This article explains how robots learn manipulation tasks through simulation-based reinforcement learning, detailing the role of physics engines, the observation-action-reward loop, domain randomization for sim-to-real transfer, and why training environments are becoming critical products in robotics.
Duke PhD student Zhou Yufa presents Athanor, a system where four Opus 5.5 coding agents autonomously post-trained Qwen3-4B-Base on 8-GPU nodes for 104 hours, raising benchmark scores from 15.9 to 67.3 — matching Qwen's official Instruct model — without human guidance or teacher models.
Hong Kong University researchers identify implicit advantage symmetry as the root cause of GRPO's poor exploration and difficulty adaptation, proposing A-GRAE to dynamically break symmetry and achieve simultaneous gains in Pass@1 accuracy and Pass@k diversity across math and multimodal benchmarks without extra compute.
Tencent's Hunyuan Hy4 Flash model reduces average per-task cost to 1/20th of Hy4 and 1/13th of Hy4 preview on internal WorkIndex benchmarks, while achieving satisfaction scores near Claude Haiku 5.5 at a fraction of the cost through mixed attention architecture, distillation, RL token reduction, and 4-bit inference acceleration.
Researchers from BUPT, Peking University, and BAAI propose CineEvolve, a framework where a single visual language model acts as both director and critic to plan, diagnose, and revise storyboards via alternating reinforcement learning, achieving superior multi-shot consistency and narrative coherence in generated videos.
ByteDance and university researchers introduce Axion 1.0, an agent training method using just two-stage supervised fine-tuning and reinforcement learning, achieving 59% on Humanity's Last Exam and 62.2% on SEAL-0 with fully open-sourced data and code.
Tsinghua University's SkillMaster framework teaches LLM agents to create, update, and retain reusable skills through reinforcement learning, using counterfactual utility rewards and DualAdv-GRPO to jointly optimize task execution and skill management, achieving 98.7% on ALFWorld and 82.0% on WebShop with Qwen2.5-7B.
HKUST MMSense Lab introduces a visual-native agent framework that assigns unique IDs to images for tool chaining in multimodal deep search, and an on-policy data evolution method that iteratively refines question-generation rules using agent rollouts, boosting 8B and 30B model accuracy across eight benchmarks.
HKUST MMSense LabImage ReuseMultimodal Deep Search
This article details a robust gait control system for humanoid robots that estimates external forces without force sensors by combining a momentum observer with a learned residual network, enabling the robot to recover from pushes during walking.
FireFlash, a 300B-parameter terminal and coding agent trained via full-parameter RL on 64 NVIDIA B300 GPUs in under five days, achieves Terminal-Bench 4.0 Pass@1 of 46.2%—matching Fable 5 and GPT-6 Sol—while reducing per-task cost by over 30x, with gains transferring to DeepSWE and SWE-Bench benchmarks.
NVIDIA's two new papers introduce Skill2Env, which compiles 3,400 practitioner skills into 7,971 verifiable training tasks across 13 domains, and SoL-Pi, an auto-research loop that cuts token usage by 49% and API costs by 33% while retaining 93.7% performance on agent benchmarks.
Researchers introduce FireFlash, a 300B-parameter terminal and coding agent trained via full-parameter RL from GLM-5.3-Flash on 64 B300 GPUs in under five days, achieving Terminal-Bench 4.0 Pass@1 of 46.2%—matching Fable 5 and GPT-6 Sol—while reducing per-task cost by over 30x through optimized infrastructure, task filtering, reward shaping, harness design, and a corrected GRPO objective.
Researchers from Xiamen and Jilin Universities propose EvoMemState, a state-centric memory paradigm for personal agents that evolves through ontology-guided induction, reinforcement learning, and state-aware reasoning, achieving 57.6-58.9% accuracy on MemSyco-Bench and 85.6% complete retrieval rate across benchmarks.
Researchers from NTU, SJTU, and UTokyo propose UMM-Reflection, enabling unified multimodal models to self-correct generated images through native reflection and reinforcement learning, achieving 64.9% repair success (up from 20.6%) and GenEval 0.84 (from 0.72) without external verifiers at inference.
Touch-R1 introduces the first tactile-anchored physical reasoning model using reinforcement learning to enable cross-sensor tactile inference, correcting visual misjudgments by prioritizing tactile evidence, and achieves state-of-the-art results on the new TouchReason-Bench benchmark across nine metrics including physical property identification, ordinal reasoning, and cross-sensor consistency.
Firefly introduces a back-chaining pipeline that explores real MCP servers to generate verifiable tool-call tasks, builds a retrieval-augmented simulator for reproducible offline RL, and shows a 4B model reaching Claude-level performance on tool-use benchmarks.
Researchers from Microsoft Research and universities introduce WildAWM, a framework that synthesizes complete personal computer environments with CLI, GUI, and MCP interfaces using only open-source models, generates cross-interface tasks validated through a five-gate process, and trains Qwen models that improve on eight out-of-distribution benchmarks, even creating tasks that challenge GPT-6 Astra.
Agent World ModelComputer-Use AgentCross-Interface Tasks
ZTE's EmbodiedMind introduces a three-stage training paradigm (RSFT, IR-GRPO, Trie-GRPO) that enables efficient long-horizon planning for embodied agents, achieving 70.02% on 18 benchmarks with an 8B model and topping the SuperCLUE EmbodiedCLUE-VLA leaderboard with its 180B model, validated in simulation and on real robots.
Researchers from Shanghai Jiao Tong University and Xiaohongshu propose Harness-S, a framework that evolves complementary execution harnesses, uses them for selective rescue training on failed tasks, and adapts harness selection via a policy-aware router, enabling agents to internalize diverse problem-solving strategies and retain them in the base framework.
A 26-page survey from HUST, Tsinghua, and other institutions reviews how reinforcement learning integrates with diffusion models across language, vision, robotics, and scientific computing, covering methods like RLHF, DPO, GRPO, and challenges in reward design, exploration, and sparse rewards.
Former AMD AI VP Sharon Zhou, who reported to CEO Lisa Su, joins Anthropic to return to frontier model research after pioneering reinforcement learning for GPU kernel optimization at AMD, claiming the approach reduces human effort for breakthroughs by 1000x; her career spans Harvard, Stanford under Andrew Ng, founding Lamini, and MIT TR35 recognition.
StartLux-Decision introduces AutoThink, a decision model that predicts first and reasons only when confidence is below 0.9, using RLCD² reinforcement learning to optimize reasoning efficiency and probability calibration, evaluated on the new DecisionPivot benchmark.
ByteDance releases Axion 1.0, a 122B parameter model achieving 66.7 on FrontierScience Research benchmark, surpassing GPT-5.6 Sol's 63.3, using only two supervised fine-tuning stages and reinforcement learning on Qwen base models, with 35B weights open-sourced under Apache 2.0.
Researchers from Fudan University, CUHK MMLab, and Shanghai Jiao Tong University introduce ReTerm, a recursive self-improvement framework where code agents learn to diagnose and recover from their own terminal failures, turning self-generated error trajectories into training data that boosts 9B model performance on Terminal-Bench, SWE-bench, and OpenThoughts-TBLite by over 13 percentage points.
CMU researchers propose a unified training paradigm where pre-training, mid-training, supervised fine-tuning, and reinforcement learning share a single maximum likelihood objective, using data mixing and scheduling instead of stage-specific algorithms, and show improved reasoning and continual learning on puzzles, math, and code benchmarks.
LongCat-2.5 is a 1.6-trillion-parameter native multimodal mixture-of-experts model that activates 48B parameters per token, combining multi-level discrete visual tokens with large-scale long-horizon reinforcement learning to excel at visual coding, complex repository development, and creative 3D tasks, while achieving 86% higher training throughput on domestic AI accelerators and 43% faster inference via W4A4 quantization.
This article summarizes a paper introducing world models for generalist agents, framing them within sequential decision-making (POMDPs) and distinguishing prediction from action, covering control, estimation, policy learning, latent space dynamics, joint world-action modeling, and evaluation criteria for effective decision-making.
Researchers from Tianjin University and Alibaba propose MIPU, a reinforcement learning method that addresses training-inference mismatch in LLMs by directly optimizing inference-time performance, achieving higher accuracy and more stable training on Qwen3 models across math benchmarks.
Shanghai Jiao Tong University researchers introduce ZwZ2 and the EvidenceConnect framework, which synthesizes 20.4K single-image training samples requiring models to connect evidence across multiple regions, yielding 4B, 9B, and 27B models that surpass base models on 25 benchmarks and transfer gains to multi-image, spatial reasoning, and agent tasks.
Tsinghua researchers propose ActionOPD, an action-level online distillation method that aligns student agents with teacher decisions at the semantic action level rather than token level, enabling black-box cross-family distillation and self-distillation with improved decision-making on benchmarks like τ²-bench.
Cutting-edge AI projects now automate the most labor-intensive quant research steps — factor mining, synthesis, risk control, and portfolio management — using reinforcement learning, GFlowNets, and LLM-based closed-loop agents that target final portfolio performance, while independent audits reveal significant gaps between reported and replicable results.
Meituan and Peking University propose ClawLoop, a Harness-native framework closing the loop of task generation, model training, and diagnostic feedback, enabling agents to self-improve without stronger teacher models; experiments on Qwen3.6-27B show consistent gains across three benchmarks after three iterations, with SkillsBench rising from 46.3 to 57.2.
Fudan and NTU researchers propose SimpleVLN, a vision-language navigation framework that unlocks state-of-the-art performance by using native multimodal large language models without architectural changes, combining observation selection, action chunking, and a three-stage training pipeline for real-world robot deployment.
NUS and Kling AI researchers introduce DiffusionART, an online diffusion RL algorithm that eliminates forward-reverse consistency forces and uses a two-update scheme with trust-region masking to achieve 10-12x faster convergence than DiffusionNFT and Flow-GRPO-Fast on HunyuanVideo-13B, while supporting pairwise rewards with tiny group sizes.
NV-CoT replaces discrete tokenized bounding boxes with continuous numerical coordinate prediction via four lightweight linear heads and a Laplace policy for reinforcement learning, raising region localization IoU from 47.3 to 59.5 and improving high-resolution VQA accuracy on V* Bench and HR-Bench under both supervised fine-tuning and GRPO.
Continuous CoordinatesMultimodal Large ModelsNV-CoT
Moonshot AI releases Kimi K3.1 Pro, a 3-trillion-parameter model with 100k context that matches top closed-source models on coding, agent, and 3D benchmarks, while intentionally filtering training data to reduce vulnerability-research capabilities.
Moonshot AI releases Kimi K3.1 Pro, a 3-trillion-parameter model with 100k context, achieving near-top scores on coding, 3D creation, and financial benchmarks via massive RL post-training, while deliberately reducing cyberattack capabilities through targeted data removal, and planning open-weight release by November 4.
Researchers from HKU and Tsinghua propose VAST, an offline RL framework that adaptively chooses time horizons for value updates by stitching variable-length trajectory segments, improving long-horizon decision-making across 50 OGBench tasks.
Alibaba Qwen and Ohio State University propose TRACE, a rollout-guided quantization-aware training method that aligns training-side FP4 rounding with actual rollout quantization results, enabling stable FP4 W/A and KV cache rollout for MoE models with up to 5.4× decoding throughput gains and performance matching BF16 baselines.
IQuest Research releases technical report for IQuest-Q1, a 320B MoE model with 15B active parameters, 512K context window, recursive multi-token prediction, and agent-focused post-training, achieving strong benchmarks across coding, tool use, and reasoning tasks.
Baidu Baige immediately adapted RLinf v0.4, an open-source reinforcement learning framework for embodied intelligence, providing pre-built images for three scenarios: Cosmos3 fine-tuning on LIBERO, Pi0.5 evaluation via ApxInf, and sim-to-real collaborative training using RL-Co, while detailing RLinf v0.4's modular robot interfaces, four new real-robot algorithms, expanded model support, realistic simulation benchmarks, and domestic chip compatibility, alongside Baige's full-chain AI Infra optimizations across data pipeline, training, simulation, and inference.
Bilibili's Index-Translate series introduces a family of multimodal translation models supporting 150 languages across three sizes, detailing their training recipe using WSD scheduling, multi-expert reinforcement learning with adversarial reward supervision, model merging and distillation, plus specialized models for speech translation, syllable-controlled dubbing, and long-document translation, with extensive benchmark results.
Tencent's Terminal-RSI framework uses recursive self-rewrite and policy refinement phases to train a Qwen3.8-27B terminal agent that achieves 86.52% on Terminal-Bench 2.1, surpassing GPT-6 Sol and Kimi K3, while demonstrating improved long-horizon task completion through credit assignment and reward design innovations.
Xiaohongshu and Fudan University propose DR-Rubric, a framework where models perform deep research to gather evidence and synthesize atomic constraints as rubrics for reinforcement learning rewards, achieving competitive results on six benchmarks with only 1,000 training samples while revealing non-monotonic bootstrap dynamics and reward polarization as a stopping signal.
This article explains the four machine learning paradigms—supervised, unsupervised, semi-supervised, and reinforcement learning—with a visual mind map, compares classification, regression, clustering, and dimensionality reduction tasks, provides a simulated interview Q&A covering K-Means pitfalls and core differences, and demonstrates a real-time text classification service using Java libraries Tribuo, ONNX, Spring Boot, and HanLP.
SPADER introduces a step-wise GRPO framework with diversity-aware exploration rewards to improve long-tail answer coverage in multi-answer question answering, achieving state-of-the-art results across four benchmarks and two base models by refining credit assignment and incentivizing rare entity discovery.
DeepSeek releases DeepSeek-V4.1-Flash-1017 featuring new RLSF (Reinforcement Learning with Self-Feedback) technique, achieving gains across 11 benchmarks with faster, shorter outputs, and demonstrates end-to-end project generation from 3D modeling to game development.
DeepSeek open-sources DeepSeek-V4.1-Flash-1017 with new RLSF (Reinforcement Learning with Self-Feedback) technique, showing across-the-board gains on 11 benchmarks, faster and shorter outputs, and demonstrating end-to-end project generation like 3D modeling and game development, all under MIT license.
A large-scale empirical study comparing GRPO with 10 self-distillation methods across 4 tasks and 3 model sizes finds no method consistently beats GRPO; some methods show high final scores but suffer training instability, and compute-equivalent comparisons often reverse apparent gains.
Fudan and SJTU researchers open-source A3, a reinforcement learning framework that uses AST-based structured credit assignment to train a 14B CLI agent achieving 24.6% exact match on ShellOps, doubling baselines, and outperforming a 235B model on out-of-distribution tasks with minimal computational overhead.
Xiaohongshu's AllSpark team introduces CompoWorld, a compositional environment scaling framework that combines 448 reusable services with 10,130 tools to train agents on cross-application workflows, achieving a 9.17-point average improvement across eight benchmarks and tripling success rates on AutomationBench.
This 72-page survey explores how optimal transport provides geometric tools for comparing distributions in reinforcement learning, covering imitation learning, policy optimization, model-based RL, offline RL, robustness, and transfer, while analyzing design choices in cost functions, temporal handling, and computational trade-offs.
Boston Dynamics, once the viral robot demo leader, accumulated $1.3B losses over five years under Hyundai after three ownership changes, as its hydraulic hardware and hand-crafted control algorithms proved too costly and fragile against China's mass-produced electric actuators and reinforcement learning pipelines.
Tsinghua researchers propose Jev-LDE, a 1.7B model that performs a single local edit (Keep, Delete, Replace) on retrieved demonstrations for in-context learning, improving average 1-shot classification accuracy from 81.2% to 88.1% across multiple datasets and target LLMs with only 11% inference overhead.
Andrej Karpathy reveals a steep AI capability funnel where roughly 5,000 people globally use the most advanced internal AI systems — capable of week-long agent collaboration, zero-day discovery, and scientific research — while 6 billion have never touched an LLM and most professionals see only limited chatbot utility.
Xiaohongshu's AllSpark team and university researchers propose AdaptEvo, a framework combining confidence-adaptive reinforcement learning with evolving decision knowledge to teach AI agents evidence-based contextual decision-making for rule-governed tasks like content moderation, outperforming Gemini 3.8-Flash and GRPO on both current and changed rules.
Renmin University introduces AEWM, an editing world model that classifies and revises noisy agent actions in real time, improving long-horizon task performance across search, terminal, and software engineering benchmarks without fabricating tool feedback.
Amap had 16 papers accepted at NeurIPS 2026, covering multimodal large models, spatial intelligence, visual reasoning, and reinforcement learning, with highlights including a decomposed verifier for generative world modeling, gradient projection for continual multimodal learning, map-free transit route generation, and benchmarks for coding agents and travel planning.
3D reconstructionEmbodied AIMultimodal Large Models
Zhejiang University's HybridCUA framework enables computer-use agents to autonomously orchestrate GUI interactions and CLI commands through a unified bash action format, mixed-trajectory training data, and dual-reward reinforcement learning, achieving 53.6% accuracy on OSWorld with 56% fewer steps.
This article chronicles DeepSeek's paper timeline from 2023 to 2026, analyzing how each model generation optimized training compute, KV cache, and throughput via MoE, MLA, sparse attention, and conditional memory, shifting long-context competition from maximum length to per-token byte cost.
This article summarizes two high-impact ACL 2026 papers: DRP reduces reasoning token usage by 64% while improving accuracy via skill-aware step pruning, and SAGE doubles skill reuse success rates in agents through sequential rollout and skill-integrated rewards.
Meta FAIR veteran and Recursive Superintelligence co-founder Tian Yuandong recounts his journey from AlexNet debates to Go AI, Coconut, and grokking research, revealing how co-authoring a paper with GPT-5 boosted efficiency 6-10x and led him to predict AI will replace researchers within five years, while arguing humans must still initiate and verify ideas.
HOMURA introduces a reinforcement learning framework with KL regularization and dynamic syllable-ratio rewards to achieve precise length control in LLM translation for subtitles and dubbing, outperforming baselines on the Sand-Glass benchmark and deployed at scale on Bilibili.
The article details a rigorous agent evaluation framework for after-sales operations using GRPO, revealing why a 93% overall success rate masks critical failures in specific scenarios and safety guardrails, and how to structure evaluations to withstand interviewer scrutiny.
VideoGen-Agent, a Qwen3-VL-8B-based multimodal agent, learns via supervised fine-tuning and reinforcement learning to autonomously invoke retrieval, simulation, and verification tools for video generation, scoring 75.6 on VABench (19.1 points above base) and 86.1 with upgraded tools, while humans prefer its outputs 84.3% over Seedance 2.0.
SocialCoach, an open-source AI system from HKUST(GZ) and USTC, simulates high-pressure social scenarios with timed responses, quotes users' exact words for post-practice review, and uses a Qwen3-8B reinforcement-learning scheduler to personalize practice sequences, achieving 4.22/5 goal-match score in evaluations.
The article analyzes Xiaomi's MiMo model progress, clarifying its autonomous training foundation, distinguishing open-source reuse types, examining RL training costs, and evaluating talent impact versus commercial returns, arguing that technical output and organizational capability matter more than hype.
This survey introduces a five-level hierarchy (L0–L4) for agentic visual generation systems, classifying them by how far the controller can influence generation decisions—from fixed support to cross-task experience reuse—and proposes a level-conditioned evaluation protocol to isolate causal contributions of control scope.
ScienceBuddy is an open-source AI research agent that uses a recursive dual-loop architecture to continuously improve its harness and model weights from real user interactions, achieving 42.2% to 73.3% accuracy gains on scientific benchmarks across genomics, molecular biology, and pharmacology domains.
Agility Robotics' Digit humanoid robots are already moving boxes in warehouses for customers like Amazon and Spanx, using learning-from-demonstration and large-scale simulation training to achieve adaptive manipulation, with the company claiming 20% lower operating costs than human labor over five years and preparing a $2.5 billion SPAC listing.
This article breaks down six high-frequency GRPO interview questions from top Chinese tech companies, covering GRPO vs PPO trade-offs, group-relative advantage calculation with concrete numbers, handling all-correct/all-wrong sample groups, KL constraint mechanics, convergence monitoring priorities, and reward hacking detection via shadow evaluation.
An Oxford PhD thesis introduces three methods—VCD, COMET, and SPARTAN—that embed causal inductive biases into world models, achieving better adaptability, robustness, and interpretability across simulation, reinforcement learning, and language modeling tasks.
OmniVChat introduces a synthetic data pipeline (OmniVChat-Studio), benchmark (OmniVChat-Bench), and RL reward design (OmniVChat-RL) to train multimodal models for native audio-visual dialogue, achieving strong generalization from synthetic to real human interactions.
ULTRA unifies motion tracking and sparse goal control in a single policy for humanoid robots, combining physics-driven motion retargeting, policy distillation, and RL fine-tuning, validated on Unitree G1 with first-person point cloud perception.
TypeSafe AI's Jev model, marketed as a breakthrough non-autoregressive decision model, faces controversy after independent researcher Nandakishor Mukkunnoth reveals he published similar work a year earlier and released open-source Laya outperforming Jev on benchmarks, sparking debate on Hacker News about priority, open source vs proprietary AI, and attention asymmetry.
HSImul3R introduces a physics-in-the-loop framework that reconstructs simulation-ready human-scene interactions from sparse views, using scene-targeted reinforcement learning and direct simulation reward optimization to achieve stable physical interactions, validated on HSIBench and deployed on Unitree G1 robot.
VBVR-Pro introduces a comprehensive framework for native visual reasoning, featuring 300 tasks, 1.25M training samples in video and interleaved formats, verifiable scorers for 100 tasks, benchmarking of 30+ models, and demonstration that verifiable rewards enable reinforcement learning to improve visual reasoning capabilities.
This guide walks through training custom locomotion policies for the $399 Microduck bipedal robot using reinforcement learning in MuJoCo simulation on UCloud GPU instances, covering environment setup, walking and running policy training, ONNX export, simulation validation, and Hugging Face deployment.
OpenAI announced GPT-5.5 deprecation starting October 14, launched gift cards for subscription access, and executives teased a major release week, while user reports indicate GPT-6 Sol is already in gray-scale testing with strong reinforcement learning performance and speed advantages over Anthropic's Opus models.
This article analyzes five major paradigm shifts in large model post-training over the past six months, covering expert distillation, online distillation as RL alternative, RLVR refinements, SFT-RL distribution mismatch, and data quality as irreducible constraint, with specific papers and metrics.
Tsinghua-led research shows On-Policy Distillation (OPD) achieves over 70% of full-data performance with just one training example, revealing that data coverage saturates quickly while algorithmic absorption speed becomes the bottleneck.
Absorption RateData EfficiencyLarge Language Models
FinSMART introduces a market-aligned reinforcement learning framework that trains LLMs for financial sentiment analysis using realized market returns instead of human annotations, employing GRPO with a dual-filter trading reward to achieve 220% higher cumulative returns than the previous state-of-the-art FinDPO.
Researchers from Shanghai Jiao Tong University and Shanghai AI Lab introduce REAL, an ECCV 2026 framework that enables embodied agents to actively explore unknown environments, disambiguate vague user instructions through dialogue, and execute mobile manipulation tasks via a unified MCP tool interface, achieving 78.3% real-world success on a dual-arm robot.
Alibaba and Zhejiang University's Astar learns from AI systems' own Git history to propose evolution strategies, achieving 54–68% single-shot success rates versus 31% for GPT-5.5 and 32% for human experts, and delivering 23.6% offline HitRate and 4.86% GMV gains in Lazada's ad system through 20 fully automated iterations.
Researchers from Beihang University and Nanyang Technological University propose PhyFilter, a lightweight plug-and-play module that uses physical differential structures and real-time feedback to correct learning residuals, allowing robots to generalize across unseen terrains and disturbances without massive datasets.
MySQL 9.0 introduces three major features: native JSON multi-valued indexes delivering 370x faster array queries, Parallel Query V2 with work-stealing achieving 7-8x speedups on analytical workloads, and an ML-based AI query optimizer that cuts index misselection from 12% to 2% and boosts complex query performance by 25-40%.
AI query optimizerDatabase PerformanceJSON multi-valued index
SafeEvolve introduces a co-evolution framework where Agent Harness and Policy jointly learn from execution trajectories, reducing attack success rates to 0.79% on AgentDojo and 2.42% on Qwen3-4B while improving task utility, enabling continuous safety improvement from real-world experience.
vivo AI Lab introduces CAPO-SLT, a confidence-aware policy optimization method that stabilizes sign language translation by giving each token a dynamic clipping bound based on old-policy confidence, achieving state-of-the-art results on Chinese and American sign language benchmarks using only pose input.
SmartPhotoCrafter introduces a unified reasoning-to-generation framework for automatic photo enhancement: an Image Critic analyzes quality via chain-of-thought reasoning, then a Photographic Artist performs high-fidelity restoration and aesthetic retouching, unifying dehazing, deblurring, and color grading without explicit user instructions, trained via a three-stage strategy with multi-level rewards.
This article details AReaL-Ascend v1.0.5's LoRA RL capabilities, explaining how low-rank adaptation reduces memory overhead for large model reinforcement learning, describing two Megatron LoRA weight update modes (adapter sync vs. merge), and covering cross-node LoRA RL, MoE support, XCCL communication, and Qwen3.6-27B examples for practical deployment.
The paper introduces BehR, a behavior consistency reward for training text-based world models, showing that optimizing for agent decision alignment rather than text fidelity improves trajectory-level consistency across 16 configurations, reduces false positives in offline evaluation from 42.5% to 9.5%, and enhances lookahead planning for weaker agents.
AReaL-Ascend v1.0.5 introduces two major innovations: colocated training and inference on shared NPUs via Ray scheduling and AWEX IPC zero-copy, and Multi-Teacher On-Policy Distillation (MOPD) that fuses domain expert models into a single student using token-level teacher signals, demonstrated on Qwen3.6-27B with GRPO.
This article analyzes DeepSeek Harness from a post-training perspective, revealing how its architecture bakes enterprise evolution into environment shaping and trajectory sedimentation, validated by a self-evolution POC that identifies interface design, contract shape, and feedback quality as critical bottlenecks.
MOON 3.0 shifts e-commerce multimodal representation from direct encoding to explicit attribute reasoning, using Multi-head Modality Fusion, joint contrastive and reinforcement learning, and fine-grained residual enhancement to improve fine-grained retrieval and classification across multiple benchmarks.
EASE introduces evidence-anchored spatial attention supervision to multimodal reinforcement learning with verifiable rewards, using annotated evidence bounding boxes to guide model attention toward relevant image regions during training, improving accuracy on visual reasoning benchmarks without inference overhead.
The NTU PINE Lab introduces Facet-0, a multimodal robot foundation model that achieves 82% success in five real computer‑assembly tasks with 0.5 mm placement accuracy, reduces human intervention from 47% to 24%, and learns to recover from contact failures using a force‑synchronized dataset and reinforcement‑learning post‑training.
UniSteer introduces a noise‑inversion interface that lets human corrections directly train a lightweight noise actor, enabling a Vision‑Language‑Action model to improve real‑world task success from 20% to 90% within 66 minutes and outperforming DSRL and DAgger baselines.