Yitao Liang

Zihao Wang, Shaofei Cai, Guanzhou Chen et al. · pku

We investigate the challenge of task planning for multi-task embodied agents in open-world environments. Two main difficulties are identified: 1) executing plans in an open-world environment (e.g., Minecraft) necessitates accurate and multi-step reasoning due to the long-term nature of tasks, and 2) as vanilla planners do not consider how easy the current agent can achieve a given sub-task when ordering parallel sub-goals within a complicated plan, the resulting plan could be inefficient or even infeasible. To this end, we propose "$\underline{D}$escribe, $\underline{E}$xplain, $\underline{P}$lan and $\underline{S}$elect" ($\textbf{DEPS}$), an interactive planning approach based on Large Language Models (LLMs). DEPS facilitates better error correction on initial LLM-generated $\textit{plan}$ by integrating $\textit{description}$ of the plan execution process and providing self-$\textit{explanation}$ of feedback when encountering failures during the extended planning phases. Furthermore, it includes a goal $\textit{selector}$, which is a trainable module that ranks parallel candidate sub-goals based on the estimated steps of completion, consequently refining the initial plan. Our experiments mark the milestone of the first zero-shot multi-task agent that can robustly accomplish 70+ Minecraft tasks and nearly double the overall performances. Further testing reveals our method's general effectiveness in popularly adopted non-open-ended domains as well (i.e., ALFWorld and tabletop manipulation). The ablation and exploratory studies detail how our design beats the counterparts and provide a promising update on the $\texttt{ObtainDiamond}$ grand challenge with our approach. The code is released at https://github.com/CraftJarvis/MC-Planner.

Xinyue Zheng, Haowei Lin, Kaichen He et al. · pku

Developing AI agents capable of interacting with open-world environments to solve diverse tasks is a compelling challenge. However, evaluating such open-ended agents remains difficult, with current benchmarks facing scalability limitations. To address this, we introduce Minecraft Universe (MCU), a comprehensive evaluation framework set within the open-world video game Minecraft. MCU incorporates three key components: (1) an expanding collection of 3,452 composable atomic tasks that encompasses 11 major categories and 41 subcategories of challenges; (2) a task composition mechanism capable of generating infinite diverse tasks with varying difficulty; and (3) a general evaluation framework that achieves 91.5\% alignment with human ratings for open-ended task assessment. Empirical results reveal that even state-of-the-art foundation agents struggle with the increasing diversity and complexity of tasks. These findings highlight the necessity of MCU as a robust benchmark to drive progress in AI agent development within open-ended environments. Our evaluation code and scripts are available at https://github.com/CraftJarvis/MCU.

Ceyao Zhang, Kaijie Yang, Siyi Hu et al. · pku

Building agents with adaptive behavior in cooperative tasks stands as a paramount goal in the realm of multi-agent systems. Current approaches to developing cooperative agents rely primarily on learning-based methods, whose policy generalization depends heavily on the diversity of teammates they interact with during the training phase. Such reliance, however, constrains the agents' capacity for strategic adaptation when cooperating with unfamiliar teammates, which becomes a significant challenge in zero-shot coordination scenarios. To address this challenge, we propose ProAgent, a novel framework that harnesses large language models (LLMs) to create proactive agents capable of dynamically adapting their behavior to enhance cooperation with teammates. ProAgent can analyze the present state, and infer the intentions of teammates from observations. It then updates its beliefs in alignment with the teammates' subsequent actual behaviors. Moreover, ProAgent exhibits a high degree of modularity and interpretability, making it easily integrated into various of coordination scenarios. Experimental evaluations conducted within the Overcooked-AI environment unveil the remarkable performance superiority of ProAgent, outperforming five methods based on self-play and population-based training when cooperating with AI agents. Furthermore, in partnered with human proxy models, its performance exhibits an average improvement exceeding 10% compared to the current state-of-the-art method. For more information about our project, please visit~\url{https://pku-proagent.github.io}.

Xiaojian Ma, Silong Yong, Zilong Zheng et al.

We propose a new task to benchmark scene understanding of embodied agents: Situated Question Answering in 3D Scenes (SQA3D). Given a scene context (e.g., 3D scan), SQA3D requires the tested agent to first understand its situation (position, orientation, etc.) in the 3D scene as described by text, then reason about its surrounding environment and answer a question under that situation. Based upon 650 scenes from ScanNet, we provide a dataset centered around 6.8k unique situations, along with 20.4k descriptions and 33.4k diverse reasoning questions for these situations. These questions examine a wide spectrum of reasoning capabilities for an intelligent agent, ranging from spatial relation comprehension to commonsense understanding, navigation, and multi-hop reasoning. SQA3D imposes a significant challenge to current multi-modal especially 3D reasoning models. We evaluate various state-of-the-art approaches and find that the best one only achieves an overall score of 47.20%, while amateur human participants can reach 90.06%. We believe SQA3D could facilitate future embodied AI research with stronger situation understanding and reasoning capability.

Zihao Wang, Shaofei Cai, Anji Liu et al. · pku

Achieving human-like planning and control with multimodal observations in an open world is a key milestone for more functional generalist agents. Existing approaches can handle certain long-horizon tasks in an open world. However, they still struggle when the number of open-world tasks could potentially be infinite and lack the capability to progressively enhance task completion as game time progresses. We introduce JARVIS-1, an open-world agent that can perceive multimodal input (visual observations and human instructions), generate sophisticated plans, and perform embodied control, all within the popular yet challenging open-world Minecraft universe. Specifically, we develop JARVIS-1 on top of pre-trained multimodal language models, which map visual observations and textual instructions to plans. The plans will be ultimately dispatched to the goal-conditioned controllers. We outfit JARVIS-1 with a multimodal memory, which facilitates planning using both pre-trained knowledge and its actual game survival experiences. JARVIS-1 is the existing most general agent in Minecraft, capable of completing over 200 different tasks using control and observation space similar to humans. These tasks range from short-horizon tasks, e.g., "chopping trees" to long-horizon tasks, e.g., "obtaining a diamond pickaxe". JARVIS-1 performs exceptionally well in short-horizon tasks, achieving nearly perfect performance. In the classic long-term task of $\texttt{ObtainDiamondPickaxe}$, JARVIS-1 surpasses the reliability of current state-of-the-art agents by 5 times and can successfully complete longer-horizon and more challenging tasks. The project page is available at https://craftjarvis.org/JARVIS-1

Xiaojuan Tang, Song-Chun Zhu, Yitao Liang et al.

Knowledge graph (KG) reasoning is an important problem for knowledge graphs. In this paper, we propose a novel and principled framework called \textbf{RulE} (stands for {Rul}e {E}mbedding) to effectively leverage logical rules to enhance KG reasoning. Unlike knowledge graph embedding (KGE) methods, RulE learns rule embeddings from existing triplets and first-order {rules} by jointly representing \textbf{entities}, \textbf{relations} and \textbf{logical rules} in a unified embedding space. Based on the learned rule embeddings, a confidence score can be calculated for each rule, reflecting its consistency with the observed triplets. This allows us to perform logical rule inference in a soft way, thus alleviating the brittleness of logic. On the other hand, RulE injects prior logical rule information into the embedding space, enriching and regularizing the entity/relation embeddings. This makes KGE alone perform better too. RulE is conceptually simple and empirically effective. We conduct extensive experiments to verify each component of RulE. Results on multiple benchmarks reveal that our model outperforms the majority of existing embedding-based and rule-based approaches.

Shaofei Cai, Zihao Wang, Xiaojian Ma et al. · pku

We study the problem of learning goal-conditioned policies in Minecraft, a popular, widely accessible yet challenging open-ended environment for developing human-level multi-task agents. We first identify two main challenges of learning such policies: 1) the indistinguishability of tasks from the state distribution, due to the vast scene diversity, and 2) the non-stationary nature of environment dynamics caused by partial observability. To tackle the first challenge, we propose Goal-Sensitive Backbone (GSB) for the policy to encourage the emergence of goal-relevant visual state representations. To tackle the second challenge, the policy is further fueled by an adaptive horizon prediction module that helps alleviate the learning uncertainty brought by the non-stationary dynamics. Experiments on 20 Minecraft tasks show that our method significantly outperforms the best baseline so far; in many of them, we double the performance. Our ablation and exploratory studies then explain how our approach beat the counterparts and also unveil the surprising bonus of zero-shot generalization to new scenes (biomes). We hope our agent could help shed some light on learning goal-conditioned, multi-task agents in challenging, open-ended environments like Minecraft.

Haotian Ye, Haowei Lin, Jiaqi Han et al. · pku

Given an unconditional diffusion model and a predictor for a target property of interest (e.g., a classifier), the goal of training-free guidance is to generate samples with desirable target properties without additional training. Existing methods, though effective in various individual applications, often lack theoretical grounding and rigorous testing on extensive benchmarks. As a result, they could even fail on simple tasks, and applying them to a new problem becomes unavoidably difficult. This paper introduces a novel algorithmic framework encompassing existing methods as special cases, unifying the study of training-free guidance into the analysis of an algorithm-agnostic design space. Via theoretical and empirical investigation, we propose an efficient and effective hyper-parameter searching strategy that can be readily applied to any downstream task. We systematically benchmark across 7 diffusion models on 16 tasks with 40 targets, and improve performance by 8.5% on average. Our framework and benchmark offer a solid foundation for conditional generation in a training-free manner.

Shaofei Cai, Bowei Zhang, Zihao Wang et al. · pku

We study the problem of building a controller that can follow open-ended instructions in open-world environments. We propose to follow reference videos as instructions, which offer expressive goal specifications while eliminating the need for expensive text-gameplay annotations. A new learning framework is derived to allow learning such instruction-following controllers from gameplay videos while producing a video instruction encoder that induces a structured goal space. We implement our agent GROOT in a simple yet effective encoder-decoder architecture based on causal transformers. We evaluate GROOT against open-world counterparts and human players on a proposed Minecraft SkillForge benchmark. The Elo ratings clearly show that GROOT is closing the human-machine gap as well as exhibiting a 70% winning rate over the best generalist agent baseline. Qualitative analysis of the induced goal space further demonstrates some interesting emergent properties, including the goal composition and complex gameplay behavior synthesis. The project page is available at https://craftjarvis-groot.github.io.

Qing Li, Yixin Zhu, Yitao Liang et al.

Current learning models often struggle with human-like systematic generalization, particularly in learning compositional rules from limited data and extrapolating them to novel combinations. We introduce the Neural-Symbolic Recursive Machine (NSR), whose core is a Grounded Symbol System (GSS), allowing for the emergence of combinatorial syntax and semantics directly from training data. The NSR employs a modular design that integrates neural perception, syntactic parsing, and semantic reasoning. These components are synergistically trained through a novel deduction-abduction algorithm. Our findings demonstrate that NSR's design, imbued with the inductive biases of equivariance and compositionality, grants it the expressiveness to adeptly handle diverse sequence-to-sequence tasks and achieve unparalleled systematic generalization. We evaluate NSR's efficacy across four challenging benchmarks designed to probe systematic generalization capabilities: SCAN for semantic parsing, PCFG for string manipulation, HINT for arithmetic reasoning, and a compositional machine translation task. The results affirm NSR's superiority over contemporary neural and hybrid models in terms of generalization and transferability.

Xuejie Liu, Anji Liu, Guy Van den Broeck et al.

Probabilistic Circuits (PCs) are a general and unified computational framework for tractable probabilistic models that support efficient computation of various inference tasks (e.g., computing marginal probabilities). Towards enabling such reasoning capabilities in complex real-world tasks, Liu et al. (2022) propose to distill knowledge (through latent variable assignments) from less tractable but more expressive deep generative models. However, it is still unclear what factors make this distillation work well. In this paper, we theoretically and empirically discover that the performance of a PC can exceed that of its teacher model. Therefore, instead of performing distillation from the most expressive deep generative model, we study what properties the teacher model and the PC should have in order to achieve good distillation performance. This leads to a generic algorithmic improvement as well as other data-type-specific ones over the existing latent variable distillation pipeline. Empirically, we outperform SoTA TPMs by a large margin on challenging image modeling benchmarks. In particular, on ImageNet32, PCs achieve 4.06 bits-per-dimension, which is only 0.34 behind variational diffusion models (Kingma et al., 2021).

Kaichen He, Zihao Wang, Muyao Li et al.

The paradigm of agentic AI is shifting from engineered complex workflows to post-training native models. However, existing agents are typically confined to static, predefined action spaces--such as exclusively using APIs, GUI events, or robotic commands. This rigidity limits their adaptability in dynamic environments where the optimal granularity of interaction varies contextually. To bridge this gap, we propose CrossAgent, a unified agentic model that masters heterogeneous action spaces and autonomously selects the most effective interface for each step of a trajectory. We introduce a comprehensive training pipeline that integrates cold-start supervised fine-tuning with a Multi-Turn Group Relative Policy Optimization (GRPO) algorithm. This approach enables the agent to learn adaptive action switching--balancing high-level efficiency with low-level precision--without human-specified rules. Extensive experiments on over 800 tasks in the open-world Minecraft environment demonstrate that CrossAgent achieves state-of-the-art performance. By dynamically leveraging the strengths of diverse action spaces, our model significantly outperforms fixed-action baselines, exhibiting superior generalization and efficiency in long-horizon reasoning. All code and models are available at https://github.com/CraftJarvis/OpenHA

Zhizhou Ren, Anji Liu, Yitao Liang et al.

Learning new task-specific skills from a few trials is a fundamental challenge for artificial intelligence. Meta reinforcement learning (meta-RL) tackles this problem by learning transferable policies that support few-shot adaptation to unseen tasks. Despite recent advances in meta-RL, most existing methods require the access to the environmental reward function of new tasks to infer the task objective, which is not realistic in many practical applications. To bridge this gap, we study the problem of few-shot adaptation in the context of human-in-the-loop reinforcement learning. We develop a meta-RL algorithm that enables fast policy adaptation with preference-based feedback. The agent can adapt to new tasks by querying human's preference between behavior trajectories instead of using per-step numeric rewards. By extending techniques from information theory, our approach can design query sequences to maximize the information gain from human interactions while tolerating the inherent error of non-expert human oracle. In experiments, we extensively evaluate our method, Adaptation with Noisy OracLE (ANOLE), on a variety of meta-RL benchmark tasks and demonstrate substantial improvement over baseline algorithms in terms of both feedback efficiency and error tolerance.

Shaofei Cai, Yulei Qin, Haojia Lin et al.

Agentic reinforcement learning (RL) holds great promise for the development of autonomous agents under complex GUI tasks, but its scalability remains severely hampered by the verification of task completion. Existing task verification is treated as a passive, post-hoc process: a verifier (i.e., rule-based scoring script, reward or critic model, and LLM-as-a-Judge) analyzes the agent's entire interaction trajectory to determine if the agent succeeds. Such processing of verbose context that contains irrelevant, noisy history poses challenges to the verification protocols and therefore leads to prohibitive cost and low reliability. To overcome this bottleneck, we propose SmartSnap, a paradigm shift from this passive, post-hoc verification to proactive, in-situ self-verification by the agent itself. We introduce the Self-Verifying Agent, a new type of agent designed with dual missions: to not only complete a task but also to prove its accomplishment with curated snapshot evidences. Guided by our proposed 3C Principles (Completeness, Conciseness, and Creativity), the agent leverages its accessibility to the online environment to perform self-verification on a minimal, decisive set of snapshots. Such evidences are provided as the sole materials for a general LLM-as-a-Judge verifier to determine their validity and relevance. Experiments on mobile tasks across model families and scales demonstrate that our SmartSnap paradigm allows training LLM-driven agents in a scalable manner, bringing performance gains up to 26.08% and 16.66% respectively to 8B and 30B models. The synergizing between solution finding and evidence seeking facilitates the cultivation of efficient, self-verifying agents with competitive performance against DeepSeek V3.1 and Qwen3-235B-A22B. Code is available at: https://github.com/TencentYoutuResearch/SmartSnap

Xuejie Liu, Anji Liu, Guy Van den Broeck et al.

A popular paradigm for offline Reinforcement Learning (RL) tasks is to first fit the offline trajectories to a sequence model, and then prompt the model for actions that lead to high expected return. In addition to obtaining accurate sequence models, this paper highlights that tractability, the ability to exactly and efficiently answer various probabilistic queries, plays an important role in offline RL. Specifically, due to the fundamental stochasticity from the offline data-collection policies and the environment dynamics, highly non-trivial conditional/constrained generation is required to elicit rewarding actions. it is still possible to approximate such queries, we observe that such crude estimates significantly undermine the benefits brought by expressive sequence models. To overcome this problem, this paper proposes Trifle (Tractable Inference for Offline RL), which leverages modern Tractable Probabilistic Models (TPMs) to bridge the gap between good sequence models and high expected returns at evaluation time. Empirically, Trifle achieves the most state-of-the-art scores in 9 Gym-MuJoCo benchmarks against strong baselines. Further, owing to its tractability, Trifle significantly outperforms prior approaches in stochastic environments and safe RL tasks (e.g. with action constraints) with minimum algorithmic modifications.

Xuejie Liu, Yap Vit Chun, Yitao Liang et al.

Diffusion Large Language Models (dLLMs) support arbitrary-order generation, yet their inference performance critically depends on the unmasking order. Existing strategies rely on heuristics that greedily optimize local confidence, offering limited guidance for identifying unmasking paths that are globally consistent and accurate. To bridge this gap, we introduce path log-likelihood (Path LL), a trajectory-conditioned objective that strongly correlates with downstream accuracy and enables principled selection of unmasking paths. To optimize Path LL at inference time, we propose POKE, an efficient value estimator that predicts the expected future Path LL of a partial decoding trajectory. We then integrate this lookahead signal into POKE-SMC, a Sequential Monte Carlo-based search framework for dynamically identifying optimal unmasking paths. Extensive experiments across 6 reasoning tasks show that POKE-SMC consistently improves accuracy, achieving 2%--3% average gains over strong decoding-time scaling baselines at comparable inference overhead on LLaDA models and advancing the accuracy--compute Pareto frontier.

Muyao Li, Zihao Wang, Kaichen He et al. · pku

Recently, action-based decision-making in open-world environments has gained significant attention. Visual Language Action (VLA) models, pretrained on large-scale web datasets, have shown promise in decision-making tasks. However, previous work has primarily focused on action post-training, often neglecting enhancements to the foundational model itself. In response, we introduce a novel approach, Act from Visual Language Post-Training, which refines Visual Language Models (VLMs) through visual and linguistic guidance in a self-supervised manner. This enhancement improves the models' capabilities in world knowledge, visual recognition, and spatial grounding in open-world environments. Following the above post-training paradigms, we obtain the first VLA models in Minecraft that can follow human instructions on over 1k different atomic tasks, including crafting, smelting, cooking, mining, and killing. Our experiments demonstrate that post-training on non-trajectory tasks leads to a significant 40% improvement over the best agent baseline on a diverse set of atomic tasks. Furthermore, we demonstrate that our approach surpasses traditional imitation learning-based policies in Minecraft, achieving state-of-the-art performance. We have open-sourced the code, models, and datasets to foster further research. The project page can be found in https://craftjarvis.github.io/JarvisVLA.

Sunbowen Lee, Junting Zhou, Chang Ao et al.

Model distillation is a fundamental technique in building large language models (LLMs), transferring knowledge from a teacher model to a student model. However, distillation can lead to model homogenization, reducing diversity among models and impairing their ability to robustly handle complex or novel tasks. These limitations underscore the need to systematically quantify the distillation process and its impact. In this work, we propose a framework to evaluate and quantify model distillation. Our method addresses two key aspects: (1) Identifying identity cognition contradictions to assess discrepancies in how models perceive and represent identity-related information, and (2) Analyzing multi-granularity response similarities across models to measure the extent of homogenization. Experimental results demonstrate two key insights: (1) Well-known closed-source and open-source LLMs usually exhibit high distillation degrees, except for Claude, Doubao, and Gemini. (2) Base LLMs show higher distillation degrees compared to aligned LLMs. By offering a systematic approach to improve the transparency of LLM data distillation, we call for LLMs with more independent development and more transparent technical reports to improve LLMs' robustness and safety. The code and data are available under https://github.com/Aegis1863/LLMs-Distillation-Quantification.

Shaofei Cai, Zhancun Mu, Kaichen He et al. · pku

Minecraft's complexity and diversity as an open world make it a perfect environment to test if agents can learn, adapt, and tackle a variety of unscripted tasks. However, the development and validation of novel agents in this setting continue to face significant engineering challenges. This paper presents MineStudio, an open-source software package designed to streamline the development of autonomous agents in Minecraft. MineStudio represents the first comprehensive integration of seven critical engineering components: simulator, data, model, offline pre-training, online fine-tuning, inference, and benchmark, thereby allowing users to concentrate their efforts on algorithm innovation. We provide a user-friendly API design accompanied by comprehensive documentation and tutorials. Our project is released at https://github.com/CraftJarvis/MineStudio.

Guangyu Zhao, Ceyao Zhang, Chengdong Ma et al.

Long-horizon tabletop games pose a distinct systems challenge for robotics: small perceptual or execution errors can invalidate accumulated task state, propagate across decision-making modules, and ultimately derail interaction. This paper studies how to maintain internal state consistency in turn-based, multi-human robotic tabletop games through deliberate system design rather than isolated component improvement. Using Mahjong as a representative long-horizon setting, we present an integrated architecture that explicitly maintains perceptual, execution, and interaction state, partitions high-level semantic reasoning from time-critical perception and control, and incorporates verified action primitives with tactile-triggered recovery to prevent premature state corruption. We further introduce interaction-level monitoring mechanisms to detect turn violations and hidden-information breaches that threaten execution assumptions. Beyond demonstrating complete-game operation, we provide an empirical characterization of failure modes, recovery effectiveness, cross-module error propagation, and hardware-algorithm trade-offs observed during deployment. Our results show that explicit partitioning, monitored state transitions, and recovery mechanisms are critical for sustaining executable consistency over extended play, whereas monolithic or unverified pipelines lead to measurable degradation in end-to-end reliability. The proposed system serves as an empirical platform for studying system-level design principles in long-horizon, turn-based interaction.

Zixia Jia, Jiaqi Li, Yipeng Kang et al.

Memory plays a foundational role in augmenting the reasoning, adaptability, and contextual fidelity of modern Large Language Models and Multi-Modal LLMs. As these models transition from static predictors to interactive systems capable of continual learning and personalized inference, the incorporation of memory mechanisms has emerged as a central theme in their architectural and functional evolution. This survey presents a comprehensive and structured synthesis of memory in LLMs and MLLMs, organizing the literature into a cohesive taxonomy comprising implicit, explicit, and agentic memory paradigms. Specifically, the survey delineates three primary memory frameworks. Implicit memory refers to the knowledge embedded within the internal parameters of pre-trained transformers, encompassing their capacity for memorization, associative retrieval, and contextual reasoning. Recent work has explored methods to interpret, manipulate, and reconfigure this latent memory. Explicit memory involves external storage and retrieval components designed to augment model outputs with dynamic, queryable knowledge representations, such as textual corpora, dense vectors, and graph-based structures, thereby enabling scalable and updatable interaction with information sources. Agentic memory introduces persistent, temporally extended memory structures within autonomous agents, facilitating long-term planning, self-consistency, and collaborative behavior in multi-agent systems, with relevance to embodied and interactive AI. Extending beyond text, the survey examines the integration of memory within multi-modal settings, where coherence across vision, language, audio, and action modalities is essential. Key architectural advances, benchmark tasks, and open challenges are discussed, including issues related to memory capacity, alignment, factual consistency, and cross-system interoperability.

Zihao Wang, Muyao Li, Kaichen He et al. · pku

The choice of action spaces is a critical yet unresolved challenge in developing capable, end-to-end trainable agents. This paper first presents a large-scale, systematic comparison of prominent abstracted action spaces and tokenizers for Vision-Language-Action (VLA) or hierarchical agent models in the open-ended Minecraft. Our analysis reveals that no single action space is universally optimal; instead, the most effective abstraction is highly task-dependent, creating a dilemma for building generalist agents. To resolve this, we introduce Chain of Action (CoA), a novel framework that unifies high-level planning and low-level control within a single, monolithic VLA model. CoA treats an abstracted action not as a command for a separate policy, but as an intermediate reasoning step--akin to a chain of thought--that guides the generation of the final, executable action. Furthermore, we demonstrate that an All-in-One agent trained on a diverse mixture of action spaces using the CoA paradigm learns a more robust and generalizable policy. This unified agent achieves a new state-of-the-art, improving the overall task success rate over strong, specialized baselines. To foster reproducible research, we release the OpenHA (Open Hierarchical Agents) suite, which includes our comprehensive benchmark of over 800 distinct tasks, curated datasets, source code, and all pretrained model checkpoints at https://github.com/CraftJarvis/OpenHA

Junting Zhou, Tingjia Miao, Yiyan Liao et al.

Advancement in Large Language Models (LLMs) reasoning capabilities enables them to solve scientific problems with enhanced efficacy. Thereby, a high-quality benchmark for comprehensive and appropriate assessment holds significance, while existing ones either confront the risk of data contamination or lack involved disciplines. To be specific, due to the data source overlap of LLMs training and static benchmark, the keys or number pattern of answers inadvertently memorized (i.e. data contamination), leading to systematic overestimation of their reasoning capabilities, especially numerical reasoning. We propose SciDA, a multidisciplinary benchmark that consists exclusively of over 1k Olympic-level numerical computation problems, allowing randomized numerical initializations for each inference round to avoid reliance on fixed numerical patterns. We conduct a series of experiments with both closed-source and open-source top-performing LLMs, and it is observed that the performance of LLMs drop significantly under random numerical initialization. Thus, we provide truthful and unbiased assessments of the numerical reasoning capabilities of LLMs. The data is available at https://huggingface.co/datasets/m-a-p/SciDA

Kewei Lian, Shaofei Cai, Yilun Du et al. · pku

The ability to simulate the world in a spatially consistent manner is a crucial requirements for effective world models. Such a model enables high-quality visual generation, and also ensures the reliability of world models for downstream tasks such as simulation and planning. Designing a memory module is a crucial component for addressing spatial consistency: such a model must not only retain long-horizon observational information, but also enables the construction of explicit or implicit internal spatial representations. However, there are no dataset designed to promote the development of memory modules by explicitly enforcing spatial consistency constraints. Furthermore, most existing benchmarks primarily emphasize visual coherence or generation quality, neglecting the requirement of long-range spatial consistency. To bridge this gap, we construct a dataset and corresponding benchmark by sampling 150 distinct locations within the open-world environment of Minecraft, collecting about 250 hours (20 million frames) of loop-based navigation videos with actions. Our dataset follows a curriculum design of sequence lengths, allowing models to learn spatial consistency on increasingly complex navigation trajectories. Furthermore, our data collection pipeline is easily extensible to new Minecraft environments and modules. Four representative world model baselines are evaluated on our benchmark. Dataset, benchmark, and code are open-sourced to support future research.

Haotian Zhang, Junting Zhou, Haowei Lin et al.

Continual Learning (CL) poses a significant challenge in Artificial Intelligence, aiming to mirror the human ability to incrementally acquire knowledge and skills. While extensive research has focused on CL within the context of classification tasks, the advent of increasingly powerful generative models necessitates the exploration of Continual Learning of Generative models (CLoG). This paper advocates for shifting the research focus from classification-based CL to CLoG. We systematically identify the unique challenges presented by CLoG compared to traditional classification-based CL. We adapt three types of existing CL methodologies, replay-based, regularization-based, and parameter-isolation-based methods to generative tasks and introduce comprehensive benchmarks for CLoG that feature great diversity and broad task coverage. Our benchmarks and results yield intriguing insights that can be valuable for developing future CLoG methods. Additionally, we will release a codebase designed to facilitate easy benchmarking and experimentation in CLoG publicly at https://github.com/linhaowei1/CLoG. We believe that shifting the research focus to CLoG will benefit the continual learning community and illuminate the path for next-generation AI-generated content (AIGC) in a lifelong learning paradigm.

Xiaojuan Tang, Zilong Zheng, Jiaqi Li et al.

The emergent few-shot reasoning capabilities of Large Language Models (LLMs) have excited the natural language and machine learning community over recent years. Despite of numerous successful applications, the underlying mechanism of such in-context capabilities still remains unclear. In this work, we hypothesize that the learned \textit{semantics} of language tokens do the most heavy lifting during the reasoning process. Different from human's symbolic reasoning process, the semantic representations of LLMs could create strong connections among tokens, thus composing a superficial logical chain. To test our hypothesis, we decouple semantics from the language reasoning process and evaluate three kinds of reasoning abilities, i.e., deduction, induction and abduction. Our findings reveal that semantics play a vital role in LLMs' in-context reasoning -- LLMs perform significantly better when semantics are consistent with commonsense but struggle to solve symbolic or counter-commonsense reasoning tasks by leveraging in-context new knowledge. The surprising observations question whether modern LLMs have mastered the inductive, deductive and abductive reasoning abilities as in human intelligence, and motivate research on unveiling the magic existing within the black-box LLMs. On the whole, our analysis provides a novel perspective on the role of semantics in developing and evaluating language models' reasoning abilities. Code is available at {\url{https://github.com/XiaojuanTang/ICSR}}.

Albert Zhao, Tong He, Yitao Liang et al.

We describe a policy learning approach to map visual inputs to driving controls conditioned on turning command that leverages side tasks on semantics and object affordances via a learned representation trained for driving. To learn this representation, we train a squeeze network to drive using annotations for the side task as input. This representation encodes the driving-relevant information associated with the side task while ideally throwing out side task-relevant but driving-irrelevant nuisances. We then train a mimic network to drive using only images as input and use the squeeze network's latent representation to supervise the mimic network via a mimicking loss. Notably, we do not aim to achieve the side task nor to learn features for it; instead, we aim to learn, via the mimicking loss, a representation of the side task annotations directly useful for driving. We test our approach using the CARLA simulator. In addition, we introduce a more challenging but realistic evaluation protocol that considers a run that reaches the destination successful only if it does not violate common traffic rules. A video summarizing this work is available at https://youtu.be/ipKAMzmJpMs , and code is available at https://github.com/twsq/sam-driving .

Jason Gauci, Edoardo Conti, Yitao Liang et al.

In this paper we present Horizon, Facebook's open source applied reinforcement learning (RL) platform. Horizon is an end-to-end platform designed to solve industry applied RL problems where datasets are large (millions to billions of observations), the feedback loop is slow (vs. a simulator), and experiments must be done with care because they don't run in a simulator. Unlike other RL platforms, which are often designed for fast prototyping and experimentation, Horizon is designed with production use cases as top of mind. The platform contains workflows to train popular deep RL algorithms and includes data preprocessing, feature transformation, distributed training, counterfactual policy evaluation, optimized serving, and a model-based data understanding tool. We also showcase and describe real examples where reinforcement learning models trained with Horizon significantly outperformed and replaced supervised learning systems at Facebook.

Zihao Wang, Anji Liu, Haowei Lin et al. · pku

We explore how iterative revising a chain of thoughts with the help of information retrieval significantly improves large language models' reasoning and generation ability in long-horizon generation tasks, while hugely mitigating hallucination. In particular, the proposed method -- *retrieval-augmented thoughts* (RAT) -- revises each thought step one by one with retrieved information relevant to the task query, the current and the past thought steps, after the initial zero-shot CoT is generated. Applying RAT to GPT-3.5, GPT-4, and CodeLLaMA-7b substantially improves their performances on various long-horizon generation tasks; on average of relatively increasing rating scores by 13.63% on code generation, 16.96% on mathematical reasoning, 19.2% on creative writing, and 42.78% on embodied task planning. The demo page can be found at https://craftjarvis.github.io/RAT

Yifan Zhong, Xuchuan Huang, Ruochong Li et al.

Dexterous grasping remains a fundamental yet challenging problem in robotics. A general-purpose robot must be capable of grasping diverse objects in arbitrary scenarios. However, existing research typically relies on restrictive assumptions, such as single-object settings or limited environments, showing constrained generalization. We present DexGraspVLA, a hierarchical framework for robust generalization in language-guided general dexterous grasping and beyond. It utilizes a pre-trained Vision-Language model as the high-level planner and learns a diffusion-based low-level Action controller. The key insight to achieve generalization lies in iteratively transforming diverse language and visual inputs into domain-invariant representations via foundation models, where imitation learning can be effectively applied due to the alleviation of domain shift. Notably, our method achieves a 90+% dexterous grasping success rate under thousands of challenging unseen cluttered scenes. Empirical analysis confirms the consistency of internal model behavior across environmental variations, validating our design. DexGraspVLA also, for the first time, simultaneously demonstrates free-form long-horizon prompt execution, robustness to adversarial objects and human disturbance, and failure recovery. Extended application to nonprehensile grasping further proves its generality. Project website: https://dexgraspvla.github.io.

Haowei Lin, Baizhou Huang, Haotian Ye et al. · pku

The ever-growing ecosystem of LLMs has posed a challenge in selecting the most appropriate pre-trained model to fine-tune amidst a sea of options. Given constrained resources, fine-tuning all models and making selections afterward is unrealistic. In this work, we formulate this resource-constrained selection task into predicting fine-tuning performance and illustrate its natural connection with Scaling Law. Unlike pre-training, we find that the fine-tuning scaling curve includes not just the well-known "power phase" but also the previously unobserved "pre-power phase". We also explain why existing Scaling Law fails to capture this phase transition phenomenon both theoretically and empirically. To address this, we introduce the concept of "pre-learned data size" into our Rectified Scaling Law, which overcomes theoretical limitations and fits experimental results much better. By leveraging our law, we propose a novel LLM selection algorithm that selects the near-optimal model with hundreds of times less resource consumption, while other methods may provide negatively correlated selection. The project page is available at rectified-scaling-law.github.io.

Zhou Ziheng, Huacong Tang, Jinyuan Zhang et al.

Discovering causal regularities and applying them to build functional systems--the discovery-to-application loop--is a hallmark of general intelligence, yet evaluating this capacity has been hindered by the vast complexity gap between scientific discovery and real-world engineering. We introduce SciCrafter, a Minecraft-based benchmark that operationalizes this loop through parameterized redstone circuit tasks. Agents must ignite lamps in specified patterns (e.g., simultaneously or in timed sequences); scaling target parameters substantially increases construction complexity and required knowledge, forcing genuine discovery rather than reliance on memorized solutions. Evaluating frontier models including GPT-5.2, Gemini-3-Pro, and Claude-Opus-4.5 under a general-purpose code agent scaffold, we find that all plateau at approximately 26% success rate. To diagnose these failures, we decompose the loop into four capacities--knowledge gap identification, experimental discovery, knowledge consolidation, and knowledge application--and design targeted interventions whose marginal contributions serve as proxies for corresponding gaps. Our analysis reveals that although the general knowledge application capability still remains as the biggest gap across all models, for frontier models the knowledge gap identification starts to become a major hurdle--indicating the bottleneck is shifting from solving problems right to raising the right problems for current AI. We release SciCrafter as a diagnostic probe for future research on AI systems that navigate the full discovery-to-application loop.

Xiangcheng Zhang, Haowei Lin, Haotian Ye et al.

Classical search algorithms have long underpinned modern artificial intelligence. In this work, we tackle the challenge of inference-time control in diffusion models -- adapting generated outputs to meet diverse test-time objectives -- using principles from classical search. We propose a general framework that orchestrates local and global search to efficiently navigate the generative space. It employs a theoretically grounded local search via annealed Langevin MCMC and performs compute-efficient global exploration using breadth-first and depth-first tree search. We evaluate our approach on a range of challenging domains, including planning, offline reinforcement learning, and image generation. Across all tasks, we observe significant gains in both performance and efficiency. These results show that classical search provides a principled and practical foundation for inference-time scaling in diffusion models. Project page at https://diffusion-inference-scaling.github.io/.

Shaofei Cai, Bowei Zhang, Zihao Wang et al. · pku

Developing agents that can follow multimodal instructions remains a fundamental challenge in robotics and AI. Although large-scale pre-training on unlabeled datasets (no language instruction) has enabled agents to learn diverse behaviors, these agents often struggle with following instructions. While augmenting the dataset with instruction labels can mitigate this issue, acquiring such high-quality annotations at scale is impractical. To address this issue, we frame the problem as a semi-supervised learning task and introduce GROOT-2, a multimodal instructable agent trained using a novel approach that combines weak supervision with latent variable models. Our method consists of two key components: constrained self-imitating, which utilizes large amounts of unlabeled demonstrations to enable the policy to learn diverse behaviors, and human intention alignment, which uses a smaller set of labeled demonstrations to ensure the latent space reflects human intentions. GROOT-2's effectiveness is validated across four diverse environments, ranging from video games to robotic manipulation, demonstrating its robust multimodal instruction-following capabilities.

Shaofei Cai, Zihao Wang, Kewei Lian et al. · pku

Vision-language models (VLMs) have excelled in multimodal tasks, but adapting them to embodied decision-making in open-world environments presents challenges. One critical issue is bridging the gap between discrete entities in low-level observations and the abstract concepts required for effective planning. A common solution is building hierarchical agents, where VLMs serve as high-level reasoners that break down tasks into executable sub-tasks, typically specified using language. However, language suffers from the inability to communicate detailed spatial information. We propose visual-temporal context prompting, a novel communication protocol between VLMs and policy models. This protocol leverages object segmentation from past observations to guide policy-environment interactions. Using this approach, we train ROCKET-1, a low-level policy that predicts actions based on concatenated visual observations and segmentation masks, supported by real-time object tracking from SAM-2. Our method unlocks the potential of VLMs, enabling them to tackle complex tasks that demand spatial reasoning. Experiments in Minecraft show that our approach enables agents to achieve previously unattainable tasks, with a $\mathbf{76}\%$ absolute improvement in open-world interaction performance. Codes and demos are now available on the project page: https://craftjarvis.github.io/ROCKET-1.

Shaofei Cai, Zhancun Mu, Anji Liu et al. · pku

We aim to develop a goal specification method that is semantically clear, spatially sensitive, domain-agnostic, and intuitive for human users to guide agent interactions in 3D environments. Specifically, we propose a novel cross-view goal alignment framework that allows users to specify target objects using segmentation masks from their camera views rather than the agent's observations. We highlight that behavior cloning alone fails to align the agent's behavior with human intent when the human and agent camera views differ significantly. To address this, we introduce two auxiliary objectives: cross-view consistency loss and target visibility loss, which explicitly enhance the agent's spatial reasoning ability. According to this, we develop ROCKET-2, a state-of-the-art agent trained in Minecraft, achieving an improvement in the efficiency of inference 3x to 6x compared to ROCKET-1. We show that ROCKET-2 can directly interpret goals from human camera views, enabling better human-agent interaction. Remarkably, ROCKET-2 demonstrates zero-shot generalization capabilities: despite being trained exclusively on the Minecraft dataset, it can adapt and generalize to other 3D environments like Doom, DMLab, and Unreal through a simple action space mapping.

Chi Zhang, Jiajun Song, Siyu Li et al.

Mathematics olympiads are prestigious competitions, with problem proposing and solving highly honored. Building artificial intelligence that proposes and solves olympiads presents an unresolved challenge in automated theorem discovery and proving, especially in geometry for its combination of numerical and spatial elements. We introduce TongGeometry, a Euclidean geometry system supporting tree-search-based guided problem proposing and solving. The efficient geometry system establishes the most extensive repository of geometry theorems to date: within the same computational budget as the existing state-of-the-art, TongGeometry discovers 6.7 billion geometry theorems requiring auxiliary constructions, including 4.1 billion exhibiting geometric symmetry. Among them, 10 theorems were proposed to regional mathematical olympiads with 3 of TongGeometry's proposals selected in real competitions, earning spots in a national team qualifying exam or a top civil olympiad in China and the US. Guided by fine-tuned large language models, TongGeometry solved all International Mathematical Olympiad geometry in IMO-AG-30, outperforming gold medalists for the first time. It also surpasses the existing state-of-the-art across a broader spectrum of olympiad-level problems. The full capabilities of the system can be utilized on a consumer-grade machine, making the model more accessible and fostering widespread democratization of its use. By analogy, unlike existing systems that merely solve problems like students, TongGeometry acts like a geometry coach, discovering, presenting, and proving theorems.

Guangyu Zhao, Kewei Lian, Haowei Lin et al. · pku

A glowing body of work has emerged focusing on instruction-following policies for open-world agents, aiming to better align the agent's behavior with human intentions. However, the performance of these policies is highly susceptible to the initial prompt, which leads to extra efforts in selecting the best instructions. We propose a framework named Preference Goal Tuning (PGT). PGT allows an instruction following policy to interact with the environment to collect several trajectories, which will be categorized into positive and negative samples based on preference. Then we use preference learning to fine-tune the initial goal latent representation with the categorized trajectories while keeping the policy backbone frozen. The experiment result shows that with minimal data and training, PGT achieves an average relative improvement of 72.0% and 81.6% over 17 tasks in 2 different foundation policies respectively, and outperforms the best human-selected instructions. Moreover, PGT surpasses full fine-tuning in the out-of-distribution (OOD) task-execution environments by 13.4%, indicating that our approach retains strong generalization capabilities. Since our approach stores a single latent representation for each task independently, it can be viewed as an efficient method for continual learning, without the risk of catastrophic forgetting or task interference. In short, PGT enhances the performance of agents across nearly all tasks in the Minecraft Skillforge benchmark and demonstrates robustness to the execution environment.

Kareem Ahmed, Stefano Teso, Paolo Morettin et al. · amazon-science

Structured output prediction problems are ubiquitous in machine learning. The prominent approach leverages neural networks as powerful feature extractors, otherwise assuming the independence of the outputs. These outputs, however, jointly encode an object, e.g. a path in a graph, and are therefore related through the structure underlying the output space. We discuss the semantic loss, which injects knowledge about such structure, defined symbolically, into training by minimizing the network's violation of such dependencies, steering the network towards predicting distributions satisfying the underlying structure. At the same time, it is agnostic to the arrangement of the symbols, and depends only on the semantics expressed thereby, while also enabling efficient end-to-end training and inference. We also discuss key improvements and applications of the semantic loss. One limitations of the semantic loss is that it does not exploit the association of every data point with certain features certifying its membership in a target class. We should therefore prefer minimum-entropy distributions over valid structures, which we obtain by additionally minimizing the neuro-symbolic entropy. We empirically demonstrate the benefits of this more refined formulation. Moreover, the semantic loss is designed to be modular and can be combined with both discriminative and generative neural models. This is illustrated by integrating it into generative adversarial networks, yielding constrained adversarial networks, a novel class of deep generative models able to efficiently synthesize complex objects obeying the structure of the underlying domain.

Xuejie Liu, Anji Liu, Guy Van den Broeck et al.

Masked generative models (MGMs) have emerged as a powerful framework for image synthesis, combining parallel decoding with strong bidirectional context modeling. However, generating high-quality samples typically requires many iterative decoding steps, resulting in high inference costs. A straightforward way to speed up generation is by decoding more tokens in each step, thereby reducing the total number of steps. However, when many tokens are decoded simultaneously, the model can only estimate the univariate marginal distributions independently, failing to capture the dependency among them. As a result, reducing the number of steps significantly compromises generation fidelity. In this work, we introduce ReCAP (Reused Context-Aware Prediction), a plug-and-play module that accelerates inference in MGMs by constructing low-cost steps via reusing feature embeddings from previously decoded context tokens. ReCAP interleaves standard full evaluations with lightweight steps that cache and reuse context features, substantially reducing computation while preserving the benefits of fine-grained, iterative generation. We demonstrate its effectiveness on top of three representative MGMs (MaskGIT, MAGE, and MAR), including both discrete and continuous token spaces and covering diverse architectural designs. In particular, on ImageNet256 class-conditional generation, ReCAP achieves up to 2.4x faster inference than the base model with minimal performance drop, and consistently delivers better efficiency-fidelity trade-offs under various generation settings.

Yi Zeng, Feifei Zhao, Yuwei Wang et al.

As Artificial Intelligence (AI) advances toward Artificial General Intelligence (AGI) and eventually Artificial Superintelligence (ASI), it may potentially surpass human control, deviate from human values, and even lead to irreversible catastrophic consequences in extreme cases. This looming risk underscores the critical importance of the "superalignment" problem - ensuring that AI systems which are much smarter than humans, remain aligned with human (compatible) intentions and values. While current scalable oversight and weak-to-strong generalization methods demonstrate certain applicability, they exhibit fundamental flaws in addressing the superalignment paradigm - notably, the unidirectional imposition of human values cannot accommodate superintelligence's autonomy or ensure AGI/ASI's stable learning. We contend that the values for sustainable symbiotic society should be co-shaped by humans and living AI together, achieving "Super Co-alignment." Guided by this vision, we propose a concrete framework that integrates external oversight and intrinsic proactive alignment. External oversight superalignment should be grounded in human-centered ultimate decision, supplemented by interpretable automated evaluation and correction, to achieve continuous alignment with humanity's evolving values. Intrinsic proactive superalignment is rooted in a profound understanding of the Self, others, and society, integrating self-awareness, self-reflection, and empathy to spontaneously infer human intentions, distinguishing good from evil and proactively prioritizing human well-being. The integration of externally-driven oversight with intrinsically-driven proactive alignment will co-shape symbiotic values and rules through iterative human-ASI co-alignment, paving the way for achieving safe and beneficial AGI and ASI for good, for human, and for a symbiotic ecology.

Haowei Lin, Xiangyu Wang, Ruilin Yan et al. · pku

With powerful large language models (LLMs) demonstrating superhuman reasoning capabilities, a critical question arises: Do LLMs genuinely reason, or do they merely recall answers from their extensive, web-scraped training datasets? Publicly released benchmarks inevitably become contaminated once incorporated into subsequent LLM training sets, undermining their reliability as faithful assessments. To address this, we introduce KUMO, a generative evaluation framework designed specifically for assessing reasoning in LLMs. KUMO synergistically combines LLMs with symbolic engines to dynamically produce diverse, multi-turn reasoning tasks that are partially observable and adjustable in difficulty. Through an automated pipeline, KUMO continuously generates novel tasks across open-ended domains, compelling models to demonstrate genuine generalization rather than memorization. We evaluated 23 state-of-the-art LLMs on 5,000 tasks across 100 domains created by KUMO, benchmarking their reasoning abilities against university students. Our findings reveal that many LLMs have outperformed university-level performance on easy reasoning tasks, and reasoning-scaled LLMs reach university-level performance on complex reasoning challenges. Moreover, LLM performance on KUMO tasks correlates strongly with results on newly released real-world reasoning benchmarks, underscoring KUMO's value as a robust, enduring assessment tool for genuine LLM reasoning capabilities.

Shuqi Lu, Haowei Lin, Lin Yao et al. · pku

3D structure modeling is essential across scales, enabling applications from fluid simulation and 3D reconstruction to protein folding and molecular docking. Yet, despite shared 3D spatial patterns, current approaches remain fragmented, with models narrowly specialized for specific domains and unable to generalize across tasks or scales. We propose Uni-3DAR, a unified autoregressive framework for cross-scale 3D generation and understanding. At its core is a coarse-to-fine tokenizer based on octree data structures, which compresses diverse 3D structures into compact 1D token sequences. We further propose a two-level subtree compression strategy, which reduces the octree token sequence by up to 8x. To address the challenge of dynamically varying token positions introduced by compression, we introduce a masked next-token prediction strategy that ensures accurate positional modeling, significantly boosting model performance. Extensive experiments across multiple 3D generation and understanding tasks, including small molecules, proteins, polymers, crystals, and macroscopic 3D objects, validate its effectiveness and versatility. Notably, Uni-3DAR surpasses previous state-of-the-art diffusion models by a substantial margin, achieving up to 256\% relative improvement while delivering inference speeds up to 21.8x faster.

Jingwen Deng, Zihao Wang, Shaofei Cai et al. · pku

Learning skills in open-world environments is essential for developing agents capable of handling a variety of tasks by combining basic skills. Online demonstration videos are typically long but unsegmented, making them difficult to segment and label with skill identifiers. Unlike existing methods that rely on sequence sampling or human labeling, we have developed a self-supervised learning-based approach to segment these long videos into a series of semantic-aware and skill-consistent segments. Drawing inspiration from human cognitive event segmentation theory, we introduce Skill Boundary Detection (SBD), an annotation-free temporal video segmentation algorithm. SBD detects skill boundaries in a video by leveraging prediction errors from a pretrained unconditional action-prediction model. This approach is based on the assumption that a significant increase in prediction error indicates a shift in the skill being executed. We evaluated our method in Minecraft, a rich open-world simulator with extensive gameplay videos available online. Our SBD-generated segments improved the average performance of conditioned policies by 63.7% and 52.1% on short-term atomic skill tasks, and their corresponding hierarchical agents by 11.3% and 20.8% on long-horizon tasks. Our method can leverage the diverse YouTube videos to train instruction-following agents. The project page can be found in https://craftjarvis.github.io/SkillDiscovery.

Haowei Lin, Shanda Li, Haotian Ye et al.

Given an unconditional generative model and a predictor for a target property (e.g., a classifier), the goal of training-free guidance is to generate samples with desirable target properties without additional training. As a highly efficient technique for steering generative models toward flexible outcomes, training-free guidance has gained increasing attention in diffusion models. However, existing methods only handle data in continuous spaces, while many scientific applications involve both continuous and discrete data (referred to as multimodality). Another emerging trend is the growing use of the simple and general flow matching framework in building generative foundation models, where guided generation remains under-explored. To address this, we introduce TFG-Flow, a novel training-free guidance method for multimodal generative flow. TFG-Flow addresses the curse-of-dimensionality while maintaining the property of unbiased sampling in guiding discrete variables. We validate TFG-Flow on four molecular design tasks and show that TFG-Flow has great potential in drug design by generating molecules with desired properties.

Yang Chen, Long Yang, Yitao Liang et al.

Low-Dimension-to-High-Dimension (LDHD) generalization is a special case of Out-of-Distribution (OOD) generalization, where the training data are restricted to a low-dimensional subspace of the high-dimensional testing space. Assuming that each instance is generated from a latent variable and the dimension of the latent variable reflects the problem scale, the inherent scaling challenge in length generalization can be captured by the LDHD generalization in the latent space. We theoretically demonstrate that LDHD generalization is generally unattainable without exploiting prior knowledge to provide appropriate inductive bias. Specifically, we explore LDHD generalization in Boolean functions. We verify that different architectures trained with (S)GD converge to \emph{min-degree interpolators w.r.t. different independent sets}. LDHD generalization is achievable if and only if the target function coincides with this inductive bias. Applying the insights from LDHD generalization to length generalization, we explain the effectiveness of CoT as changing the structure latent space to enable better LDHD generalization. We also propose a principle for position embedding design to handle both the inherent LDHD generalization and the nuisances such as the data format. Following the principle, we propose a novel position embedding called RPE-Square that remedies the RPE for dealing with the data format nuisance.

Jie Ying, Haowei Lin, Chao Yue et al.

In this study, we unveil a new AI model, termed PhyE2E, to discover physical formulas through symbolic regression. PhyE2E simplifies symbolic regression by decomposing it into sub-problems using the second-order derivatives of an oracle neural network, and employs a transformer model to translate data into symbolic formulas in an end-to-end manner. The resulting formulas are refined through Monte-Carlo Tree Search and Genetic Programming. We leverage a large language model to synthesize extensive symbolic expressions resembling real physics, and train the model to recover these formulas directly from data. A comprehensive evaluation reveals that PhyE2E outperforms existing state-of-the-art approaches, delivering superior symbolic accuracy, precision in data fitting, and consistency in physical units. We deployed PhyE2E to five applications in space physics, including the prediction of sunspot numbers, solar rotational angular velocity, emission line contribution functions, near-Earth plasma pressure, and lunar-tide plasma signals. The physical formulas generated by AI demonstrate a high degree of accuracy in fitting the experimental data from satellites and astronomical telescopes. We have successfully upgraded the formula proposed by NASA in 1993 regarding solar activity, and for the first time, provided the explanations for the long cycle of solar activity in an explicit form. We also found that the decay of near-Earth plasma pressure is proportional to r^2 to Earth, where subsequent mathematical derivations are consistent with satellite data from another independent study. Moreover, we found physical formulas that can describe the relationships between emission lines in the extreme ultraviolet spectrum of the Sun, temperatures, electron densities, and magnetic fields. The formula obtained is consistent with the properties that physicists had previously hypothesized it should possess.

Zihao Wang, Xujing Li, Yining Ye et al. · pku

We present Game-TARS, a generalist game agent trained with a unified, scalable action space anchored to human-aligned native keyboard-mouse inputs. Unlike API- or GUI-based approaches, this paradigm enables large-scale continual pre-training across heterogeneous domains, including OS, web, and simulation games. Game-TARS is pre-trained on over 500B tokens with diverse trajectories and multimodal data. Key techniques include a decaying continual loss to reduce causal confusion and an efficient Sparse-Thinking strategy that balances reasoning depth and inference cost. Experiments show that Game-TARS achieves about 2 times the success rate over the previous sota model on open-world Minecraft tasks, is close to the generality of fresh humans in unseen web 3d games, and outperforms GPT-5, Gemini-2.5-Pro, and Claude-4-Sonnet in FPS benchmarks. Scaling results on training-time and test-time confirm that the unified action space sustains improvements when scaled to cross-game and multimodal data. Our results demonstrate that simple, scalable action representations combined with large-scale pre-training provide a promising path toward generalist agents with broad computer-use abilities.

Yang Chen, Yitao Liang, Zhouchen Lin

In Transformers, Position Embeddings (PEs) significantly influence Length Generalization (LG) performance, yet their fundamental role remains unclear. In this work, we investigate the limitations and capabilities of PEs in achieving LG. We theoretically analyze PEs in Position-Only Linear Attentions (POLAs), introducing Linear Representation Complexity (LRC) to characterize when PEs enable LG. Our analysis shows that PEs do not expand computational capabilities but structure learned computations across positions. Extending to practical Transformers, we propose Sequential Representation Complexity (SRC) and conjecture that LG is possible if and only if SRC remains invariant across scales. We support this hypothesis with empirical evidence in various reasoning tasks. To enhance LG, we introduce Scale Hint, allowing flexible instance scaling, and a Learning-Based Position Embedding framework that automatically learns positional relations. Our work provides theoretical insights and practical strategies for improving LG in Transformers.

Shaofei Cai, Zhancun Mu, Haiwen Xia et al. · pku

While Reinforcement Learning (RL) has achieved remarkable success in language modeling, its triumph hasn't yet fully translated to visuomotor agents. A primary challenge in RL models is their tendency to overfit specific tasks or environments, thereby hindering the acquisition of generalizable behaviors across diverse settings. This paper provides a preliminary answer to this challenge by demonstrating that RL-finetuned visuomotor agents in Minecraft can achieve zero-shot generalization to unseen worlds. Specifically, we explore RL's potential to enhance generalizable spatial reasoning and interaction capabilities in 3D worlds. To address challenges in multi-task RL representation, we analyze and establish cross-view goal specification as a unified multi-task goal space for visuomotor policies. Furthermore, to overcome the significant bottleneck of manual task design, we propose automated task synthesis within the highly customizable Minecraft environment for large-scale multi-task RL training, and we construct an efficient distributed RL framework to support this. Experimental results show RL significantly boosts interaction success rates by $4\times$ and enables zero-shot generalization of spatial reasoning across diverse environments, including real-world settings. Our findings underscore the immense potential of RL training in 3D simulated environments, especially those amenable to large-scale task generation, for significantly advancing visuomotor agents' spatial reasoning.