This article reviews an open-source book 'Deep Understanding of AI Agent' that structures AI Agent engineering around the formula Agent = LLM + Context + Tools, offering 10 chapters and 103 runnable Python experiments covering context engineering, memory, tool use, multi-agent collaboration, and model post-training.
The article introduces OpenMLE, an open‑source full‑stack system for recursive self‑improvement (RSI) research, showing how a 35B Frontis‑MA1 model improves its Medal Average from 39.39% to 71.21% on MLE‑Bench Lite, surpasses GPT‑5.5+Codex, and details the mechanism hierarchy, task‑curation gym, trainable evolution operators, and experimental evidence that training and search gains combine additively.
This article presents Vime-Ascend, an open‑source reinforcement‑learning framework that merges slime's training design with vLLM's inference engine, details its three‑stage decoupled architecture, showcases NPU performance gains on Qwen3‑4B GRPO experiments, and provides step‑by‑step ModelArts deployment instructions.
AgentDoG 1.5 expands AI‑agent safety from final replies to complete execution trajectories, introducing the ATBench family for fine‑grained evaluation, a taxonomy‑guided DataEngine for high‑quality data generation, and demonstrating substantial safety gains in both SFT/RL training and online guardrail deployment with lightweight models.
VeRL-Omni introduces a universal reinforcement‑learning post‑training framework that extends the verl and vLLM‑Omni stacks to support diffusion transformers, hybrid AR‑DiT, and unified understanding‑generation models, offering high‑throughput multimodal rollout, flexible reward engines, modular trainers, and broad hardware compatibility.
VeRL-Omni is an open‑source RL post‑training framework built on verl and vLLM‑Omni that enables efficient, high‑throughput rollout and flexible reward computation for diffusion, AR‑DiT, and unified multimodal generation models, supporting diverse hardware, modular trainers, and demonstrating up to 14% latency reduction and high training throughput in benchmark experiments.
The article explains how Hermes structures the runtime execution of agents—capturing tool calls, context changes, results, and rewards—so that these trajectories can be evaluated, fine‑tuned, and fed into reinforcement‑learning loops, creating a continuous improvement cycle.
The authors show that high SFT accuracy does not guarantee strong RL performance because over‑fitting reduces output diversity, and they propose Adaptive Early‑Stop Loss (AESL), a diversity‑aware early‑stopping objective that dynamically weights token and subsequence losses, yielding consistently better RL results on multiple LLMs and math benchmarks.
This article shares practical insights and pitfalls from six months of applying reinforcement learning to fine‑tune large language models, covering exploration efficiency, training stability, model selection, and special considerations for thinking‑oriented agents.
Recent research shows that applying reinforcement learning to large language models can dramatically improve inference performance, but its effectiveness depends on the token distribution produced during pre‑training, prompting a novel rewrite of cross‑entropy as a single‑step policy gradient with controllable entropy parameters.
This article systematically reviews the main reinforcement‑learning algorithms—PPO, GRPO, DAPO, GSPO, and SAPO—used for fine‑tuning large language models, explaining why supervised fine‑tuning precedes RL, how each method improves training efficiency and stability, and what trade‑offs they entail.
This article examines the necessity of integrating Supervised Fine‑Tuning (SFT) with Reinforcement Learning (RL) for large language models, surveys alternating, sample‑reuse, simultaneous, and hint‑guided fusion methods, presents the underlying loss functions, and discusses practical trade‑offs such as entropy collapse and importance‑sampling corrections.
The article analyzes OpenAI’s five‑layer AGI roadmap, compares it with DeepMind’s ECEVS framework, and examines the technical progress from L1 to L5—including RL‑enhanced chain‑of‑thought, ReAct agents, deep research, and upcoming innovations—while highlighting the commercial implications of each stage.
This article examines how reinforcement learning fine‑tuning influences large language model reasoning, revealing that RL primarily amplifies pre‑trained capabilities, suffers from entropy collapse, and fails to push the model’s reasoning boundary, supported by extensive experiments on scaling laws, entropy analysis, and mitigation techniques.
The paper identifies that GRPO’s token‑level importance weighting introduces high‑variance noise causing instability in large‑scale language model RL training, and proposes GSPO, a sequence‑level importance sampling method that aligns with reward definitions, improves gradient stability, and yields higher training efficiency and better performance across benchmarks.
The article analyzes the DAPO technical report, showing how dynamic‑sampling pipelines and token‑level loss handling in SFT and RL training outperform ad‑hoc algorithm tricks, and compares the training dynamics of reinforce_baseline and GRPO with concrete code examples.
This article reviews a new risk‑sensitive trading agent that combines reinforcement learning with large language models to extract stock recommendations and news‑based risk scores, describes the extended CVaR‑PPO algorithm, presents extensive experiments on the FNSPID dataset, and discusses the resulting performance gains and future work.
This article analytically reconstructs OpenAI o1’s architecture, training pipeline, and inference workflow, exploring its reinforcement‑learning‑enhanced hidden chain‑of‑thought, multi‑model composition, scaling laws, reward modeling, and potential implications for future AI safety and small‑model strategies.
This article examines the limitations of traditional learning‑to‑rank for Ctrip hotel sorting, introduces reinforcement learning as a remedy, outlines three progressive implementation plans (A, B, C) with algorithm choices and engineering trade‑offs, and presents early experimental findings that demonstrate RL's potential to improve conversion rates.
