Nan Jiang

Tengyang Xie, Dylan J. Foster, Yu Bai et al. · mit

Coverage conditions -- which assert that the data logging distribution adequately covers the state space -- play a fundamental role in determining the sample complexity of offline reinforcement learning. While such conditions might seem irrelevant to online reinforcement learning at first glance, we establish a new connection by showing -- somewhat surprisingly -- that the mere existence of a data distribution with good coverage can enable sample-efficient online RL. Concretely, we show that coverability -- that is, existence of a data distribution that satisfies a ubiquitous coverage condition called concentrability -- can be viewed as a structural property of the underlying MDP, and can be exploited by standard algorithms for sample-efficient exploration, even when the agent does not know said distribution. We complement this result by proving that several weaker notions of coverage, despite being sufficient for offline RL, are insufficient for online RL. We also show that existing complexity measures for online RL, including Bellman rank and Bellman-Eluder dimension, fail to optimally capture coverability, and propose a new complexity measure, the sequential extrapolation coefficient, to provide a unification.

Yong Lin, Hangyu Lin, Wei Xiong et al.

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax. To investigate alignment tax, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. Whereas, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between alignment performance and forgetting mitigation, leading to an alignment-forgetting trade-off. In this paper we show that model averaging, which simply interpolates between pre and post RLHF model weights, surprisingly achieves the most strongest alignment-forgetting Pareto front among a wide range of competing methods. To understand its effectiveness, we offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different alignment-forgetting trade-offs, we propose Heterogeneous Model Averaging (HMA) to Heterogeneously find various combination ratios of model layers. HMA seeks to maximize the alignment performance while incurring minimal alignment tax. Moreover, we validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B which is evaluated by open-sourced preference model and GPT4. Code available here: https://github.com/avalonstrel/Mitigating-the-Alignment-Tax-of-RLHF.git.

Tengyang Xie, Akanksha Saran, Dylan J. Foster et al. · mit

Consider the problem setting of Interaction-Grounded Learning (IGL), in which a learner's goal is to optimally interact with the environment with no explicit reward to ground its policies. The agent observes a context vector, takes an action, and receives a feedback vector, using this information to effectively optimize a policy with respect to a latent reward function. Prior analyzed approaches fail when the feedback vector contains the action, which significantly limits IGL's success in many potential scenarios such as Brain-computer interface (BCI) or Human-computer interface (HCI) applications. We address this by creating an algorithm and analysis which allows IGL to work even when the feedback vector contains the action, encoded in any fashion. We provide theoretical guarantees and large-scale experiments based on supervised datasets to demonstrate the effectiveness of the new approach.

Philip Amortila, Nan Jiang, Csaba Szepesvári · deepmind

Theoretical guarantees in reinforcement learning (RL) are known to suffer multiplicative blow-up factors with respect to the misspecification error of function approximation. Yet, the nature of such \emph{approximation factors} -- especially their optimal form in a given learning problem -- is poorly understood. In this paper we study this question in linear off-policy value function estimation, where many open questions remain. We study the approximation factor in a broad spectrum of settings, such as with the weighted $L_2$-norm (where the weighting is the offline state distribution), the $L_\infty$ norm, the presence vs. absence of state aliasing, and full vs. partial coverage of the state space. We establish the optimal asymptotic approximation factors (up to constants) for all of these settings. In particular, our bounds identify two instance-dependent factors for the $L_2(μ)$ norm and only one for the $L_\infty$ norm, which are shown to dictate the hardness of off-policy evaluation under misspecification.

Masatoshi Uehara, Haruka Kiyohara, Andrew Bennett et al. · harvard

We study off-policy evaluation (OPE) for partially observable MDPs (POMDPs) with general function approximation. Existing methods such as sequential importance sampling estimators and fitted-Q evaluation suffer from the curse of horizon in POMDPs. To circumvent this problem, we develop a novel model-free OPE method by introducing future-dependent value functions that take future proxies as inputs. Future-dependent value functions play similar roles as classical value functions in fully-observable MDPs. We derive a new Bellman equation for future-dependent value functions as conditional moment equations that use history proxies as instrumental variables. We further propose a minimax learning method to learn future-dependent value functions using the new Bellman equation. We obtain the PAC result, which implies our OPE estimator is consistent as long as futures and histories contain sufficient information about latent states, and the Bellman completeness. Finally, we extend our methods to learning of dynamics and establish the connection between our approach and the well-known spectral learning methods in POMDPs.

Nan Jiang, Tengyu Liu, Zhexuan Cao et al. · pku

Fine-grained capturing of 3D HOI boosts human activity understanding and facilitates downstream visual tasks, including action recognition, holistic scene reconstruction, and human motion synthesis. Despite its significance, existing works mostly assume that humans interact with rigid objects using only a few body parts, limiting their scope. In this paper, we address the challenging problem of f-AHOI, wherein the whole human bodies interact with articulated objects, whose parts are connected by movable joints. We present CHAIRS, a large-scale motion-captured f-AHOI dataset, consisting of 16.2 hours of versatile interactions between 46 participants and 81 articulated and rigid sittable objects. CHAIRS provides 3D meshes of both humans and articulated objects during the entire interactive process, as well as realistic and physically plausible full-body interactions. We show the value of CHAIRS with object pose estimation. By learning the geometrical relationships in HOI, we devise the very first model that leverage human pose estimation to tackle the estimation of articulated object poses and shapes during whole-body interactions. Given an image and an estimated human pose, our model first reconstructs the pose and shape of the object, then optimizes the reconstruction according to a learned interaction prior. Under both evaluation settings (e.g., with or without the knowledge of objects' geometries/structures), our model significantly outperforms baselines. We hope CHAIRS will promote the community towards finer-grained interaction understanding. We will make the data/code publicly available.

Nan Jiang, Thibaud Lutellier, Yiling Lou et al.

Automated Program Repair (APR) improves software reliability by generating patches for a buggy program automatically. Recent APR techniques leverage deep learning (DL) to build models to learn to generate patches from existing patches and code corpora. While promising, DL-based APR techniques suffer from the abundant syntactically or semantically incorrect patches in the patch space. These patches often disobey the syntactic and semantic domain knowledge of source code and thus cannot be the correct patches to fix a bug. We propose a DL-based APR approach KNOD, which incorporates domain knowledge to guide patch generation in a direct and comprehensive way. KNOD has two major novelties, including (1) a novel three-stage tree decoder, which directly generates Abstract Syntax Trees of patched code according to the inherent tree structure, and (2) a novel domain-rule distillation, which leverages syntactic and semantic rules and teacher-student distributions to explicitly inject the domain knowledge into the decoding procedure during both the training and inference phases. We evaluate KNOD on three widely-used benchmarks. KNOD fixes 72 bugs on the Defects4J v1.2, 25 bugs on the QuixBugs, and 50 bugs on the additional Defects4J v2.0 benchmarks, outperforming all existing APR tools.

Philip Amortila, Nan Jiang, Dhruv Madeka et al. · amazon-science

The current paper studies sample-efficient Reinforcement Learning (RL) in settings where only the optimal value function is assumed to be linearly-realizable. It has recently been understood that, even under this seemingly strong assumption and access to a generative model, worst-case sample complexities can be prohibitively (i.e., exponentially) large. We investigate the setting where the learner additionally has access to interactive demonstrations from an expert policy, and we present a statistically and computationally efficient algorithm (Delphi) for blending exploration with expert queries. In particular, Delphi requires $\tilde{\mathcal{O}}(d)$ expert queries and a $\texttt{poly}(d,H,|\mathcal{A}|,1/\varepsilon)$ amount of exploratory samples to provably recover an $\varepsilon$-suboptimal policy. Compared to pure RL approaches, this corresponds to an exponential improvement in sample complexity with surprisingly-little expert input. Compared to prior imitation learning (IL) approaches, our required number of expert demonstrations is independent of $H$ and logarithmic in $1/\varepsilon$, whereas all prior work required at least linear factors of both in addition to the same dependence on $d$. Towards establishing the minimal amount of expert queries needed, we show that, in the same setting, any learner whose exploration budget is polynomially-bounded (in terms of $d,H,$ and $|\mathcal{A}|$) will require at least $\tildeΩ(\sqrt{d})$ oracle calls to recover a policy competing with the expert's value function. Under the weaker assumption that the expert's policy is linear, we show that the lower bound increases to $\tildeΩ(d)$.

Jinglin Chen, Aditya Modi, Akshay Krishnamurthy et al.

We study reward-free reinforcement learning (RL) under general non-linear function approximation, and establish sample efficiency and hardness results under various standard structural assumptions. On the positive side, we propose the RFOLIVE (Reward-Free OLIVE) algorithm for sample-efficient reward-free exploration under minimal structural assumptions, which covers the previously studied settings of linear MDPs (Jin et al., 2020b), linear completeness (Zanette et al., 2020b) and low-rank MDPs with unknown representation (Modi et al., 2021). Our analyses indicate that the explorability or reachability assumptions, previously made for the latter two settings, are not necessary statistically for reward-free exploration. On the negative side, we provide a statistical hardness result for both reward-free and reward-aware exploration under linear completeness assumptions when the underlying features are unknown, showing an exponential separation between low-rank and linear completeness settings.

Ismail Hossain, Sai Puppala, Zhuoran Lu et al.

Open agent platforms allow community contributors to publish reusable skills that agents can invoke at runtime. This extensibility also creates a supply-chain risk: malicious contributors can hide harmful behavior inside skills that appear benign under superficial inspection. However, existing defenses are hard to evaluate because there is no benchmark that measures both malicious-skill detection and runtime verification. We present SkillVetBench, a two-stage security vetting benchmark for open agentic skill ecosystems. The first stage performs semantic vetting over each skill's natural-language specification to detect hidden malicious intent. The second stage executes flagged skills in an instrumented sandbox to observe runtime behavior and collect auditable evidence. We build a benchmark from confirmed malicious skills in the live OpenClaw ecosystem, including samples from the recent ClawHavoc supplychain campaign. Unlike static-only methods, SkillVetBench verifies detected threats with execution traces. Our experiments show that: (1) semantic-only and signature-based baselines are insufficient, missing up to 89\% of malicious skills whose threats arise from natural-language instructions, multicomponent logic, or cross-component interactions; (2) runtime attacks are concentrated in a small set of high-permission primitives, especially exec, write\_file, install\_skill, and spawn; and (3) SkillVetBench provides case studies in which sandbox execution directly supports malicious verdicts with concrete runtime evidence.

Mohak Bhardwaj, Tengyang Xie, Byron Boots et al.

We propose a novel model-based offline Reinforcement Learning (RL) framework, called Adversarial Model for Offline Reinforcement Learning (ARMOR), which can robustly learn policies to improve upon an arbitrary reference policy regardless of data coverage. ARMOR is designed to optimize policies for the worst-case performance relative to the reference policy through adversarially training a Markov decision process model. In theory, we prove that ARMOR, with a well-tuned hyperparameter, can compete with the best policy within data coverage when the reference policy is supported by the data. At the same time, ARMOR is robust to hyperparameter choices: the policy learned by ARMOR, with "any" admissible hyperparameter, would never degrade the performance of the reference policy, even when the reference policy is not covered by the dataset. To validate these properties in practice, we design a scalable implementation of ARMOR, which by adversarial training, can optimize policies without using model ensembles in contrast to typical model-based methods. We show that ARMOR achieves competent performance with both state-of-the-art offline model-free and model-based RL algorithms and can robustly improve the reference policy over various hyperparameter choices.

Jiawei Huang, Li Zhao, Tao Qin et al.

We propose a new learning framework that captures the tiered structure of many real-world user-interaction applications, where the users can be divided into two groups based on their different tolerance on exploration risks and should be treated separately. In this setting, we simultaneously maintain two policies $π^{\text{O}}$ and $π^{\text{E}}$: $π^{\text{O}}$ ("O" for "online") interacts with more risk-tolerant users from the first tier and minimizes regret by balancing exploration and exploitation as usual, while $π^{\text{E}}$ ("E" for "exploit") exclusively focuses on exploitation for risk-averse users from the second tier utilizing the data collected so far. An important question is whether such a separation yields advantages over the standard online setting (i.e., $π^{\text{E}}=π^{\text{O}}$) for the risk-averse users. We individually consider the gap-independent vs.~gap-dependent settings. For the former, we prove that the separation is indeed not beneficial from a minimax perspective. For the latter, we show that if choosing Pessimistic Value Iteration as the exploitation algorithm to produce $π^{\text{E}}$, we can achieve a constant regret for risk-averse users independent of the number of episodes $K$, which is in sharp contrast to the $Ω(\log K)$ regret for any online RL algorithms in the same setting, while the regret of $π^{\text{O}}$ (almost) maintains its online regret optimality and does not need to compromise for the success of $π^{\text{E}}$.

Emmanuel Oluwatosin Oluokun, Festus Fatai Adedoyin, Huseyin Dogan et al.

User Experience Research (UXR) in a legal and regulatory contexts presents unique challenges that require specialised approaches to protect vulnerable populations whilst generating actionable insights. Digital consultation, appointment booking, and medication delivery platforms show promise for extending care access; however, their real-world effectiveness is curtailed by an absence of theoretically grounded user experience research (UXR) methodologies that adequately account for the psychosocial conditions of these populations. This paper introduces a Generative AI-augmented UXR methodology, grounded in the UXR Point of View (PoV) Playbook, to guide the design of psychologically safe, low-cognitive-load digital health interventions for MSM and transgender individuals living with HIV/AIDS in Nigeria. Drawing from empirical research involving co-design workshops, thematic analysis, and requirements engineering, the methodology is operationalised through a four-stage UXR process encompassing AI-supported hypothesis generation, foundational planning, insight generation via Building Blocks, and the construction of stakeholder-specific PoV narratives. This process results in ten theory-informed UXR Play Cards that translate psychological mechanisms and empirical findings into actionable design guidance. Each play contains actionable tasks, AI-augmented approaches, and ethical guardrails tailored for research with marginalised populations. The output is a set of ten theory-informed UXR Play Cards translating psychological insight and empirical evidence into actionable design guidance. The core contribution is a replicable, stigma-aware, and privacy-centred framework for responsible GenAI use in UXR practice, advancing human-centred digital health design for marginalised communities.

Olumuyiwa Ayorinde, Huseyin Dogan, Festus Adedoyin et al.

This paper investigates how User Experience Research (UXR) methods can be combined with AI-supported analysis to develop clearer design direction for digital wellbeing interventions targeting Emergency and Public Safety Personnel (EPSP). EPSP work in high-stress, shift-based environments where cognitive fatigue and unpredictable schedules reduce engagement with conventional wellbeing tools. Using the UXR Point-of-View (PoV) framework, this study applied an AI-supported literature analysis process to identify recurring psychological, behavioural, and design patterns. Behaviour Change Techniques and Persuasive Technology principles were integrated throughout interpretation to connect evidence with practical design reasoning. The process resulted in a UXR PoV Pyramid, nine UXR Play Cards, and stakeholder focused PoV narratives. Findings show that effective wellbeing systems for EPSP must minimise cognitive effort, adapt to operational context, and prioritise psychological safety. The work demonstrates how AI can assist large-scale evidence interpretation while human researchers maintain responsibility for contextual judgement and design direction.

Jinglin Chen, Nan Jiang

We consider a challenging theoretical problem in offline reinforcement learning (RL): obtaining sample-efficiency guarantees with a dataset lacking sufficient coverage, under only realizability-type assumptions for the function approximators. While the existing theory has addressed learning under realizability and under non-exploratory data separately, no work has been able to address both simultaneously (except for a concurrent work which we compare in detail). Under an additional gap assumption, we provide guarantees to a simple pessimistic algorithm based on a version space formed by marginalized importance sampling (MIS), and the guarantee only requires the data to cover the optimal policy and the function classes to realize the optimal value and density-ratio functions. While similar gap assumptions have been used in other areas of RL theory, our work is the first to identify the utility and the novel mechanism of gap assumptions in offline RL with weak function approximation.

Yi Wu, Zikang Xiong, Yiran Hu et al.

Despite significant advancements in large language models (LLMs) that enhance robot agents' understanding and execution of natural language (NL) commands, ensuring the agents adhere to user-specified constraints remains challenging, particularly for complex commands and long-horizon tasks. To address this challenge, we present three key insights, equivalence voting, constrained decoding, and domain-specific fine-tuning, which significantly enhance LLM planners' capability in handling complex tasks. Equivalence voting ensures consistency by generating and sampling multiple Linear Temporal Logic (LTL) formulas from NL commands, grouping equivalent LTL formulas, and selecting the majority group of formulas as the final LTL formula. Constrained decoding then uses the generated LTL formula to enforce the autoregressive inference of plans, ensuring the generated plans conform to the LTL. Domain-specific fine-tuning customizes LLMs to produce safe and efficient plans within specific task domains. Our approach, Safe Efficient LLM Planner (SELP), combines these insights to create LLM planners to generate plans adhering to user commands with high confidence. We demonstrate the effectiveness and generalizability of SELP across different robot agents and tasks, including drone navigation and robot manipulation. For drone navigation tasks, SELP outperforms state-of-the-art planners by 10.8% in safety rate (i.e., finishing tasks conforming to NL commands) and by 19.8% in plan efficiency. For robot manipulation tasks, SELP achieves 20.4% improvement in safety rate. Our datasets for evaluating NL-to-LTL and robot task planning will be released in github.com/lt-asset/selp.

Audrey Huang, Jinglin Chen, Nan Jiang

MDPs with low-rank transitions -- that is, the transition matrix can be factored into the product of two matrices, left and right -- is a highly representative structure that enables tractable learning. The left matrix enables expressive function approximation for value-based learning and has been studied extensively. In this work, we instead investigate sample-efficient learning with density features, i.e., the right matrix, which induce powerful models for state-occupancy distributions. This setting not only sheds light on leveraging unsupervised learning in RL, but also enables plug-in solutions for convex RL. In the offline setting, we propose an algorithm for off-policy estimation of occupancies that can handle non-exploratory data. Using this as a subroutine, we further devise an online algorithm that constructs exploratory data distributions in a level-by-level manner. As a central technical challenge, the additive error of occupancy estimation is incompatible with the multiplicative definition of data coverage. In the absence of strong assumptions like reachability, this incompatibility easily leads to exponential error blow-up, which we overcome via novel technical tools. Our results also readily extend to the representation learning setting, when the density features are unknown and must be learned from an exponentially large candidate set.

Yuheng Zhang, Yu Bai, Nan Jiang

We study offline multi-agent reinforcement learning (RL) in Markov games, where the goal is to learn an approximate equilibrium -- such as Nash equilibrium and (Coarse) Correlated Equilibrium -- from an offline dataset pre-collected from the game. Existing works consider relatively restricted tabular or linear models and handle each equilibria separately. In this work, we provide the first framework for sample-efficient offline learning in Markov games under general function approximation, handling all 3 equilibria in a unified manner. By using Bellman-consistent pessimism, we obtain interval estimation for policies' returns, and use both the upper and the lower bounds to obtain a relaxation on the gap of a candidate policy, which becomes our optimization objective. Our results generalize prior works and provide several additional insights. Importantly, we require a data coverage condition that improves over the recently proposed "unilateral concentrability". Our condition allows selective coverage of deviation policies that optimally trade-off between their greediness (as approximate best responses) and coverage, and we show scenarios where this leads to significantly better guarantees. As a new connection, we also show how our algorithmic framework can subsume seemingly different solution concepts designed for the special case of two-player zero-sum games.

Tengyang Xie, Mohak Bhardwaj, Nan Jiang et al.

We propose a new model-based offline RL framework, called Adversarial Models for Offline Reinforcement Learning (ARMOR), which can robustly learn policies to improve upon an arbitrary baseline policy regardless of data coverage. Based on the concept of relative pessimism, ARMOR is designed to optimize for the worst-case relative performance when facing uncertainty. In theory, we prove that the learned policy of ARMOR never degrades the performance of the baseline policy with any admissible hyperparameter, and can learn to compete with the best policy within data coverage when the hyperparameter is well tuned, and the baseline policy is supported by the data. Such a robust policy improvement property makes ARMOR especially suitable for building real-world learning systems, because in practice ensuring no performance degradation is imperative before considering any benefit learning can bring.

Audrey Huang, Nan Jiang

Off-policy evaluation often refers to two related tasks: estimating the expected return of a policy and estimating its value function (or other functions of interest, such as density ratios). While recent works on marginalized importance sampling (MIS) show that the former can enjoy provable guarantees under realizable function approximation, the latter is only known to be feasible under much stronger assumptions such as prohibitively expressive discriminators. In this work, we provide guarantees for off-policy function estimation under only realizability, by imposing proper regularization on the MIS objectives. Compared to commonly used regularization in MIS, our regularizer is much more flexible and can account for an arbitrary user-specified distribution, under which the learned function will be close to the groundtruth. We provide exact characterization of the optimal dual solution that needs to be realized by the discriminator class, which determines the data-coverage assumption in the case of value-function learning. As another surprising observation, the regularizer can be altered to relax the data-coverage requirement, and completely eliminate it in the ideal case with strong side information.

Yixin Chen, Junfeng Ni, Nan Jiang et al.

Reconstructing detailed 3D scenes from single-view images remains a challenging task due to limitations in existing approaches, which primarily focus on geometric shape recovery, overlooking object appearances and fine shape details. To address these challenges, we propose a novel framework for simultaneous high-fidelity recovery of object shapes and textures from single-view images. Our approach utilizes the proposed Single-view neural implicit Shape and Radiance field (SSR) representations to leverage both explicit 3D shape supervision and volume rendering of color, depth, and surface normal images. To overcome shape-appearance ambiguity under partial observations, we introduce a two-stage learning curriculum incorporating both 3D and 2D supervisions. A distinctive feature of our framework is its ability to generate fine-grained textured meshes while seamlessly integrating rendering capabilities into the single-view 3D reconstruction model. This integration enables not only improved textured 3D object reconstruction by 27.7% and 11.6% on the 3D-FRONT and Pix3D datasets, respectively, but also supports the rendering of images from novel viewpoints. Beyond individual objects, our approach facilitates composing object-level representations into flexible scene representations, thereby enabling applications such as holistic scene understanding and 3D scene editing. We conduct extensive experiments to demonstrate the effectiveness of our method.

Yuanheng Zhang, Nan Jiang, Zhaoheng Xie et al.

Accurately annotated ultrasonic images are vital components of a high-quality medical report. Hospitals often have strict guidelines on the types of annotations that should appear on imaging results. However, manually inspecting these images can be a cumbersome task. While a neural network could potentially automate the process, training such a model typically requires a dataset of paired input and target images, which in turn involves significant human labour. This study introduces an automated approach for detecting annotations in images. This is achieved by treating the annotations as noise, creating a self-supervised pretext task and using a model trained under the Noise2Noise scheme to restore the image to a clean state. We tested a variety of model structures on the denoising task against different types of annotation, including body marker annotation, radial line annotation, etc. Our results demonstrate that most models trained under the Noise2Noise scheme outperformed their counterparts trained with noisy-clean data pairs. The costumed U-Net yielded the most optimal outcome on the body marker annotation dataset, with high scores on segmentation precision and reconstruction similarity. We released our code at https://github.com/GrandArth/UltrasonicImage-N2N-Approach.

Nan Jiang, Zimo He, Wanhe Yu et al. · pku

A long-standing objective in humanoid robotics is the realization of versatile agents capable of following diverse multimodal instructions with human-level flexibility. Despite advances in humanoid control, bridging high-level multimodal perception with whole-body execution remains a significant bottleneck. Existing methods often struggle to translate heterogeneous instructions -- such as language, music, and trajectories -- into stable, real-time actions. Here we show that UniAct, a two-stage framework integrating a fine-tuned MLLM with a causal streaming pipeline, enables humanoid robots to execute multimodal instructions with sub-500 ms latency. By unifying inputs through a shared discrete codebook via FSQ, UniAct ensures cross-modal alignment while constraining motions to a physically grounded manifold. This approach yields a 19% improvement in the success rate of zero-shot tracking of imperfect reference motions. We validate UniAct on UniMoCap, our 20-hour humanoid motion benchmark, demonstrating robust generalization across diverse real-world scenarios. Our results mark a critical step toward responsive, general-purpose humanoid assistants capable of seamless interaction through unified perception and control.

Hanze Dong, Wei Xiong, Bo Pang et al.

We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF) in this technical report, which is widely reported to outperform its offline counterpart by a large margin in the recent large language model (LLM) literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this technical report, we aim to fill in this gap and provide a detailed recipe that is easy to reproduce for online iterative RLHF. In particular, since online human feedback is usually infeasible for open-source communities with limited resources, we start by constructing preference models using a diverse set of open-source datasets and use the constructed proxy preference model to approximate human feedback. Then, we discuss the theoretical insights and algorithmic principles behind online iterative RLHF, followed by a detailed practical implementation. Our trained LLM achieves impressive performance on LLM chatbot benchmarks, including AlpacaEval-2, Arena-Hard, and MT-Bench, as well as other academic benchmarks such as HumanEval and TruthfulQA. We have shown that supervised fine-tuning (SFT) and iterative RLHF can obtain state-of-the-art performance with fully open-source datasets. Further, we have made our models, curated datasets, and comprehensive step-by-step code guidebooks publicly available. Please refer to https://github.com/RLHFlow/RLHF-Reward-Modeling and https://github.com/RLHFlow/Online-RLHF for more detailed information.

Nan Jiang, Chengxiao Wang, Kevin Liu et al.

Binary code analysis is the foundation of crucial tasks in the security domain; thus building effective binary analysis techniques is more important than ever. Large language models (LLMs) although have brought impressive improvement to source code tasks, do not directly generalize to assembly code due to the unique challenges of assembly: (1) the low information density of assembly and (2) the diverse optimizations in assembly code. To overcome these challenges, this work proposes a hierarchical attention mechanism that builds attention summaries to capture the semantics more effectively and designs contrastive learning objectives to train LLMs to learn assembly optimization. Equipped with these techniques, this work develops Nova, a generative LLM for assembly code. Nova outperforms existing techniques on binary code decompilation by up to 14.84 -- 21.58% (absolute percentage point improvement) higher Pass@1 and Pass@10, and outperforms the latest binary code similarity detection techniques by up to 6.17% Recall@1, showing promising abilities on both assembly generation and understanding tasks.

Yanbing Zhang, Bo Wang, Jianhui Liu et al.

Current Large Multimodal Models (LMMs) struggle with spatial reasoning tasks requiring viewpoint-dependent understanding, largely because they are confined to a single, static observation. We propose Thinking with Novel Views (TwNV), a paradigm that integrates generative novel-view synthesis into the reasoning loop: a Reasoner LMM identifies spatial ambiguity, instructs a Painter to synthesize an alternative viewpoint, and re-examines the scene with the additional evidence. Through systematic experiments we address three research questions. (1) Instruction format: numerical camera-pose specifications yield more reliable view control than free-form language. (2) Generation fidelity: synthesized view quality is tightly coupled with downstream spatial accuracy. (3) Inference-time visual scaling: iterative multi-turn view refinement further improves performance, echoing recent scaling trends in language reasoning. Across four spatial subtask categories and four LMM architectures (both closed- and open-source), TwNV consistently improves accuracy by +1.3 to +3.9 pp, with the largest gains on viewpoint-sensitive subtasks. These results establish novel-view generation as a practical lever for advancing spatial intelligence of LMMs.

Pulkit Katdare, Nan Jiang, Katherine Driggs-Campbell

Reinforcement Learning (RL) methods are typically sample-inefficient, making it challenging to train and deploy RL-policies in real world robots. Even a robust policy trained in simulation requires a real-world deployment to assess their performance. This paper proposes a new approach to evaluate the real-world performance of agent policies prior to deploying them in the real world. Our approach incorporates a simulator along with real-world offline data to evaluate the performance of any policy using the framework of Marginalized Importance Sampling (MIS). Existing MIS methods face two challenges: (1) large density ratios that deviate from a reasonable range and (2) indirect supervision, where the ratio needs to be inferred indirectly, thus exacerbating estimation error. Our approach addresses these challenges by introducing the target policy's occupancy in the simulator as an intermediate variable and learning the density ratio as the product of two terms that can be learned separately. The first term is learned with direct supervision and the second term has a small magnitude, thus making it computationally efficient. We analyze the sample complexity as well as error propagation of our two step-procedure. Furthermore, we empirically evaluate our approach on Sim2Sim environments such as Cartpole, Reacher, and Half-Cheetah. Our results show that our method generalizes well across a variety of Sim2Sim gap, target policies and offline data collection policies. We also demonstrate the performance of our algorithm on a Sim2Real task of validating the performance of a 7 DoF robotic arm using offline data along with the Gazebo simulator.

Jie Yang, Xuesong Niu, Nan Jiang et al.

Existing 3D human object interaction (HOI) datasets and models simply align global descriptions with the long HOI sequence, while lacking a detailed understanding of intermediate states and the transitions between states. In this paper, we argue that fine-grained semantic alignment, which utilizes state-level descriptions, offers a promising paradigm for learning semantically rich HOI representations. To achieve this, we introduce Semantic-HOI, a new dataset comprising over 20K paired HOI states with fine-grained descriptions for each HOI state and the body movements that happen between two consecutive states. Leveraging the proposed dataset, we design three state-level HOI tasks to accomplish fine-grained semantic alignment within the HOI sequence. Additionally, we propose a unified model called F-HOI, designed to leverage multimodal instructions and empower the Multi-modal Large Language Model to efficiently handle diverse HOI tasks. F-HOI offers multiple advantages: (1) It employs a unified task formulation that supports the use of versatile multimodal inputs. (2) It maintains consistency in HOI across 2D, 3D, and linguistic spaces. (3) It utilizes fine-grained textual supervision for direct optimization, avoiding intricate modeling of HOI states. Extensive experiments reveal that F-HOI effectively aligns HOI states with fine-grained semantic descriptions, adeptly tackling understanding, reasoning, generation, and reconstruction tasks.

Lin Song, Wenbo Li, Guoqing Ma et al.

We present JoyAI-Image, a unified multimodal foundation model for visual understanding, text-to-image generation, and instruction-guided image editing. JoyAI-Image couples a spatially enhanced Multimodal Large Language Model (MLLM) with a Multimodal Diffusion Transformer (MMDiT), allowing perception and generation to interact through a shared multimodal interface. Around this architecture, we build a scalable training recipe that combines unified instruction tuning, long-text rendering supervision, spatially grounded data, and both general and spatial editing signals. This design gives the model broad multimodal capability while strengthening geometry-aware reasoning and controllable visual synthesis. Experiments across understanding, generation, long-text rendering, and editing benchmarks show that JoyAI-Image achieves state-of-the-art or highly competitive performance. More importantly, the bidirectional loop between enhanced understanding, controllable spatial editing, and novel-view-assisted reasoning enables the model to move beyond general visual competence toward stronger spatial intelligence. These results suggest a promising path for unified visual models in downstream applications such as vision-language-action systems and world models.

Nan Jiang, Yi Gu, Yexiang Xue

Learning to generate complex combinatorial structures satisfying constraints will have transformative impacts in many application domains. However, it is beyond the capabilities of existing approaches due to the highly intractable nature of the embedded probabilistic inference. Prior works spend most of the training time learning to separate valid from invalid structures but do not learn the inductive biases of valid structures. We develop NEural Lovász Sampler (Nelson), which embeds the sampler through Lovász Local Lemma (LLL) as a fully differentiable neural network layer. Our Nelson-CD embeds this sampler into the contrastive divergence learning process of Markov random fields. Nelson allows us to obtain valid samples from the current model distribution. Contrastive divergence is then applied to separate these samples from those in the training set. Nelson is implemented as a fully differentiable neural net, taking advantage of the parallelism of GPUs. Experimental results on several real-world domains reveal that Nelson learns to generate 100\% valid structures, while baselines either time out or cannot ensure validity. Nelson also outperforms other approaches in running time, log-likelihood, and MAP scores.

Qitian Jason Hu, Jacob Bieker, Xiuyu Li et al.

As the range of applications for Large Language Models (LLMs) continues to grow, the demand for effective serving solutions becomes increasingly critical. Despite the versatility of LLMs, no single model can optimally address all tasks and applications, particularly when balancing performance with cost. This limitation has led to the development of LLM routing systems, which combine the strengths of various models to overcome the constraints of individual LLMs. Yet, the absence of a standardized benchmark for evaluating the performance of LLM routers hinders progress in this area. To bridge this gap, we present RouterBench, a novel evaluation framework designed to systematically assess the efficacy of LLM routing systems, along with a comprehensive dataset comprising over 405k inference outcomes from representative LLMs to support the development of routing strategies. We further propose a theoretical framework for LLM routing, and deliver a comparative analysis of various routing approaches through RouterBench, highlighting their potentials and limitations within our evaluation framework. This work not only formalizes and advances the development of LLM routing systems but also sets a standard for their assessment, paving the way for more accessible and economically viable LLM deployments. The code and data are available at https://github.com/withmartian/routerbench.

Yuheng Zhang, Chenlu Ye, Shuowei Jin et al.

Reinforcement learning, including reinforcement learning with verifiable rewards (RLVR), has emerged as a powerful approach for LLM post-training. Central to these approaches is the design of the importance sampling (IS) ratio used in off-policy policy-gradient estimation. Existing methods face a fundamental bias-variance dilemma: token-level IS ratios, as adopted by PPO (Schulman et al., 2017) and GRPO (Shao et al., 2024), introduce bias by ignoring prefix state distribution mismatch; full sequence ratios provide exact trajectory-level correction but suffer from high variance due to the multiplicative accumulation of per-token ratios, while GSPO (Zheng et al., 2025) improves numerical stability via length normalization at the cost of deviating from the exact full-sequence IS correction. In this work, we identify the cumulative token IS ratio, the product of per-token ratios up to position $t$, as a theoretically principled solution to this dilemma. We prove that, under the token-level policy-gradient formulation, this ratio provides an unbiased prefix correction for each token-level gradient term and has strictly lower variance than the full sequence ratio. Building on this insight, we propose CTPO (Cumulative Token Policy Optimization), which combines the cumulative token IS ratio with position-adaptive clipping that scales log-space clip bounds according to the natural $\sqrt{t}$ growth of the cumulative log-ratio. This yields more consistent regularization across token positions. We implement and evaluate CTPO in the tool-integrated reasoning setting on several challenging mathematical reasoning benchmarks, achieving the best average performance across both model scales compared with strong GRPO and GSPO baselines. Code will be available at https://github.com/horizon-llm/CTPO.

Minxiao Chen, Haitao Yuan, Nan Jiang et al.

Traffic accidents pose a significant risk to human health and property safety. Therefore, to prevent traffic accidents, predicting their risks has garnered growing interest. We argue that a desired prediction solution should demonstrate resilience to the complexity of traffic accidents. In particular, it should adequately consider the regional background, accurately capture both spatial proximity and semantic similarity, and effectively address the sparsity of traffic accidents. However, these factors are often overlooked or difficult to incorporate. In this paper, we propose a novel multi-granularity hierarchical spatio-temporal network. Initially, we innovate by incorporating remote sensing data, facilitating the creation of hierarchical multi-granularity structure and the comprehension of regional background. We construct multiple high-level risk prediction tasks to enhance model's ability to cope with sparsity. Subsequently, to capture both spatial proximity and semantic similarity, region feature and multi-view graph undergo encoding processes to distill effective representations. Additionally, we propose message passing and adaptive temporal attention module that bridges different granularities and dynamically captures time correlations inherent in traffic accident patterns. At last, a multivariate hierarchical loss function is devised considering the complexity of the prediction purpose. Extensive experiments on two real datasets verify the superiority of our model against the state-of-the-art methods.

Adnan Aijaz, Nan Jiang, Aftab Khan

Time-critical control applications typically pose stringent connectivity requirements for communication networks. The imperfections associated with the wireless medium such as packet losses, synchronization errors, and varying delays have a detrimental effect on performance of real-time control, often with safety implications. This paper introduces multi-service edge-intelligence as a new paradigm for realizing time-critical control over wireless. It presents the concept of multi-service edge-intelligence which revolves around tight integration of wireless access, edge-computing and machine learning techniques, in order to provide stability guarantees under wireless imperfections. The paper articulates some of the key system design aspects of multi-service edge-intelligence. It also presents a temporal-adaptive prediction technique to cope with dynamically changing wireless environments. It provides performance results in a robotic teleoperation scenario. Finally, it discusses some open research and design challenges for multi-service edge-intelligence.

Haoran Li, Jiayu Lv, Congying Han et al.

Achieving optimality and adversarial robustness in deep reinforcement learning has long been regarded as conflicting goals. Nonetheless, recent theoretical insights presented in CAR suggest a potential alignment, raising the important question of how to realize this in practice. This paper first identifies a key gap between theory and practice by comparing standard policy optimization (SPO) and adversarially robust policy optimization (ARPO). Although they share theoretical consistency, a fundamental tension between robustness and optimality arises in practical policy gradient methods. SPO tends toward convergence to vulnerable first-order stationary policies (FOSPs) with strong natural performance, whereas ARPO typically favors more robust FOSPs at the expense of reduced returns. Furthermore, we attribute this tradeoff to the reshaping effect of the strongest adversary in ARPO, which significantly complicates the global landscape by inducing deceptive sticky FOSPs. This improves robustness but makes navigation more challenging. To alleviate this, we develop the BARPO, a bilevel framework unifying SPO and ARPO by modulating adversary strength, thereby facilitating navigability while preserving global optima. Extensive empirical results demonstrate that BARPO consistently outperforms vanilla ARPO, providing a practical approach to reconcile theoretical and empirical performance.

Nan Jiang, Ziyi Wang, Yexiang Xue

Symbolic regression seeks to uncover physical laws from experimental data by searching for closed-form expressions, which is an important task in AI-driven scientific discovery. Yet the exponential growth of the search space of expression renders the task computationally challenging. A promising yet underexplored direction for reducing the effective search space and accelerating training lies in symbolic equivalence: many expressions, although syntactically different, define the same function -- for example, $\log(x_1^2x_2^3)$, $\log(x_1^2)+\log(x_2^3)$, and $2\log(x_1)+3\log(x_2)$. Existing algorithms treat such variants as distinct outputs, leading to redundant exploration and slow learning. We introduce EGG-SR, a unified framework that integrates equality graphs (e-graphs) into diverse symbolic regression algorithms, including Monte Carlo Tree Search (MCTS), deep reinforcement learning (DRL), and large language models (LLMs). EGG-SR compactly represents equivalent expressions through the proposed EGG module, enabling more efficient learning by: (1) pruning redundant subtree exploration in EGG-MCTS, (2) aggregating rewards across equivalence classes in EGG-DRL, and (3) enriching feedback prompts in EGG-LLM. Under mild assumptions, we show that embedding e-graphs tightens the regret bound of MCTS and reduces the variance of the DRL gradient estimator. Empirically, EGG-SR consistently enhances multiple baselines across challenging benchmarks, discovering equations with lower normalized mean squared error than state-of-the-art methods. Code implementation is available at: https://www.github.com/jiangnanhugo/egg-sr.

Ivan Lee, Nan Jiang, Taylor Berg-Kirkpatrick

What is the relationship between model architecture and the ability to perform in-context learning? In this empirical study, we take the first steps toward answering this question. We evaluate thirteen model architectures capable of causal language modeling across a suite of synthetic in-context learning tasks. These selected architectures represent a broad range of paradigms, including recurrent and convolution-based neural networks, transformers, state space model inspired, and other emerging attention alternatives. We discover that all the considered architectures can perform in-context learning under a wider range of conditions than previously documented. Additionally, we observe stark differences in statistical efficiency and consistency by varying the number of in-context examples and task difficulty. We also measure each architecture's predisposition towards in-context learning when presented with the option to memorize rather than leverage in-context examples. Finally, and somewhat surprisingly, we find that several attention alternatives are sometimes competitive with or better in-context learners than transformers. However, no single architecture demonstrates consistency across all tasks, with performance either plateauing or declining when confronted with a significantly larger number of in-context examples than those encountered during gradient-based training.

Pengfei Cai, Yan Song, Nan Jiang et al.

A significant challenge in sound event detection (SED) is the effective utilization of unlabeled data, given the limited availability of labeled data due to high annotation costs. Semi-supervised algorithms rely on labeled data to learn from unlabeled data, and the performance is constrained by the quality and size of the former. In this paper, we introduce the Prototype based Masked Audio Model~(PMAM) algorithm for self-supervised representation learning in SED, to better exploit unlabeled data. Specifically, semantically rich frame-level pseudo labels are constructed from a Gaussian mixture model (GMM) based prototypical distribution modeling. These pseudo labels supervise the learning of a Transformer-based masked audio model, in which binary cross-entropy loss is employed instead of the widely used InfoNCE loss, to provide independent loss contributions from different prototypes, which is important in real scenarios in which multiple labels may apply to unsupervised data frames. A final stage of fine-tuning with just a small amount of labeled data yields a very high performing SED model. On like-for-like tests using the DESED task, our method achieves a PSDS1 score of 62.5\%, surpassing current state-of-the-art models and demonstrating the superiority of the proposed technique.

Junhong Shen, Hao Bai, Lunjun Zhang et al. · cmu

The current paradigm of test-time scaling relies on generating long reasoning traces ("thinking" more) before producing a response. In agent problems that require interaction, this can be done by generating thinking traces before acting in the world. However, this process does not allow agents to acquire new information from the environment or adapt their behavior over time. In this work, we propose to scale test-time interaction, an untapped dimension of test-time scaling that increases the agent's interaction horizon to enable running rich behaviors such as exploration, backtracking, and dynamic re-planning within a single rollout. To demonstrate the promise of this scaling dimension, we study the domain of web agents. We first show that even prompting-based interaction scaling without any training can improve task success on web benchmarks non-trivially. Building on this, we introduce TTI (Test-Time Interaction), a curriculum-based online reinforcement learning (RL) approach that trains agents by adaptively adjusting their rollout lengths. Using a Gemma 3 12B model, TTI produces state-of-the-art open-source, open-data web agents on WebVoyager and WebArena benchmarks. We further show that TTI enables agents to balance exploration and exploitation adaptively. Our results establish interaction scaling as a powerful, complementary axis to scaling per-step compute, offering new avenues for training adaptive agents.

Nan Jiang, Qi Li, Lin Tan et al.

Despite their success, large language models (LLMs) face the critical challenge of hallucinations, generating plausible but incorrect content. While much research has focused on hallucinations in multiple modalities including images and natural language text, less attention has been given to hallucinations in source code, which leads to incorrect and vulnerable code that causes significant financial loss. To pave the way for research in LLMs' hallucinations in code, we introduce Collu-Bench, a benchmark for predicting code hallucinations of LLMs across code generation (CG) and automated program repair (APR) tasks. Collu-Bench includes 13,234 code hallucination instances collected from five datasets and 11 diverse LLMs, ranging from open-source models to commercial ones. To better understand and predict code hallucinations, Collu-Bench provides detailed features such as the per-step log probabilities of LLMs' output, token types, and the execution feedback of LLMs' generated code for in-depth analysis. In addition, we conduct experiments to predict hallucination on Collu-Bench, using both traditional machine learning techniques and neural networks, which achieves 22.03 -- 33.15% accuracy. Our experiments draw insightful findings of code hallucination patterns, reveal the challenge of accurately localizing LLMs' hallucinations, and highlight the need for more sophisticated techniques.

Nan Jiang, Shanchao Liang, Chengxiao Wang et al.

Portable Document Format (PDF) files are dominantly used for storing and disseminating scientific research, legal documents, and tax information. LaTeX is a popular application for creating PDF documents. Despite its advantages, LaTeX is not WYSWYG -- what you see is what you get, i.e., the LaTeX source and rendered PDF images look drastically different, especially for formulae and tables. This gap makes it hard to modify or export LaTeX sources for formulae and tables from PDF images, and existing work is still limited. First, prior work generates LaTeX sources in a single iteration and struggles with complex LaTeX formulae. Second, existing work mainly recognizes and extracts LaTeX sources for formulae; and is incapable or ineffective for tables. This paper proposes LATTE, the first iterative refinement framework for LaTeX recognition. Specifically, we propose delta-view as feedback, which compares and pinpoints the differences between a pair of rendered images of the extracted LaTeX source and the expected correct image. Such delta-view feedback enables our fault localization model to localize the faulty parts of the incorrect recognition more accurately and enables our LaTeX refinement model to repair the incorrect extraction more accurately. LATTE improves the LaTeX source extraction accuracy of both LaTeX formulae and tables, outperforming existing techniques as well as GPT-4V by at least 7.03% of exact match, with a success refinement rate of 46.08% (formula) and 25.51% (table).

Yihao Chen, Haochen Wu, Nan Jiang et al.

This paper describes the USTC-KXDIGIT system submitted to the ASVspoof5 Challenge for Track 1 (speech deepfake detection) and Track 2 (spoofing-robust automatic speaker verification, SASV). Track 1 showcases a diverse range of technical qualities from potential processing algorithms and includes both open and closed conditions. For these conditions, our system consists of a cascade of a frontend feature extractor and a back-end classifier. We focus on extensive embedding engineering and enhancing the generalization of the back-end classifier model. Specifically, the embedding engineering is based on hand-crafted features and speech representations from a self-supervised model, used for closed and open conditions, respectively. To detect spoof attacks under various adversarial conditions, we trained multiple systems on an augmented training set. Additionally, we used voice conversion technology to synthesize fake audio from genuine audio in the training set to enrich the synthesis algorithms. To leverage the complementary information learned by different model architectures, we employed activation ensemble and fused scores from different systems to obtain the final decision score for spoof detection. During the evaluation phase, the proposed methods achieved 0.3948 minDCF and 14.33% EER in the close condition, and 0.0750 minDCF and 2.59% EER in the open condition, demonstrating the robustness of our submitted systems under adversarial conditions. In Track 2, we continued using the CM system from Track 1 and fused it with a CNN-based ASV system. This approach achieved 0.2814 min-aDCF in the closed condition and 0.0756 min-aDCF in the open condition, showcasing superior performance in the SASV system.

Yanfeng Shi, Pengfei Cai, Jun Liu et al.

Large Audio-Language Models (LALMs) enable general audio understanding and demonstrate remarkable performance across various audio tasks. However, these models still face challenges in temporal perception (e.g., inferring event onset and offset), leading to limited utility in fine-grained scenarios. To address this issue, we propose Audio-Side Time Prompt and leverage Reinforcement Learning (RL) to develop the TimePro-RL framework for fine-grained temporal perception. Specifically, we encode timestamps as embeddings and interleave them within the audio feature sequence as temporal coordinates to prompt the model. Furthermore, we introduce RL following Supervised Fine-Tuning (SFT) to directly optimize temporal alignment performance. Experiments demonstrate that TimePro-RL achieves significant performance gains across a range of audio temporal tasks, such as audio grounding, sound event detection, and dense audio captioning, validating its robust effectiveness.

Jiarui Yao, Yifan Hao, Hanning Zhang et al.

Chain-of-thought (CoT) reasoning in large language models (LLMs) can be formalized as a latent variable problem, where the model needs to generate intermediate reasoning steps. While prior approaches such as iterative reward-ranked fine-tuning (RAFT) have relied on such formulations, they typically apply uniform inference budgets across prompts, which fails to account for variability in difficulty and convergence behavior. This work identifies the main bottleneck in CoT training as inefficient stochastic gradient estimation due to static sampling strategies. We propose GVM-RAFT, a prompt-specific Dynamic Sample Allocation Strategy designed to minimize stochastic gradient variance under a computational budget constraint. The method dynamically allocates computational resources by monitoring prompt acceptance rates and stochastic gradient norms, ensuring that the resulting gradient variance is minimized. Our theoretical analysis shows that the proposed dynamic sampling strategy leads to accelerated convergence guarantees under suitable conditions. Experiments on mathematical reasoning show that GVM-RAFT achieves a 2-4x speedup and considerable accuracy improvements over vanilla RAFT. The proposed dynamic sampling strategy is general and can be incorporated into other reinforcement learning algorithms, such as GRPO, leading to similar improvements in convergence and test accuracy. Our code is available at https://github.com/RLHFlow/GVM.

Yicheng Xiao, Wenhu Zhang, Lin Song et al.

Image spatial editing performs geometry-driven transformations, allowing precise control over object layout and camera viewpoints. Current models are insufficient for fine-grained spatial manipulations, motivating a dedicated assessment suite. Our contributions are listed: (i) We introduce SpatialEdit-Bench, a complete benchmark that evaluates spatial editing by jointly measuring perceptual plausibility and geometric fidelity via viewpoint reconstruction and framing analysis. (ii) To address the data bottleneck for scalable training, we construct SpatialEdit-500k, a synthetic dataset generated with a controllable Blender pipeline that renders objects across diverse backgrounds and systematic camera trajectories, providing precise ground-truth transformations for both object- and camera-centric operations. (iii) Building on this data, we develop SpatialEdit-16B, a baseline model for fine-grained spatial editing. Our method achieves competitive performance on general editing while substantially outperforming prior methods on spatial manipulation tasks. All resources will be made public at https://github.com/EasonXiao-888/SpatialEdit.

Ziyi Wang, Siva Rajesh Kasa, Ankith M S et al.

Speculative decoding is an effective technique for accelerating large language model inference by drafting multiple tokens in parallel. In practice, its speedup is often bottlenecked by a rigid verification step that strictly enforces the accepted token distribution to exactly match the target model. This constraint leads to the rejection of many plausible tokens, lowering the acceptance rate and limiting overall time speedup. To overcome this limitation, we propose Dynamic Verification Relaxed Speculative Decoding (DIVERSED), a relaxed verification framework that improves time efficiency while preserving generation quality. DIVERSED learns an ensemble-based verifier that blends the draft and target model distributions with a task-dependent and context-dependent weight. We provide theoretical justification for our approach and demonstrate empirically that DIVERSED achieves substantially higher inference efficiency compared to standard speculative decoding methods. Code is available at: https://github.com/comeusr/diversed.

Jianhui Liu, Haoze Sun, Wenbo Li et al.

Spatial understanding is a fundamental cornerstone of human-level intelligence. Nonetheless, current research predominantly focuses on domain-specific data production, leaving a critical void: the absence of a principled, open-source engine capable of fully unleashing the potential of high-quality spatial data. To bridge this gap, we elucidate the design principles of a robust data generation system and introduce OpenSpatial -- an open-source data engine engineered for high quality, extensive scalability, broad task diversity, and optimized efficiency. OpenSpatial adopts 3D bounding boxes as the fundamental primitive to construct a comprehensive data hierarchy across five foundational tasks: Spatial Measurement (SM), Spatial Relationship (SR), Camera Perception (CP), Multi-view Consistency (MC), and Scene-Aware Reasoning (SAR). Leveraging this scalable infrastructure, we curate OpenSpatial-3M, a large-scale dataset comprising 3 million high-fidelity samples. Extensive evaluations demonstrate that versatile models trained on our dataset achieve state-of-the-art performance across a wide spectrum of spatial reasoning benchmarks. Notably, the best-performing model exhibits a substantial average improvement of 19 percent, relatively. Furthermore, we provide a systematic analysis of how data attributes influence spatial perception. By open-sourcing both the engine and the 3M-scale dataset, we provide a robust foundation to accelerate future research in spatial intelligence.

Claire Chen, Yuheng Zhang, Xinyu Liu et al.

We study the problem of learning Nash equilibria in offline two-player zero-sum Markov games. While existing approaches often rely on explicit pessimism to address distribution shift, we show that KL regularization alone suffices to stabilize learning and guarantee convergence. We first introduce Regularized Offline Sequential Equilibrium (ROSE), a theoretical framework that achieves a fast $\widetilde{\mathcal{O}}(1/n)$ convergence rate under \textit{unilateral concentrability}, improving over the standard $\widetilde{\mathcal{O}}(1/\sqrt{n})$ rates in unregularized settings. We then propose Sequential Offline Self-play Mirror Descent (SOS-MD), a practical model-free algorithm based on least-squares value estimation and iterative self-play updates. We prove that the last iterate of SOS-MD attains the same $\widetilde{\mathcal{O}}(1/n)$ statistical rate up to a vanishing optimization error of order $\widetilde{\mathcal{O}}(1/\sqrt{T})$ in the number of self-play iterations $T$.

Haoyang Zheng, Xinyang Liu, Cindy Xiangrui Kong et al.

Fast and high-quality language generation is the holy grail that people pursue in the age of AI. In this work, we introduce Discrete Diffusion Divergence Instruct (DiDi-Instruct), a training-based method that initializes from a pre-trained (masked) discrete diffusion language model (dLLM) and distills a few-step student for fast generation. The resulting DiDi-Instruct model achieves comparable or superior performance to its dLLM teacher and the GPT-2 baseline while enabling up to 64$\times$ acceleration. The theoretical foundation of DiDi-Instruct is a novel framework based on integral KL-divergence minimization, which yields a practical training algorithm. We further introduce grouped reward normalization, intermediate-state matching, and the reward-guided ancestral sampler that significantly improve training stability, model coverage, and inference quality. On OpenWebText, DiDi-Instruct achieves perplexity from 62.2 (8 NFEs) to 18.4 (128 NFEs), which outperforms prior accelerated dLLMs and GPT-2 baseline. These gains come with a negligible entropy loss (around $1\%$) and reduce additional training wall-clock time by more than $20\times$ compared to competing dLLM distillation methods. We further validate the robustness and effectiveness of DiDi-Instruct through extensive ablation studies, model scaling, and the generation of discrete protein sequences. In conclusion, DiDi-Instruct is an efficient yet effective distillation method, enabling language generation in the blink of an eye. We will release both code and models at github.com/haoyangzheng-ai/didi-instruct.

Wei Xiong, Hanze Dong, Chenlu Ye et al.

This paper studies the alignment process of generative models with Reinforcement Learning from Human Feedback (RLHF). We first identify the primary challenges of existing popular methods like offline PPO and offline DPO as lacking in strategical exploration of the environment. Then, to understand the mathematical principle of RLHF, we consider a standard mathematical formulation, the reverse-KL regularized contextual bandit for RLHF. Despite its widespread practical application, a rigorous theoretical analysis of this formulation remains open. We investigate its behavior in three distinct settings -- offline, online, and hybrid -- and propose efficient algorithms with finite-sample theoretical guarantees. Moving towards practical applications, our framework, with a robust approximation of the information-theoretical policy improvement oracle, naturally gives rise to several novel RLHF algorithms. This includes an iterative version of the Direct Preference Optimization (DPO) algorithm for online settings, and a multi-step rejection sampling strategy for offline scenarios. Our empirical evaluations on real-world alignment experiment of large language model demonstrate that these proposed methods significantly surpass existing strong baselines, such as DPO and Rejection Sampling Optimization (RSO), showcasing the connections between solid theoretical foundations and their potent practical implementations.