Beyond Tokens: Semantic-Aware Speculative Decoding for Efficient Inference by Probing Internal States

Large Language Models (LLMs) achieve strong performance across many tasks but suffer from high inference latency due to autoregressive decoding. The issue is exacerbated in Large Reasoning Models (LRMs), which generate lengthy chains of thought. While speculative decoding accelerates inference by drafting and verifying multiple tokens in parallel, existing methods operate at the token level and ignore semantic equivalence (i.e., different token sequences expressing the same meaning), leading to inefficient rejections. We propose SemanticSpec, a semantic-aware speculative decoding framework that verifies entire semantic sequences instead of tokens. SemanticSpec introduces a semantic probability estimation mechanism that probes the model's internal hidden states to assess the likelihood of generating sequences with specific meanings.Experiments on four benchmarks show that SemanticSpec achieves up to 2.7x speedup on DeepSeekR1-32B and 2.1x on QwQ-32B, consistently outperforming token-level and sequence-level baselines in both efficiency and effectiveness.

Learning Legged MPC with Smooth Neural Surrogates

Deep learning and model predictive control (MPC) can play complementary roles in legged robotics. However, integrating learned models with online planning remains challenging. When dynamics are learned with neural networks, three key difficulties arise: (1) stiff transitions from contact events may be inherited from the data; (2) additional non-physical local nonsmoothness can occur; and (3) training datasets can induce non-Gaussian model errors due to rapid state changes. We address (1) and (2) by introducing the smooth neural surrogate, a neural network with tunable smoothness designed to provide informative predictions and derivatives for trajectory optimization through contact. To address (3), we train these models using a heavy-tailed likelihood that better matches the empirical error distributions observed in legged-robot dynamics. Together, these design choices substantially improve the reliability, scalability, and generalizability of learned legged MPC. Across zero-shot locomotion tasks of increasing difficulty, smooth neural surrogates with robust learning yield consistent reductions in cumulative cost on simple, well-conditioned behaviors (typically 10-50%), while providing substantially larger gains in regimes where standard neural dynamics often fail outright. In these regimes, smoothing enables reliable execution (from 0/5 to 5/5 success) and produces about 2-50x lower cumulative cost, reflecting orders-of-magnitude absolute improvements in robustness rather than incremental performance gains.

Large Video Planner Enables Generalizable Robot Control

General-purpose robots require decision-making models that generalize across diverse tasks and environments. Recent works build robot foundation models by extending multimodal large language models (MLLMs) with action outputs, creating vision-language-action (VLA) systems. These efforts are motivated by the intuition that MLLMs' large-scale language and image pretraining can be effectively transferred to the action output modality. In this work, we explore an alternative paradigm of using large-scale video pretraining as a primary modality for building robot foundation models. Unlike static images and language, videos capture spatio-temporal sequences of states and actions in the physical world that are naturally aligned with robotic behavior. We curate an internet-scale video dataset of human activities and task demonstrations, and train, for the first time at a foundation-model scale, an open video model for generative robotics planning. The model produces zero-shot video plans for novel scenes and tasks, which we post-process to extract executable robot actions. We evaluate task-level generalization through third-party selected tasks in the wild and real-robot experiments, demonstrating successful physical execution. Together, these results show robust instruction following, strong generalization, and real-world feasibility. We release both the model and dataset to support open, reproducible video-based robot learning. Our website is available at https://www.boyuan.space/large-video-planner/.

Time-Aware Policy Learning for Adaptive and Punctual Robot Control

Temporal awareness underlies intelligent behavior in both animals and humans, guiding how actions are sequenced, paced, and adapted to changing goals and environments. Yet most robot learning algorithms remain blind to time. We introduce time-aware policy learning, a reinforcement learning framework that enables robots to explicitly perceive and reason with time as a first-class variable. The framework augments conventional reinforcement policies with two complementary temporal signals, the remaining time and a time ratio, which allow a single policy to modulate its behavior continuously from rapid and dynamic to cautious and precise execution. By jointly optimizing punctuality and stability, the robot learns to balance efficiency, robustness, resiliency, and punctuality without re-training or reward adjustment. Across diverse manipulation domains from long-horizon pick and place, to granular-media pouring, articulated-object handling, and multi-agent object delivery, the time-aware policy produces adaptive behaviors that outperform standard reinforcement learning baselines by up to 48% in efficiency, 8 times more robust in sim-to-real transfer, and 90% in acoustic quietness while maintaining near-perfect success rates. Explicit temporal reasoning further enables real-time human-in-the-loop control and multi-agent coordination, allowing robots to recover from disturbances, re-synchronize after delays, and align motion tempo with human intent. By treating time not as a constraint but as a controllable dimension of behavior, time-aware policy learning provides a unified foundation for efficient, robust, resilient, and human-aligned robot autonomy.

TumorMap: A Laser-based Surgical Platform for 3D Tumor Mapping and Fully-Automated Tumor Resection

Surgical resection of malignant solid tumors is critically dependent on the surgeon's ability to accurately identify pathological tissue and remove the tumor while preserving surrounding healthy structures. However, building an intraoperative 3D tumor model for subsequent removal faces major challenges due to the lack of high-fidelity tumor reconstruction, difficulties in developing generalized tissue models to handle the inherent complexities of tumor diagnosis, and the natural physical limitations of bimanual operation, physiologic tremor, and fatigue creep during surgery. To overcome these challenges, we introduce "TumorMap", a surgical robotic platform to formulate intraoperative 3D tumor boundaries and achieve autonomous tissue resection using a set of multifunctional lasers. TumorMap integrates a three-laser mechanism (optical coherence tomography, laser-induced endogenous fluorescence, and cutting laser scalpel) combined with deep learning models to achieve fully-automated and noncontact tumor resection. We validated TumorMap in murine osteoscarcoma and soft-tissue sarcoma tumor models, and established a novel histopathological workflow to estimate sensor performance. With submillimeter laser resection accuracy, we demonstrated multimodal sensor-guided autonomous tumor surgery without any human intervention.

How Well do Diffusion Policies Learn Kinematic Constraint Manifolds?

Diffusion policies have shown impressive results in robot imitation learning, even for tasks that require satisfaction of kinematic equality constraints. However, task performance alone is not a reliable indicator of the policy's ability to precisely learn constraints in the training data. To investigate, we analyze how well diffusion policies discover these manifolds with a case study on a bimanual pick-and-place task that encourages fulfillment of a kinematic constraint for success. We study how three factors affect trained policies: dataset size, dataset quality, and manifold curvature. Our experiments show diffusion policies learn a coarse approximation of the constraint manifold with learning affected negatively by decreases in both dataset size and quality. On the other hand, the curvature of the constraint manifold showed inconclusive correlations with both constraint satisfaction and task success. A hardware evaluation verifies the applicability of our results in the real world. Project website with additional results and visuals: https://diffusion-learns-kinematic.github.io

Sym2Real: Symbolic Dynamics with Residual Learning for Data-Efficient Adaptive Control

We present Sym2Real, a fully data-driven framework that provides a principled way to train low-level adaptive controllers in a highly data-efficient manner. Using only about 10 trajectories, we achieve robust control of both a quadrotor and a racecar in the real world, without expert knowledge or simulation tuning. Our approach achieves this data efficiency by bringing symbolic regression to real-world robotics while addressing key challenges that prevent its direct application, including noise sensitivity and model degradation that lead to unsafe control. Our key observation is that the underlying physics is often shared for a system regardless of internal or external changes. Hence, we strategically combine low-fidelity simulation data with targeted real-world residual learning. Through experimental validation on quadrotor and racecar platforms, we demonstrate consistent data-efficient adaptation across six out-of-distribution sim2sim scenarios and successful sim2real transfer across five real-world conditions. More information and videos can be found at at http://generalroboticslab.com/Sym2Real

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 .

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.