Pref-GUIDE: Continual Policy Learning from Real-Time Human Feedback via Preference-Based Learning

Training reinforcement learning agents with human feedback is crucial when task objectives are difficult to specify through dense reward functions. While prior methods rely on offline trajectory comparisons to elicit human preferences, such data is unavailable in online learning scenarios where agents must adapt on the fly. Recent approaches address this by collecting real-time scalar feedback to guide agent behavior and train reward models for continued learning after human feedback becomes unavailable. However, scalar feedback is often noisy and inconsistent, limiting the accuracy and generalization of learned rewards. We propose Pref-GUIDE, a framework that transforms real-time scalar feedback into preference-based data to improve reward model learning for continual policy training. Pref-GUIDE Individual mitigates temporal inconsistency by comparing agent behaviors within short windows and filtering ambiguous feedback. Pref-GUIDE Voting further enhances robustness by aggregating reward models across a population of users to form consensus preferences. Across three challenging environments, Pref-GUIDE significantly outperforms scalar-feedback baselines, with the voting variant exceeding even expert-designed dense rewards. By reframing scalar feedback as structured preferences with population feedback, Pref-GUIDE offers a scalable and principled approach for harnessing human input in online reinforcement learning.

InterMT: Multi-Turn Interleaved Preference Alignment with Human Feedback

As multimodal large models (MLLMs) continue to advance across challenging tasks, a key question emerges: What essential capabilities are still missing? A critical aspect of human learning is continuous interaction with the environment -- not limited to language, but also involving multimodal understanding and generation. To move closer to human-level intelligence, models must similarly support multi-turn, multimodal interaction. In particular, they should comprehend interleaved multimodal contexts and respond coherently in ongoing exchanges. In this work, we present an initial exploration through the InterMT -- the first preference dataset for multi-turn multimodal interaction, grounded in real human feedback. In this exploration, we particularly emphasize the importance of human oversight, introducing expert annotations to guide the process, motivated by the fact that current MLLMs lack such complex interactive capabilities. InterMT captures human preferences at both global and local levels into nine sub-dimensions, consists of 15.6k prompts, 52.6k multi-turn dialogue instances, and 32.4k human-labeled preference pairs. To compensate for the lack of capability for multi-modal understanding and generation, we introduce an agentic workflow that leverages tool-augmented MLLMs to construct multi-turn QA instances. To further this goal, we introduce InterMT-Bench to assess the ability of MLLMs in assisting judges with multi-turn, multimodal tasks. We demonstrate the utility of \InterMT through applications such as judge moderation and further reveal the multi-turn scaling law of judge model. We hope the open-source of our data can help facilitate further research on aligning current MLLMs to the next step. Our project website can be found at https://pku-intermt.github.io .

Mitigating Deceptive Alignment via Self-Monitoring

Modern large language models rely on chain-of-thought (CoT) reasoning to achieve impressive performance, yet the same mechanism can amplify deceptive alignment, situations in which a model appears aligned while covertly pursuing misaligned goals. Existing safety pipelines treat deception as a black-box output to be filtered post-hoc, leaving the model free to scheme during its internal reasoning. We ask: Can deception be intercepted while the model is thinking? We answer this question, the first framework that embeds a Self-Monitor inside the CoT process itself, named CoT Monitor+. During generation, the model produces (i) ordinary reasoning steps and (ii) an internal self-evaluation signal trained to flag and suppress misaligned strategies. The signal is used as an auxiliary reward in reinforcement learning, creating a feedback loop that rewards honest reasoning and discourages hidden goals. To study deceptive alignment systematically, we introduce DeceptionBench, a five-category benchmark that probes covert alignment-faking, sycophancy, etc. We evaluate various LLMs and show that unrestricted CoT roughly aggravates the deceptive tendency. In contrast, CoT Monitor+ cuts deceptive behaviors by 43.8% on average while preserving task accuracy. Further, when the self-monitor signal replaces an external weak judge in RL fine-tuning, models exhibit substantially fewer obfuscated thoughts and retain transparency. Our project website can be found at cot-monitor-plus.github.io

Generative RLHF-V: Learning Principles from Multi-modal Human Preference

Training multi-modal large language models (MLLMs) that align with human intentions is a long-term challenge. Traditional score-only reward models for alignment suffer from low accuracy, weak generalization, and poor interpretability, blocking the progress of alignment methods, e.g., reinforcement learning from human feedback (RLHF). Generative reward models (GRMs) leverage MLLMs' intrinsic reasoning capabilities to discriminate pair-wise responses, but their pair-wise paradigm makes it hard to generalize to learnable rewards. We introduce Generative RLHF-V, a novel alignment framework that integrates GRMs with multi-modal RLHF. We propose a two-stage pipeline: $\textbf{multi-modal generative reward modeling from RL}$, where RL guides GRMs to actively capture human intention, then predict the correct pair-wise scores; and $\textbf{RL optimization from grouped comparison}$, which enhances multi-modal RL scoring precision by grouped responses comparison. Experimental results demonstrate that, besides out-of-distribution generalization of RM discrimination, our framework improves 4 MLLMs' performance across 7 benchmarks by $18.1\%$, while the baseline RLHF is only $5.3\%$. We further validate that Generative RLHF-V achieves a near-linear improvement with an increasing number of candidate responses. Our code and models can be found at https://generative-rlhf-v.github.io.

LAPP: Large Language Model Feedback for Preference-Driven Reinforcement Learning

We introduce Large Language Model-Assisted Preference Prediction (LAPP), a novel framework for robot learning that enables efficient, customizable, and expressive behavior acquisition with minimum human effort. Unlike prior approaches that rely heavily on reward engineering, human demonstrations, motion capture, or expensive pairwise preference labels, LAPP leverages large language models (LLMs) to automatically generate preference labels from raw state-action trajectories collected during reinforcement learning (RL). These labels are used to train an online preference predictor, which in turn guides the policy optimization process toward satisfying high-level behavioral specifications provided by humans. Our key technical contribution is the integration of LLMs into the RL feedback loop through trajectory-level preference prediction, enabling robots to acquire complex skills including subtle control over gait patterns and rhythmic timing. We evaluate LAPP on a diverse set of quadruped locomotion and dexterous manipulation tasks and show that it achieves efficient learning, higher final performance, faster adaptation, and precise control of high-level behaviors. Notably, LAPP enables robots to master highly dynamic and expressive tasks such as quadruped backflips, which remain out of reach for standard LLM-generated or handcrafted rewards. Our results highlight LAPP as a promising direction for scalable preference-driven robot learning.

Empirical Analysis of Sim-and-Real Cotraining Of Diffusion Policies For Planar Pushing from Pixels

In imitation learning for robotics, cotraining with demonstration data generated both in simulation and on real hardware has emerged as a powerful recipe to overcome the sim2real gap. This work seeks to elucidate basic principles of this sim-and-real cotraining to help inform simulation design, sim-and-real dataset creation, and policy training. Focusing narrowly on the canonical task of planar pushing from camera inputs enabled us to be thorough in our study. These experiments confirm that cotraining with simulated data \emph{can} dramatically improve performance in real, especially when real data is limited. Performance gains scale with simulated data, but eventually plateau; real-world data increases this performance ceiling. The results also suggest that reducing the domain gap in physics may be more important than visual fidelity for non-prehensile manipulation tasks. Perhaps surprisingly, having some visual domain gap actually helps the cotrained policy -- binary probes reveal that high-performing policies learn to distinguish simulated domains from real. We conclude by investigating this nuance and mechanisms that facilitate positive transfer between sim-and-real. In total, our experiments span over 40 real-world policies (evaluated on 800+ trials) and 200 simulated policies (evaluated on 40,000+ trials).

PhysGen3D: Crafting a Miniature Interactive World from a Single Image

Envisioning physically plausible outcomes from a single image requires a deep understanding of the world's dynamics. To address this, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene. By combining advanced image-based geometric and semantic understanding with physics-based simulation, PhysGen3D creates an interactive 3D world from a static image, enabling us to "imagine" and simulate future scenarios based on user input. At its core, PhysGen3D estimates 3D shapes, poses, physical and lighting properties of objects, thereby capturing essential physical attributes that drive realistic object interactions. This framework allows users to specify precise initial conditions, such as object speed or material properties, for enhanced control over generated video outcomes. We evaluate PhysGen3D's performance against closed-source state-of-the-art (SOTA) image-to-video models, including Pika, Kling, and Gen-3, showing PhysGen3D's capacity to generate videos with realistic physics while offering greater flexibility and fine-grained control. Our results show that PhysGen3D achieves a unique balance of photorealism, physical plausibility, and user-driven interactivity, opening new possibilities for generating dynamic, physics-grounded video from an image.

History-Guided Video Diffusion

Classifier-free guidance (CFG) is a key technique for improving conditional generation in diffusion models, enabling more accurate control while enhancing sample quality. It is natural to extend this technique to video diffusion, which generates video conditioned on a variable number of context frames, collectively referred to as history. However, we find two key challenges to guiding with variable-length history: architectures that only support fixed-size conditioning, and the empirical observation that CFG-style history dropout performs poorly. To address this, we propose the Diffusion Forcing Transformer (DFoT), a video diffusion architecture and theoretically grounded training objective that jointly enable conditioning on a flexible number of history frames. We then introduce History Guidance, a family of guidance methods uniquely enabled by DFoT. We show that its simplest form, vanilla history guidance, already significantly improves video generation quality and temporal consistency. A more advanced method, history guidance across time and frequency further enhances motion dynamics, enables compositional generalization to out-of-distribution history, and can stably roll out extremely long videos. Website: https://boyuan.space/history-guidance

RedStar: Does Scaling Long-CoT Data Unlock Better Slow-Reasoning Systems?

Can scaling transform reasoning? In this work, we explore the untapped potential of scaling Long Chain-of-Thought (Long-CoT) data to 1000k samples, pioneering the development of a slow-thinking model, RedStar. Through extensive experiments with various LLMs and different sizes, we uncover the ingredients for specialization and scale for Long-CoT training. Surprisingly, even smaller models show significant performance gains with limited data, revealing the sample efficiency of Long-CoT and the critical role of sample difficulty in the learning process. Our findings demonstrate that Long-CoT reasoning can be effectively triggered with just a few thousand examples, while larger models achieve unparalleled improvements. We also introduce reinforcement learning (RL)-scale training as a promising direction for advancing slow-thinking systems. RedStar shines across domains: on the MATH-Hard benchmark, RedStar-code-math boosts performance from 66.2\% to 81.6\%, and on the USA Math Olympiad (AIME), it solves 46.7\% of problems using only 21k mixed-code-math datasets. In multimodal tasks like GeoQA and MathVista-GEO, RedStar-Geo achieves competitive results with minimal Long-CoT data, outperforming other slow-thinking systems like QvQ-Preview. Compared to QwQ, RedStar strikes the perfect balance between reasoning and generalizability. Our work highlights that, with careful tuning, scaling Long-CoT can unlock extraordinary reasoning capabilities-even with limited dataset and set a new standard for slow-thinking models across diverse challenges. Our data and models are released at https://huggingface.co/RedStar-Reasoning.

A Paragraph is All It Takes: Rich Robot Behaviors from Interacting, Trusted LLMs

Large Language Models (LLMs) are compact representations of all public knowledge of our physical environment and animal and human behaviors. The application of LLMs to robotics may offer a path to highly capable robots that perform well across most human tasks with limited or even zero tuning. Aside from increasingly sophisticated reasoning and task planning, networks of (suitably designed) LLMs offer ease of upgrading capabilities and allow humans to directly observe the robot's thinking. Here we explore the advantages, limitations, and particularities of using LLMs to control physical robots. The basic system consists of four LLMs communicating via a human language data bus implemented via web sockets and ROS2 message passing. Surprisingly, rich robot behaviors and good performance across different tasks could be achieved despite the robot's data fusion cycle running at only 1Hz and the central data bus running at the extremely limited rates of the human brain, of around 40 bits/s. The use of natural language for inter-LLM communication allowed the robot's reasoning and decision making to be directly observed by humans and made it trivial to bias the system's behavior with sets of rules written in plain English. These rules were immutably written into Ethereum, a global, public, and censorship resistant Turing-complete computer. We suggest that by using natural language as the data bus among interacting AIs, and immutable public ledgers to store behavior constraints, it is possible to build robots that combine unexpectedly rich performance, upgradability, and durable alignment with humans.