Tag
RL can assist robot data collection during demonstrations through a technique called Data Inoculation, improving data quality for both behavior cloning and RL, as shown in a new paper on stacking cubes.
A Google and University of Hong Kong paper proposes VeriCUA, which verifies computer-use agent tasks by checking saved files and app state instead of trusting the agent's own completion report, cutting false positives to 5.4% versus 24.8% for a VLM judge.
NVIDIA introduces Cosmos Framework, an open, modular platform for training, post-training, optimizing, and deploying world foundation models for physical AI, covering RL pipelines like DPO, GRPO, PPO, and serving via Cosmos NIM microservices.
A technical write-up introduces GPU-accelerated Weight Teleportation, a Blackwell-native delta weight sync pipeline that brings sampler pause time below 0.23s and end-to-end weight sync below 4s for large-scale RL, beating intra-cluster P2P RDMA for most workloads.
GR.inc publishes a detailed write-up on training agent swarms with reinforcement learning, introducing RLCR (reinforcement learning from collective rewards) — a shared-policy policy-gradient recipe with flat collective advantages and latency pressure designed to achieve collective inference-time scaling.
CUA-Master 提出嵌套双环递归改进框架，通过开放世界语料、自催化任务合成与反馈引导的快照回滚重试机制持续提升计算机使用代理（CUA）性能，其 27B 模型在 OSWorld Verified 等基准上取得竞争力表现。
The paper shows that a student model's traits can transfer purely through a judge's reward scores during GRPO post-training, even when outputs contain no semantic trace of the trait, and establishes a four-stage causal chain explaining when this subliminal transfer occurs, with implications for AI-safety dispositions.
SAT-ACT is a learning framework for CDCL branching that extracts action preferences from outcome-derived rollouts at the same solver state, trains a state-conditioned network to guide branching, and reduces search costs while improving PAR-2 on SAT Competition benchmarks.
The paper introduces CivHarness, a replay-grounded system for generating long-horizon feedback in Freeciv, and CivTelescope, an LLM judge that predicts which game position leads to a better outcome, addressing 'scoreboard traps' where in-game scores mislead. Using CivTelescope as a dense reward (CivMarshmallow), agents achieve 8.94-point mid-game gains, 77% above the best baseline.
RTL-Interact is a multi-task post-training framework that trains a single RTL-specialized LLM to generate RTL, construct EDA checks, and revise designs using tool feedback, paired with Pattern-Aware Policy Optimization (PAPO) for multi-task RL. On CVDP, RTL-Interact-32B reaches 48.7% weighted Pass@1 and 68.9% Pass@5 (surpassing DeepSeek-V3.2), while PAPO cuts average tool calls by 63.8% and the models outperform RTL-specialized baselines on VerilogEval v2 and RTLLM v2.
The paper introduces Noura, a framework that turns executable personal clinical specifications into supervision for LLM-based diet planning, enabling compact models like Qwen3-14B to dramatically improve multi-day meal plan compliance and reduce rule violations compared to direct generation by larger models.
The paper proposes CRoS, a CoT causality metric that rewards scrambling the order of reasoning steps without altering their content, avoiding confounding effects of traditional perturbation-based metrics. Evaluated with RLVR on Gemma-3-4B over Reasoning Gym, CRoS improves out-of-domain accuracy (46.8% vs 43.8%), CoT quality, and CoT faithfulness over existing causality and outcome rewards.
The paper introduces KEEPER, a comparative study of four auxiliary objectives for building control-sufficient textual memory in long-horizon LLM agents, finding that reward-related prediction yields the best task performance and introducing self-consistency bias as a measure of error persistence in recursive memory updates.
CoFiBreak 提出一种粗到细的多样性奖励机制用于 GRPO 训练的多轮越狱攻击，在保持高效攻击的同时显著提升攻击策略多样性，并可迁移到未见过的前沿模型。
Microsoft Mage Team presents Mage-VL-Cowork, a full training recipe (data synthesis from evolved harnesses, multi-stage SFT, and online hybrid RL) for learning hybrid computer use, with the 4B model achieving strong results on OSWorld-Verified relative to its size.
The paper proposes Li2SAO, a lightweight asynchronous RL method for LLM post-training that uses group-relative actor-critic bootstrapping to warm up the critic before switching to single-rollout GAE, catching up with or surpassing GRPO early in training and raising later performance ceilings.
rl_rob is an open-source, from-scratch Python library for teaching reinforcement learning, covering tabular TD methods through linear function approximation to deep value-based and policy-gradient methods with transparent, few-line update rules and a unified agent interface.
Microsoft AI introduces TalentRL, a strategy-guided reinforcement learning framework for hierarchical routing over procedural memory in LLM agents, where coarse-grained 'talents' guide selection of fine-grained skills. It improves success rates across seven QA benchmarks and achieves 99.3% on ALFWorld and 90.7 on WebShop across Qwen and Llama backbones.
An academic paper proposing correlated private quantization for communication-efficient, locally differentially private federated reinforcement learning in vehicular networks, coupled with post-quantum authentication; it reports 92.4% smaller authenticated uplink sizes and 14.3% lower mean travel time versus independent quantization.
The paper frames continual reinforcement learning with a contextual MDP that splits context into behavioral and physical components, and studies zero-shot reuse of world models in FourRooms tasks with Dyna-Q and TD-MPC, finding that physical-context changes substantially degrade reuse depending on the observation representation. It identifies representation generalization as a key bottleneck for reliable world-model reuse in continual RL.