Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning large language models with human preferences. While recent research has focused on algorithmic improvements, the importance of prompt-data construction has been overlooked. This paper addresses this gap by exploring data-driven bottlenecks in RLHF performance scaling, particularly reward hacking and decreasing response diversity. We introduce a hybrid reward system combining reasoning task verifiers (RTV) and a generative reward model (GenRM) to mitigate reward hacking. We also propose a novel prompt-selection method, Pre-PPO, to maintain response diversity and enhance learning effectiveness. Additionally, we find that prioritizing mathematical and coding tasks early in RLHF training significantly improves performance. Experiments across two model sizes validate our methods' effectiveness and scalability. Results show that RTV is most resistant to reward hacking, followed by GenRM with ground truth, and then GenRM with SFT Best-of-N responses. Our strategies enable rapid capture of subtle task-specific distinctions, leading to substantial improvements in overall RLHF performance. This work highlights the importance of careful data construction and provides practical methods to overcome performance barriers in RLHF.

Autonomous graphical user interface (GUI) agents powered by multimodal large language models have shown great promise. However, a critical yet underexplored issue persists: over-execution, where the agent executes tasks in a fully autonomous way, without adequate assessment of its action confidence to compromise an adaptive human-agent collaboration. This poses substantial risks in complex scenarios, such as those involving ambiguous user instructions, unexpected interruptions, and environmental hijacks. To address the issue, we introduce OS-Kairos, an adaptive GUI agent capable of predicting confidence levels at each interaction step and efficiently deciding whether to act autonomously or seek human intervention. OS-Kairos is developed through two key mechanisms: (i) collaborative probing that annotates confidence scores at each interaction step; (ii) confidence-driven interaction that leverages these confidence scores to elicit the ability of adaptive interaction. Experimental results show that OS-Kairos substantially outperforms existing models on our curated dataset featuring complex scenarios, as well as on established benchmarks such as AITZ and Meta-GUI, with 24.59\%$\sim$87.29\% improvements in task success rate. OS-Kairos facilitates an adaptive human-agent collaboration, prioritizing effectiveness, generality, scalability, and efficiency for real-world GUI interaction. The dataset and codes are available at https://github.com/Wuzheng02/OS-Kairos.

-- The efficient planning of stacking boxes, especially in the online setting where the sequence of item arrivals is unpredictable, remains a critical challenge in modern warehouse and logistics management. Existing solutions often address box size variations, but overlook their intrinsic and physical properties, such as density and rigidity, which are crucial for real-world applications. We use reinforcement learning (RL) to solve this problem by employing action space masking to direct the RL policy toward valid actions. Unlike previous methods that rely on heuristic stability assessments which are difficult to assess in physical scenarios, our framework utilizes online learning to dynamically train the action space mask, eliminating the need for manual heuristic design. Extensive experiments demonstrate that our proposed method outperforms existing state-of-the-arts. Furthermore, we deploy our learned task planner in a real-world robotic palletizer, validating its practical applicability in operational settings. I. INTRODUCTION In modern warehouse and logistics management, stacking boxes continues to be a common challenge. In the past, due to the smaller scale of trade and lower efficiency requirements, workers could rely on their experience to decide how each box should be placed. However, with the globalization of trade, there is a growing need for fast and stable box stacking, and a good solution for this is robotic palletization [1] [2].

Since the release of ChatGPT, large language models (LLMs) have demonstrated remarkable capabilities across various domains. A key challenge in developing these general capabilities is efficiently sourcing diverse, high-quality data. This becomes especially critical in reasoning-related tasks with sandbox checkers, such as math or code, where the goal is to generate correct solutions to specific problems with higher probability. In this work, we introduce Flaming-hot Initiation with Regular Execution (FIRE) sampling, a simple yet highly effective method to efficiently find good responses. Our empirical findings show that FIRE sampling enhances inference-time generation quality and also benefits training in the alignment stage. Furthermore, we explore how FIRE sampling improves performance by promoting diversity and analyze the impact of employing FIRE at different positions within a response.

Reinforcement Learning (RL) with unit test feedback has enhanced large language models (LLMs) code generation, but relies on sparse rewards provided only after complete code evaluation, limiting learning efficiency and incremental improvements. When generated code fails all unit tests, no learning signal is received, hindering progress on complex tasks. To address this, we propose a Process Reward Model (PRM) that delivers dense, line-level feedback on code correctness during generation, mimicking human code refinement and providing immediate guidance. We explore various strategies for training PRMs and integrating them into the RL framework, finding that using PRMs both as dense rewards and for value function initialization significantly boosts performance. Our approach increases our in-house LLM's pass rate from 28.2% to 29.8% on LiveCodeBench and from 31.8% to 35.8% on our internal benchmark. Our experimental results highlight the effectiveness of PRMs in enhancing RL-driven code generation, especially for long-horizon scenarios.

Reinforcement Learning (RL) plays a crucial role in aligning large language models (LLMs) with human preferences and improving their ability to perform complex tasks. However, current approaches either require significant computational resources due to the use of multiple models and extensive online sampling for training (e.g., PPO) or are framed as bandit problems (e.g., DPO, DRO), which often struggle with multi-step reasoning tasks, such as math problem-solving and complex reasoning that involve long chains of thought. To overcome these limitations, we introduce Direct Q-function Optimization (DQO), which formulates the response generation process as a Markov Decision Process (MDP) and utilizes the soft actor-critic (SAC) framework to optimize a Q-function directly parameterized by the language model. The MDP formulation of DQO offers structural advantages over bandit-based methods, enabling more effective process supervision. Experimental results on two math problem-solving datasets, GSM8K and MATH, demonstrate that DQO outperforms previous methods, establishing it as a promising offline reinforcement learning approach for aligning language models.

The development of robotic systems for palletization in logistics scenarios is of paramount importance, addressing critical efficiency and precision demands in supply chain management. This paper investigates the application of Reinforcement Learning (RL) in enhancing task planning for such robotic systems. Confronted with the substantial challenge of a vast action space, which is a significant impediment to efficiently apply out-of-the-shelf RL methods, our study introduces a novel method of utilizing supervised learning to iteratively prune and manage the action space effectively. By reducing the complexity of the action space, our approach not only accelerates the learning phase but also ensures the effectiveness and reliability of the task planning in robotic palletization. The experimental results underscore the efficacy of this method, highlighting its potential in improving the performance of RL applications in complex and high-dimensional environments like logistics palletization.

Efficiency and reliability are critical in robotic bin-picking as they directly impact the productivity of automated industrial processes. However, traditional approaches, demanding static objects and fixed collisions, lead to deployment limitations, operational inefficiencies, and process unreliability. This paper introduces a Dynamic Bin-Picking Framework (DBPF) that challenges traditional static assumptions. The DBPF endows the robot with the reactivity to pick multiple moving arbitrary objects while avoiding dynamic obstacles, such as the moving bin. Combined with scene-level pose generation, the proposed pose selection metric leverages the Tendency-Aware Manipulability Network optimizing suction pose determination. Heuristic task-specific designs like velocity-matching, dynamic obstacle avoidance, and the resight policy, enhance the picking success rate and reliability. Empirical experiments demonstrate the importance of these components. Our method achieves an average 84% success rate, surpassing the 60% of the most comparable baseline, crucially, with zero collisions. Further evaluations under diverse dynamic scenarios showcase DBPF's robust performance in dynamic bin-picking. Results suggest that our framework offers a promising solution for efficient and reliable robotic bin-picking under dynamics.

Reinforcement Learning (RL) offers a versatile framework for achieving long-term goals. Its generality allows us to formalize a wide range of problems that real-world intelligent systems encounter, such as dealing with delayed rewards, handling partial observability, addressing the exploration and exploitation dilemma, utilizing offline data to improve online performance, and ensuring safety constraints are met. Despite considerable progress made by the RL research community in addressing these issues, existing open-source RL libraries tend to focus on a narrow portion of the RL solution pipeline, leaving other aspects largely unattended. This paper introduces Pearl, a Production-ready RL agent software package explicitly designed to embrace these challenges in a modular fashion. In addition to presenting preliminary benchmark results, this paper highlights Pearl's industry adoptions to demonstrate its readiness for production usage.

Contact-rich manipulation is common in a wide range of robotic applications, including assembly [1, 2, 3, 4, 5, 6, 7, 8], object pivoting [9, 10, 11], grasping[12, 13, 14], and pushing [15, 16]. To accomplish these tasks, robots need to learn both the manipulation trajectory and the force control parameters. The manipulation trajectory guides the robot toward completing the task while physically engaging with the environment, whereas the force control parameters regulate the contact force. Incorrect control parameters can lead to oscillations and excessive contact forces that may damage the robot or the environment. Past works have tackled the contact-rich skill-learning problem in different ways. First, the majority of previous works [2, 10, 6, 7, 3, 17, 9, 18, 19] focus on learning the manipulation trajectories and rely on human experts to manually tune force control parameters. While this simplification has demonstrated remarkable performance in many applications, letting human labor tune control parameters is still inconvenient. Furthermore, the tuned parameter for one task may not generalize well to other task settings with different kinematic or dynamic properties. For example, assembly tasks with different clearances will require different control parameters.