such as GRPO, RLOO, DAPO, CISPO, DPPO, ScaleRL and MaxRL, highlighting their design motivations, mathematical formulations, empirical trade‑offs and open research challenges.
The article analyzes how DeepSeek’s "极" bug and OpenAI’s recurring "goblin" output stem from unclean training data and an unintended reinforcement‑learning reward bias, showing how a persona‑specific habit leaked into general model behavior and how engineers responded.
LWD (Learning While Deploying) introduces a distributed multi‑robot reinforcement‑learning framework that continuously improves VLA policies during real‑world deployment, leveraging DIVL, QAM, dynamic n‑step TD and an asynchronous actor‑learner architecture to achieve over 90% success on five‑minute tasks and outperform traditional behavior‑cloning, HG‑Dagger and RECAP baselines.
The article highlights the rapid rise of reinforcement learning across major 2026 conferences, curates 181 RL papers from eight top venues, and provides detailed summaries of innovative works such as MSRL and MedVR, offering free access to the papers and code.
The article reviews the paper "Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond", introducing a two‑axis framework (capability levels L1‑L3 and law domains) to map diverse world‑modeling systems, highlighting that most current systems stall at L1, that explicit law encoding is crucial for long‑term stability, and that L3 represents the ultimate, self‑evolving model.