Jinxin Liu

Jifan Yu, Xiaozhi Wang, Shangqing Tu et al. · tsinghua

The unprecedented performance of large language models (LLMs) necessitates improvements in evaluations. Rather than merely exploring the breadth of LLM abilities, we believe meticulous and thoughtful designs are essential to thorough, unbiased, and applicable evaluations. Given the importance of world knowledge to LLMs, we construct a Knowledge-oriented LLM Assessment benchmark (KoLA), in which we carefully design three crucial factors: (1) For \textbf{ability modeling}, we mimic human cognition to form a four-level taxonomy of knowledge-related abilities, covering $19$ tasks. (2) For \textbf{data}, to ensure fair comparisons, we use both Wikipedia, a corpus prevalently pre-trained by LLMs, along with continuously collected emerging corpora, aiming to evaluate the capacity to handle unseen data and evolving knowledge. (3) For \textbf{evaluation criteria}, we adopt a contrastive system, including overall standard scores for better numerical comparability across tasks and models and a unique self-contrast metric for automatically evaluating knowledge-creating ability. We evaluate $28$ open-source and commercial LLMs and obtain some intriguing findings. The KoLA dataset and open-participation leaderboard are publicly released at https://kola.xlore.cn and will be continuously updated to provide references for developing LLMs and knowledge-related systems.

GLM-5 Team, Aohan Zeng, Xin Lv et al. · tsinghua

We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.

Yao Lai, Jinxin Liu, Zhentao Tang et al. · tsinghua

Placement is a critical step in modern chip design, aiming to determine the positions of circuit modules on the chip canvas. Recent works have shown that reinforcement learning (RL) can improve human performance in chip placement. However, such an RL-based approach suffers from long training time and low transfer ability in unseen chip circuits. To resolve these challenges, we cast the chip placement as an offline RL formulation and present ChiPFormer that enables learning a transferable placement policy from fixed offline data. ChiPFormer has several advantages that prior arts do not have. First, ChiPFormer can exploit offline placement designs to learn transferable policies more efficiently in a multi-task setting. Second, ChiPFormer can promote effective finetuning for unseen chip circuits, reducing the placement runtime from hours to minutes. Third, extensive experiments on 32 chip circuits demonstrate that ChiPFormer achieves significantly better placement quality while reducing the runtime by 10x compared to recent state-of-the-art approaches in both public benchmarks and realistic industrial tasks. The deliverables are released at https://sites.google.com/view/chipformer/home.

Zifeng Zhuang, Kun Lei, Jinxin Liu et al.

Offline reinforcement learning (RL) is a challenging setting where existing off-policy actor-critic methods perform poorly due to the overestimation of out-of-distribution state-action pairs. Thus, various additional augmentations are proposed to keep the learned policy close to the offline dataset (or the behavior policy). In this work, starting from the analysis of offline monotonic policy improvement, we get a surprising finding that some online on-policy algorithms are naturally able to solve offline RL. Specifically, the inherent conservatism of these on-policy algorithms is exactly what the offline RL method needs to overcome the overestimation. Based on this, we propose Behavior Proximal Policy Optimization (BPPO), which solves offline RL without any extra constraint or regularization introduced compared to PPO. Extensive experiments on the D4RL benchmark indicate this extremely succinct method outperforms state-of-the-art offline RL algorithms. Our implementation is available at https://github.com/Dragon-Zhuang/BPPO.

Chaoyi Wu, Pengcheng Qiu, Jinxin Liu et al.

In this study, we present MedS-Bench, a comprehensive benchmark designed to evaluate the performance of large language models (LLMs) in clinical contexts. Unlike existing benchmarks that focus on multiple-choice question answering, MedS-Bench spans 11 high-level clinical tasks, including clinical report summarization, treatment recommendations, diagnosis, named entity recognition, and medical concept explanation, among others. We evaluated six leading LLMs, e.g., MEDITRON, Mistral, InternLM 2, Llama 3, GPT-4, and Claude-3.5 using few-shot prompting, and found that even the most sophisticated models struggle with these complex tasks. To address these limitations, we developed MedS-Ins, a large-scale instruction tuning dataset for medicine. MedS-Ins comprises 58 medically oriented language corpora, totaling 13.5 million samples across 122 tasks. To demonstrate the dataset's utility, we conducted a proof-of-concept experiment by performing instruction tuning on a lightweight, open-source medical language model. The resulting model, MMedIns-Llama 3, significantly outperformed existing models across nearly all clinical tasks. To promote further advancements in the application of LLMs to clinical challenges, we have made the MedS-Ins dataset fully accessible and invite the research community to contribute to its expansion.Additionally, we have launched a dynamic leaderboard for MedS-Bench, which we plan to regularly update the test set to track progress and enhance the adaptation of general LLMs to the medical domain. Leaderboard: https://henrychur.github.io/MedS-Bench/. Github: https://github.com/MAGIC-AI4Med/MedS-Ins.

Zhiyan Chen, Jinxin Liu, Yu Shen et al.

Despite its technological benefits, Internet of Things (IoT) has cyber weaknesses due to the vulnerabilities in the wireless medium. Machine learning (ML)-based methods are widely used against cyber threats in IoT networks with promising performance. Advanced persistent threat (APT) is prominent for cybercriminals to compromise networks, and it is crucial to long-term and harmful characteristics. However, it is difficult to apply ML-based approaches to identify APT attacks to obtain a promising detection performance due to an extremely small percentage among normal traffic. There are limited surveys to fully investigate APT attacks in IoT networks due to the lack of public datasets with all types of APT attacks. It is worth to bridge the state-of-the-art in network attack detection with APT attack detection in a comprehensive review article. This survey article reviews the security challenges in IoT networks and presents the well-known attacks, APT attacks, and threat models in IoT systems. Meanwhile, signature-based, anomaly-based, and hybrid intrusion detection systems are summarized for IoT networks. The article highlights statistical insights regarding frequently applied ML-based methods against network intrusion alongside the number of attacks types detected. Finally, open issues and challenges for common network intrusion and APT attacks are presented for future research.

Yudan Jiang, Xiao Tang, Jinxin Liu et al.

The development of 6G networks brings an increasing variety of data services, which motivates the hybrid computation paradigm that coordinates the over-the-air computation (AirComp) and edge computing for diverse and effective data processing. In this paper, we address this emerging issue of hybrid data computation from an energy-efficiency perspective, where the coexistence of both types induces resource competition and interference, and thus complicates the network management. Accordingly, we formulate the problem to minimize the overall energy consumption including the data transmission and computation, subject to the offloading capacity and aggregation accuracy. We then propose a block coordinate descent framework that decomposes and solves the subproblems including the user scheduling, power control, and transceiver scaling, which are then iterated towards a coordinated hybrid computation solution. Simulation results confirm that our coordinated approach achieves significant energy savings compared to baseline strategies, demonstrating its effectiveness in creating a well-coordinated and sustainable hybrid computing environment.

Bowei He, Zexu Sun, Jinxin Liu et al.

In offline imitation learning (IL), an agent aims to learn an optimal expert behavior policy without additional online environment interactions. However, in many real-world scenarios, such as robotics manipulation, the offline dataset is collected from suboptimal behaviors without rewards. Due to the scarce expert data, the agents usually suffer from simply memorizing poor trajectories and are vulnerable to variations in the environments, lacking the capability of generalizing to new environments. To automatically generate high-quality expert data and improve the generalization ability of the agent, we propose a framework named \underline{O}ffline \underline{I}mitation \underline{L}earning with \underline{C}ounterfactual data \underline{A}ugmentation (OILCA) by doing counterfactual inference. In particular, we leverage identifiable variational autoencoder to generate \textit{counterfactual} samples for expert data augmentation. We theoretically analyze the influence of the generated expert data and the improvement of generalization. Moreover, we conduct extensive experiments to demonstrate that our approach significantly outperforms various baselines on both \textsc{DeepMind Control Suite} benchmark for in-distribution performance and \textsc{CausalWorld} benchmark for out-of-distribution generalization. Our code is available at \url{https://github.com/ZexuSun/OILCA-NeurIPS23}.

Zhaole Wang, Xiao Tang, Naijin Liu et al.

Satellite communications are envisioned as a key enabler for ubiquitous coverage in future 6G networks, yet the broadcast nature renders them vulnerable to eavesdropping, especially given the long-distance transmissions and associated high uncertainties. In this paper, we propose the physical layer security enhancement for multi-beam satellite communications with the assistance of an aerial reconfigurable intelligent surface (ARIS). Considering the high dynamics and uncertainties of channels, we characterize the channel distribution with moment-based ambiguity sets. Accordingly, a distributionally robust secrecy rate optimization is formulated through joint design of transmit and reflection beamforming. We then introduce a conditional value-at-risk-based reformulation to convert the probabilistic constraints into deterministic forms. An alternating optimization framework is subsequently employed to iteratively update the transmit and reflective beamforming vectors until convergence. Simulation results demonstrate that the proposed distributionally robust scheme significantly enhances secrecy performance, and maintains reliable performance across various channel error distributions.

Jinxin Liu, Hongyin Zhang, Donglin Wang

Offline reinforcement learning algorithms promise to be applicable in settings where a fixed dataset is available and no new experience can be acquired. However, such formulation is inevitably offline-data-hungry and, in practice, collecting a large offline dataset for one specific task over one specific environment is also costly and laborious. In this paper, we thus 1) formulate the offline dynamics adaptation by using (source) offline data collected from another dynamics to relax the requirement for the extensive (target) offline data, 2) characterize the dynamics shift problem in which prior offline methods do not scale well, and 3) derive a simple dynamics-aware reward augmentation (DARA) framework from both model-free and model-based offline settings. Specifically, DARA emphasizes learning from those source transition pairs that are adaptive for the target environment and mitigates the offline dynamics shift by characterizing state-action-next-state pairs instead of the typical state-action distribution sketched by prior offline RL methods. The experimental evaluation demonstrates that DARA, by augmenting rewards in the source offline dataset, can acquire an adaptive policy for the target environment and yet significantly reduce the requirement of target offline data. With only modest amounts of target offline data, our performance consistently outperforms the prior offline RL methods in both simulated and real-world tasks.

Jinxin Liu, Li He, Yachen Kang et al.

In this paper, we present \textbf{C}ont\textbf{E}xtual \textbf{I}mitation \textbf{L}earning~(CEIL), a general and broadly applicable algorithm for imitation learning (IL). Inspired by the formulation of hindsight information matching, we derive CEIL by explicitly learning a hindsight embedding function together with a contextual policy using the hindsight embeddings. To achieve the expert matching objective for IL, we advocate for optimizing a contextual variable such that it biases the contextual policy towards mimicking expert behaviors. Beyond the typical learning from demonstrations (LfD) setting, CEIL is a generalist that can be effectively applied to multiple settings including: 1)~learning from observations (LfO), 2)~offline IL, 3)~cross-domain IL (mismatched experts), and 4) one-shot IL settings. Empirically, we evaluate CEIL on the popular MuJoCo tasks (online) and the D4RL dataset (offline). Compared to prior state-of-the-art baselines, we show that CEIL is more sample-efficient in most online IL tasks and achieves better or competitive performances in offline tasks.

Jinxin Liu, Ziqi Zhang, Zhenyu Wei et al.

Offline reinforcement learning (RL) aims to learn a policy using only pre-collected and fixed data. Although avoiding the time-consuming online interactions in RL, it poses challenges for out-of-distribution (OOD) state actions and often suffers from data inefficiency for training. Despite many efforts being devoted to addressing OOD state actions, the latter (data inefficiency) receives little attention in offline RL. To address this, this paper proposes the cross-domain offline RL, which assumes offline data incorporate additional source-domain data from varying transition dynamics (environments), and expects it to contribute to the offline data efficiency. To do so, we identify a new challenge of OOD transition dynamics, beyond the common OOD state actions issue, when utilizing cross-domain offline data. Then, we propose our method BOSA, which employs two support-constrained objectives to address the above OOD issues. Through extensive experiments in the cross-domain offline RL setting, we demonstrate BOSA can greatly improve offline data efficiency: using only 10\% of the target data, BOSA could achieve {74.4\%} of the SOTA offline RL performance that uses 100\% of the target data. Additionally, we also show BOSA can be effortlessly plugged into model-based offline RL and noising data augmentation techniques (used for generating source-domain data), which naturally avoids the potential dynamics mismatch between target-domain data and newly generated source-domain data.

Jinxin Liu, Lipeng Zu, Li He et al.

Offline reinforcement learning (RL) aims to learn an optimal policy from pre-collected and labeled datasets, which eliminates the time-consuming data collection in online RL. However, offline RL still bears a large burden of specifying/handcrafting extrinsic rewards for each transition in the offline data. As a remedy for the labor-intensive labeling, we propose to endow offline RL tasks with a few expert data and utilize the limited expert data to drive intrinsic rewards, thus eliminating the need for extrinsic rewards. To achieve that, we introduce \textbf{C}alibrated \textbf{L}atent g\textbf{U}idanc\textbf{E} (CLUE), which utilizes a conditional variational auto-encoder to learn a latent space such that intrinsic rewards can be directly qualified over the latent space. CLUE's key idea is to align the intrinsic rewards consistent with the expert intention via enforcing the embeddings of expert data to a calibrated contextual representation. We instantiate the expert-driven intrinsic rewards in sparse-reward offline RL tasks, offline imitation learning (IL) tasks, and unsupervised offline RL tasks. Empirically, we find that CLUE can effectively improve the sparse-reward offline RL performance, outperform the state-of-the-art offline IL baselines, and discover diverse skills from static reward-free offline data.

Jinxin Liu, Hongyin Zhang, Zifeng Zhuang et al.

In this work, we decouple the iterative bi-level offline RL (value estimation and policy extraction) from the offline training phase, forming a non-iterative bi-level paradigm and avoiding the iterative error propagation over two levels. Specifically, this non-iterative paradigm allows us to conduct inner-level optimization (value estimation) in training, while performing outer-level optimization (policy extraction) in testing. Naturally, such a paradigm raises three core questions that are not fully answered by prior non-iterative offline RL counterparts like reward-conditioned policy: (q1) What information should we transfer from the inner-level to the outer-level? (q2) What should we pay attention to when exploiting the transferred information for safe/confident outer-level optimization? (q3) What are the benefits of concurrently conducting outer-level optimization during testing? Motivated by model-based optimization (MBO), we propose DROP (design from policies), which fully answers the above questions. Specifically, in the inner-level, DROP decomposes offline data into multiple subsets, and learns an MBO score model (a1). To keep safe exploitation to the score model in the outer-level, we explicitly learn a behavior embedding and introduce a conservative regularization (a2). During testing, we show that DROP permits deployment adaptation, enabling an adaptive inference across states (a3). Empirically, we evaluate DROP on various tasks, showing that DROP gains comparable or better performance compared to prior methods.

Ji Qi, Bin Xu, Kaisheng Zeng et al.

Document-level relation extraction with graph neural networks faces a fundamental graph construction gap between training and inference - the golden graph structure only available during training, which causes that most methods adopt heuristic or syntactic rules to construct a prior graph as a pseudo proxy. In this paper, we propose $\textbf{ConstGCN}$, a novel graph convolutional network which performs knowledge-based information propagation between entities along with all specific relation spaces without any prior graph construction. Specifically, it updates the entity representation by aggregating information from all other entities along with each relation space, thus modeling the relation-aware spatial information. To control the information flow passing through the indeterminate relation spaces, we propose to constrain the propagation using transmitting scores learned from the Noise Contrastive Estimation between fact triples. Experimental results show that our method outperforms the previous state-of-the-art (SOTA) approaches on the DocRE dataset.

GLM-4. 5 Team, Aohan Zeng, Xin Lv et al.

We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks. We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters), to advance research in reasoning and agentic AI systems. Code, models, and more information are available at https://github.com/zai-org/GLM-4.5.

Yachen Kang, Li He, Jinxin Liu et al.

Preference-based reinforcement learning (PbRL) promises to learn a complex reward function with binary human preference. However, such human-in-the-loop formulation requires considerable human effort to assign preference labels to segment pairs, hindering its large-scale applications. Recent approache has tried to reuse unlabeled segments, which implicitly elucidates the distribution of segments and thereby alleviates the human effort. And consistency regularization is further considered to improve the performance of semi-supervised learning. However, we notice that, unlike general classification tasks, in PbRL there exits a unique phenomenon that we defined as similarity trap in this paper. Intuitively, human can have diametrically opposite preferredness for similar segment pairs, but such similarity may trap consistency regularization fail in PbRL. Due to the existence of similarity trap, such consistency regularization improperly enhances the consistency possiblity of the model's predictions between segment pairs, and thus reduces the confidence in reward learning, since the augmented distribution does not match with the original one in PbRL. To overcome such issue, we present a self-training method along with our proposed peer regularization, which penalizes the reward model memorizing uninformative labels and acquires confident predictions. Empirically, we demonstrate that our approach is capable of learning well a variety of locomotion and robotic manipulation behaviors using different semi-supervised alternatives and peer regularization.

Zifeng Zhuang, Dengyun Peng, Jinxin Liu et al.

As a data-driven paradigm, offline reinforcement learning (RL) has been formulated as sequence modeling that conditions on the hindsight information including returns, goal or future trajectory. Although promising, this supervised paradigm overlooks the core objective of RL that maximizes the return. This overlook directly leads to the lack of trajectory stitching capability that affects the sequence model learning from sub-optimal data. In this work, we introduce the concept of max-return sequence modeling which integrates the goal of maximizing returns into existing sequence models. We propose Reinforced Transformer (Reinformer), indicating the sequence model is reinforced by the RL objective. Reinformer additionally incorporates the objective of maximizing returns in the training phase, aiming to predict the maximum future return within the distribution. During inference, this in-distribution maximum return will guide the selection of optimal actions. Empirically, Reinformer is competitive with classical RL methods on the D4RL benchmark and outperforms state-of-the-art sequence model particularly in trajectory stitching ability. Code is public at https://github.com/Dragon-Zhuang/Reinformer.

Yao Lai, Jinxin Liu, David Z. Pan et al. · tsinghua

Across a wide range of hardware scenarios, the computational efficiency and physical size of the arithmetic units significantly influence the speed and footprint of the overall hardware system. Nevertheless, the effectiveness of prior arithmetic design techniques proves inadequate, as it does not sufficiently optimize speed and area, resulting in a reduced processing rate and larger module size. To boost the arithmetic performance, in this work, we focus on the two most common and fundamental arithmetic modules: adders and multipliers. We cast the design tasks as single-player tree generation games, leveraging reinforcement learning techniques to optimize their arithmetic tree structures. Such a tree generation formulation allows us to efficiently navigate the vast search space and discover superior arithmetic designs that improve computational efficiency and hardware size within just a few hours. For adders, our approach discovers designs of 128-bit adders that achieve Pareto optimality in theoretical metrics. Compared with the state-of-the-art PrefixRL, our method decreases computational delay and hardware size by up to 26% and 30%, respectively. For multipliers, when compared to RL-MUL, our approach increases speed and reduces size by as much as 49% and 45%. Moreover, the inherent flexibility and scalability of our method enable us to deploy our designs into cutting-edge technologies, as we show that they can be seamlessly integrated into 7nm technology. We believe our work will offer valuable insights into hardware design, further accelerating speed and reducing size through the refined search space and our tree generation methodologies. See our introduction video at https://bit.ly/ArithmeticTree. Codes are released at https://github.com/laiyao1/ArithmeticTree.

Zhidian Huang, Zijun Yao, Ji Qi et al.

Recent advances in multimodal large language models (MLLMs) mark a shift from non-thinking models to post-trained reasoning models capable of solving complex problems through thinking. However, whether such thinking mitigates hallucinations in multimodal perception and reasoning remains unclear. Self-reflective reasoning enhances robustness but introduces additional hallucinations, and subtle perceptual errors still result in incorrect or coincidentally correct answers. Existing benchmarks primarily focus on models before the emergence of reasoning MLLMs, neglecting the internal thinking process and failing to measure the hallucinations that occur during thinking. To address these challenges, we introduce MM-THEBench, a comprehensive benchmark for assessing hallucinations of intermediate CoTs in reasoning MLLMs. MM-THEBench features a fine-grained taxonomy grounded in cognitive dimensions, diverse data with verified reasoning annotations, and a multi-level automated evaluation framework. Extensive experiments on mainstream reasoning MLLMs reveal insights into how thinking affects hallucination and reasoning capability in various multimodal tasks.

Ziqi Zhang, Xiao Xiong, Zifeng Zhuang et al.

Studying how to fine-tune offline reinforcement learning (RL) pre-trained policy is profoundly significant for enhancing the sample efficiency of RL algorithms. However, directly fine-tuning pre-trained policies often results in sub-optimal performance. This is primarily due to the distribution shift between offline pre-training and online fine-tuning stages. Specifically, the distribution shift limits the acquisition of effective online samples, ultimately impacting the online fine-tuning performance. In order to narrow down the distribution shift between offline and online stages, we proposed Q conditioned state entropy (QCSE) as intrinsic reward. Specifically, QCSE maximizes the state entropy of all samples individually, considering their respective Q values. This approach encourages exploration of low-frequency samples while penalizing high-frequency ones, and implicitly achieves State Marginal Matching (SMM), thereby ensuring optimal performance, solving the asymptotic sub-optimality of constraint-based approaches. Additionally, QCSE can seamlessly integrate into various RL algorithms, enhancing online fine-tuning performance. To validate our claim, we conduct extensive experiments, and observe significant improvements with QCSE (about 13% for CQL and 8% for Cal-QL). Furthermore, we extended experimental tests to other algorithms, affirming the generality of QCSE.

Amy Xin, Jinxin Liu, Zijun Yao et al.

Despite the outstanding capabilities of large language models (LLMs), knowledge-intensive reasoning still remains a challenging task due to LLMs' limitations in compositional reasoning and the hallucination problem. A prevalent solution is to employ chain-of-thought (CoT) with retrieval-augmented generation (RAG), which first formulates a reasoning plan by decomposing complex questions into simpler sub-questions, and then applies iterative RAG at each sub-question. However, prior works exhibit two crucial problems: inadequate reasoning planning and poor incorporation of heterogeneous knowledge. In this paper, we introduce AtomR, a framework for LLMs to conduct accurate heterogeneous knowledge reasoning at the atomic level. Inspired by how knowledge graph query languages model compositional reasoning through combining predefined operations, we propose three atomic knowledge operators, a unified set of operators for LLMs to retrieve and manipulate knowledge from heterogeneous sources. First, in the reasoning planning stage, AtomR decomposes a complex question into a reasoning tree where each leaf node corresponds to an atomic knowledge operator, achieving question decomposition that is highly fine-grained and orthogonal. Subsequently, in the reasoning execution stage, AtomR executes each atomic knowledge operator, which flexibly selects, retrieves, and operates atomic level knowledge from heterogeneous sources. We also introduce BlendQA, a challenging benchmark specially tailored for heterogeneous knowledge reasoning. Experiments on three single-source and two multi-source datasets show that AtomR outperforms state-of-the-art baselines by a large margin, with F1 score improvements of 9.4% on 2WikiMultihop and 9.5% on BlendQA. We release our code and datasets.

Jinxin Liu, Zao Yang

The responses generated by Large Language Models (LLMs) can include sensitive information from individuals and organizations, leading to potential privacy leakage. This work implements Influence Functions (IFs) to trace privacy leakage back to the training data, thereby mitigating privacy concerns of Language Models (LMs). However, we notice that current IFs struggle to accurately estimate the influence of tokens with large gradient norms, potentially overestimating their influence. When tracing the most influential samples, this leads to frequently tracing back to samples with large gradient norm tokens, overshadowing the actual most influential samples even if their influences are well estimated. To address this issue, we propose Heuristically Adjusted IF (HAIF), which reduces the weight of tokens with large gradient norms, thereby significantly improving the accuracy of tracing the most influential samples. To establish easily obtained groundtruth for tracing privacy leakage, we construct two datasets, PII-E and PII-CR, representing two distinct scenarios: one with identical text in the model outputs and pre-training data, and the other where models leverage their reasoning abilities to generate text divergent from pre-training data. HAIF significantly improves tracing accuracy, enhancing it by 20.96% to 73.71% on the PII-E dataset and 3.21% to 45.93% on the PII-CR dataset, compared to the best SOTA IFs against various GPT-2 and QWen-1.5 models. HAIF also outperforms SOTA IFs on real-world pretraining data CLUECorpus2020, demonstrating strong robustness regardless prompt and response lengths.

Ziqi Zhang, Jingzehua Xu, Zifeng Zhuang et al.

Proximal Policy Optimization (PPO) has been broadly applied to robotics learning, showcasing stable training performance. However, the fixed clipping bound setting may limit the performance of PPO. Specifically, there is no theoretical proof that the optimal clipping bound remains consistent throughout the entire training process. Meanwhile, previous researches suggest that a fixed clipping bound restricts the policy's ability to explore. Therefore, many past studies have aimed to dynamically adjust the PPO clipping bound to enhance PPO's performance. However, the objective of these approaches are not directly aligned with the objective of reinforcement learning (RL) tasks, which is to maximize the cumulative Return. Unlike previous clipping approaches, we propose a bi-level proximal policy optimization objective that can dynamically adjust the clipping bound to better reflect the preference (maximizing Return) of these RL tasks. Based on this bi-level proximal policy optimization paradigm, we introduce a new algorithm named Preference based Proximal Policy Optimization (Pb-PPO). Pb-PPO utilizes a multi-armed bandit approach to refelect RL preference, recommending the clipping bound for PPO that can maximizes the current Return. Therefore, Pb-PPO results in greater stability and improved performance compared to PPO with a fixed clipping bound. We test Pb-PPO on locomotion benchmarks across multiple environments, including Gym-Mujoco and legged-gym. Additionally, we validate Pb-PPO on customized navigation tasks. Meanwhile, we conducted comparisons with PPO using various fixed clipping bounds and various of clipping approaches. The experimental results indicate that Pb-PPO demonstrates superior training performance compared to PPO and its variants. Our codebase has been released at : https://github.com/stevezhangzA/pb_ppo

Guanzhong He, Zhen Yang, Jinxin Liu et al.

Search agents have achieved significant advancements in enabling intelligent information retrieval and decision-making within interactive environments. Although reinforcement learning has been employed to train agentic models capable of more dynamic interactive retrieval, existing methods are limited by shallow tool-use depth and the accumulation of errors over multiple iterative interactions. In this paper, we present WebSeer, a more intelligent search agent trained via reinforcement learning enhanced with a self-reflection mechanism. Specifically, we construct a large dataset annotated with reflection patterns and design a two-stage training framework that unifies cold start and reinforcement learning within the self-reflection paradigm for real-world web-based environments, which enables the model to generate longer and more reflective tool-use trajectories. Our approach substantially extends tool-use chains and improves answer accuracy. Using a single 14B model, we achieve state-of-the-art results on HotpotQA and SimpleQA, with accuracies of 72.3% and 90.0%, respectively, and demonstrate strong generalization to out-of-distribution datasets. The code is available at https://github.com/99hgz/WebSeer

Yachen Kang, Diyuan Shi, Jinxin Liu et al.

This study focuses on the topic of offline preference-based reinforcement learning (PbRL), a variant of conventional reinforcement learning that dispenses with the need for online interaction or specification of reward functions. Instead, the agent is provided with fixed offline trajectories and human preferences between pairs of trajectories to extract the dynamics and task information, respectively. Since the dynamics and task information are orthogonal, a naive approach would involve using preference-based reward learning followed by an off-the-shelf offline RL algorithm. However, this requires the separate learning of a scalar reward function, which is assumed to be an information bottleneck of the learning process. To address this issue, we propose the offline preference-guided policy optimization (OPPO) paradigm, which models offline trajectories and preferences in a one-step process, eliminating the need for separately learning a reward function. OPPO achieves this by introducing an offline hindsight information matching objective for optimizing a contextual policy and a preference modeling objective for finding the optimal context. OPPO further integrates a well-performing decision policy by optimizing the two objectives iteratively. Our empirical results demonstrate that OPPO effectively models offline preferences and outperforms prior competing baselines, including offline RL algorithms performed over either true or pseudo reward function specifications. Our code is available on the project website: https://sites.google.com/view/oppo-icml-2023 .

Ji Qi, Chuchun Zhang, Xiaozhi Wang et al.

The robustness to distribution changes ensures that NLP models can be successfully applied in the realistic world, especially for information extraction tasks. However, most prior evaluation benchmarks have been devoted to validating pairwise matching correctness, ignoring the crucial measurement of robustness. In this paper, we present the first benchmark that simulates the evaluation of open information extraction models in the real world, where the syntactic and expressive distributions under the same knowledge meaning may drift variously. We design and annotate a large-scale testbed in which each example is a knowledge-invariant clique that consists of sentences with structured knowledge of the same meaning but with different syntactic and expressive forms. By further elaborating the robustness metric, a model is judged to be robust if its performance is consistently accurate on the overall cliques. We perform experiments on typical models published in the last decade as well as a popular large language model, the results show that the existing successful models exhibit a frustrating degradation, with a maximum drop of 23.43 F1 score. Our resources and code are available at https://github.com/qijimrc/ROBUST.

Pengxiang Ding, Jianfei Ma, Xinyang Tong et al.

This paper addresses the limitations of current humanoid robot control frameworks, which primarily rely on reactive mechanisms and lack autonomous interaction capabilities due to data scarcity. We propose Humanoid-VLA, a novel framework that integrates language understanding, egocentric scene perception, and motion control, enabling universal humanoid control. Humanoid-VLA begins with language-motion pre-alignment using non-egocentric human motion datasets paired with textual descriptions, allowing the model to learn universal motion patterns and action semantics. We then incorporate egocentric visual context through a parameter efficient video-conditioned fine-tuning, enabling context-aware motion generation. Furthermore, we introduce a self-supervised data augmentation strategy that automatically generates pseudoannotations directly derived from motion data. This process converts raw motion sequences into informative question-answer pairs, facilitating the effective use of large-scale unlabeled video data. Built upon whole-body control architectures, extensive experiments show that Humanoid-VLA achieves object interaction and environment exploration tasks with enhanced contextual awareness, demonstrating a more human-like capacity for adaptive and intelligent engagement.

Mohammad Ghazi Vakili, Christoph Gorgulla, AkshatKumar Nigam et al.

The discovery of small molecules with therapeutic potential is a long-standing challenge in chemistry and biology. Researchers have increasingly leveraged novel computational techniques to streamline the drug development process to increase hit rates and reduce the costs associated with bringing a drug to market. To this end, we introduce a quantum-classical generative model that seamlessly integrates the computational power of quantum algorithms trained on a 16-qubit IBM quantum computer with the established reliability of classical methods for designing small molecules. Our hybrid generative model was applied to designing new KRAS inhibitors, a crucial target in cancer therapy. We synthesized 15 promising molecules during our investigation and subjected them to experimental testing to assess their ability to engage with the target. Notably, among these candidates, two molecules, ISM061-018-2 and ISM061-22, each featuring unique scaffolds, stood out by demonstrating effective engagement with KRAS. ISM061-018-2 was identified as a broad-spectrum KRAS inhibitor, exhibiting a binding affinity to KRAS-G12D at $1.4 μM$. Concurrently, ISM061-22 exhibited specific mutant selectivity, displaying heightened activity against KRAS G12R and Q61H mutants. To our knowledge, this work shows for the first time the use of a quantum-generative model to yield experimentally confirmed biological hits, showcasing the practical potential of quantum-assisted drug discovery to produce viable therapeutics. Moreover, our findings reveal that the efficacy of distribution learning correlates with the number of qubits utilized, underlining the scalability potential of quantum computing resources. Overall, we anticipate our results to be a stepping stone towards developing more advanced quantum generative models in drug discovery.

Zhicheng Lee, Shulin Cao, Jinxin Liu et al.

Large Reasoning Models (LRMs) exhibit remarkable reasoning abilities but rely primarily on parametric knowledge, limiting factual accuracy. While recent works equip reinforcement learning (RL)-based LRMs with retrieval capabilities, they suffer from overthinking and lack robustness in reasoning, reducing their effectiveness in question answering (QA) tasks. To address this, we propose ReaRAG, a factuality-enhanced reasoning model that explores diverse queries without excessive iterations. Our solution includes a novel data construction framework with an upper bound on the reasoning chain length. Specifically, we first leverage an LRM to generate deliberate thinking, then select an action from a predefined action space (Search and Finish). For Search action, a query is executed against the RAG engine, where the result is returned as observation to guide reasoning steps later. This process iterates until a Finish action is chosen. Benefiting from ReaRAG's strong reasoning capabilities, our approach outperforms existing baselines on multi-hop QA. Further analysis highlights its strong reflective ability to recognize errors and refine its reasoning trajectory. Our study enhances LRMs' factuality while effectively integrating robust reasoning for Retrieval-Augmented Generation (RAG).

Jinxin Liu, Shulin Cao, Jiaxin Shi et al. · tsinghua

Knowledge Base Question Answering (KBQA) aims to answer natural language questions based on facts in knowledge bases. A typical approach to KBQA is semantic parsing, which translates a question into an executable logical form in a formal language. Recent works leverage the capabilities of large language models (LLMs) for logical form generation to improve performance. However, although it is validated that LLMs are capable of solving some KBQA problems, there has been little discussion on the differences in LLMs' proficiency in formal languages used in semantic parsing. In this work, we propose to evaluate the understanding and generation ability of LLMs to deal with differently structured logical forms by examining the inter-conversion of natural and formal language through in-context learning of LLMs. Extensive experiments with models of different sizes show that state-of-the-art LLMs can understand formal languages as well as humans, but generating correct logical forms given a few examples remains a challenge. Most importantly, our results also indicate that LLMs exhibit considerable sensitivity. In general, the formal language with a lower formalization level, i.e., the more similar it is to natural language, is more friendly to LLMs.

Jiaran Ye, Zijun Yao, Zhidian Huang et al. · tsinghua

Recent work suggests that large language models (LLMs) can perform multi-hop reasoning implicitly -- producing correct answers without explicitly verbalizing intermediate steps -- but the underlying mechanisms remain poorly understood. In this paper, we study how such implicit reasoning emerges by training transformers from scratch in a controlled symbolic environment. Our analysis reveals a three-stage developmental trajectory: early memorization, followed by in-distribution generalization, and eventually cross-distribution generalization. We find that training with atomic triples is not necessary but accelerates learning, and that second-hop generalization relies on query-level exposure to specific compositional structures. To interpret these behaviors, we introduce two diagnostic tools: cross-query semantic patching, which identifies semantically reusable intermediate representations, and a cosine-based representational lens, which reveals that successful reasoning correlates with the cosine-base clustering in hidden space. This clustering phenomenon in turn provides a coherent explanation for the behavioral dynamics observed across training, linking representational structure to reasoning capability. These findings provide new insights into the interpretability of implicit multi-hop reasoning in LLMs, helping to clarify how complex reasoning processes unfold internally and offering pathways to enhance the transparency of such models.

Ziqi Zhang, Jingzehua Xu, Jinxin Liu et al.

Offline reinforcement learning (RL) algorithms can learn better decision-making compared to behavior policies by stitching the suboptimal trajectories to derive more optimal ones. Meanwhile, Decision Transformer (DT) abstracts the RL as sequence modeling, showcasing competitive performance on offline RL benchmarks. However, recent studies demonstrate that DT lacks of stitching capacity, thus exploiting stitching capability for DT is vital to further improve its performance. In order to endow stitching capability to DT, we abstract trajectory stitching as expert matching and introduce our approach, ContextFormer, which integrates contextual information-based imitation learning (IL) and sequence modeling to stitch sub-optimal trajectory fragments by emulating the representations of a limited number of expert trajectories. To validate our approach, we conduct experiments from two perspectives: 1) We conduct extensive experiments on D4RL benchmarks under the settings of IL, and experimental results demonstrate ContextFormer can achieve competitive performance in multiple IL settings. 2) More importantly, we conduct a comparison of ContextFormer with various competitive DT variants using identical training datasets. The experimental results unveiled ContextFormer's superiority, as it outperformed all other variants, showcasing its remarkable performance.

Jinxin Liu, Xinghong Guo, Zifeng Zhuang et al.

In this paper, we propose a novel approach called DIffusion-guided DIversity (DIDI) for offline behavioral generation. The goal of DIDI is to learn a diverse set of skills from a mixture of label-free offline data. We achieve this by leveraging diffusion probabilistic models as priors to guide the learning process and regularize the policy. By optimizing a joint objective that incorporates diversity and diffusion-guided regularization, we encourage the emergence of diverse behaviors while maintaining the similarity to the offline data. Experimental results in four decision-making domains (Push, Kitchen, Humanoid, and D4RL tasks) show that DIDI is effective in discovering diverse and discriminative skills. We also introduce skill stitching and skill interpolation, which highlight the generalist nature of the learned skill space. Further, by incorporating an extrinsic reward function, DIDI enables reward-guided behavior generation, facilitating the learning of diverse and optimal behaviors from sub-optimal data.

Jinxin Liu, Petar Djukic, Michel Kulhandjian et al.

We propose Deep Dict, a deep learning-based lossy time series compressor designed to achieve a high compression ratio while maintaining decompression error within a predefined range. Deep Dict incorporates two essential components: the Bernoulli transformer autoencoder (BTAE) and a distortion constraint. BTAE extracts Bernoulli representations from time series data, reducing the size of the representations compared to conventional autoencoders. The distortion constraint limits the prediction error of BTAE to the desired range. Moreover, in order to address the limitations of common regression losses such as L1/L2, we introduce a novel loss function called quantized entropy loss (QEL). QEL takes into account the specific characteristics of the problem, enhancing robustness to outliers and alleviating optimization challenges. Our evaluation of Deep Dict across ten diverse time series datasets from various domains reveals that Deep Dict outperforms state-of-the-art lossy compressors in terms of compression ratio by a significant margin by up to 53.66%.

Jinxin Liu, Murat Simsek, Michele Nogueira et al.

Timely response of Network Intrusion Detection Systems (NIDS) is constrained by the flow generation process which requires accumulation of network packets. This paper introduces Multivariate Time Series (MTS) early detection into NIDS to identify malicious flows prior to their arrival at target systems. With this in mind, we first propose a novel feature extractor, Time Series Network Flow Meter (TS-NFM), that represents network flow as MTS with explainable features, and a new benchmark dataset is created using TS-NFM and the meta-data of CICIDS2017, called SCVIC-TS-2022. Additionally, a new deep learning-based early detection model called Multi-Domain Transformer (MDT) is proposed, which incorporates the frequency domain into Transformer. This work further proposes a Multi-Domain Multi-Head Attention (MD-MHA) mechanism to improve the ability of MDT to extract better features. Based on the experimental results, the proposed methodology improves the earliness of the conventional NIDS (i.e., percentage of packets that are used for classification) by 5x10^4 times and duration-based earliness (i.e., percentage of duration of the classified packets of a flow) by a factor of 60, resulting in a 84.1% macro F1 score (31% higher than Transformer) on SCVIC-TS-2022. Additionally, the proposed MDT outperforms the state-of-the-art early detection methods by 5% and 6% on ECG and Wafer datasets, respectively.

Jinxin Liu, Hao Shen, Donglin Wang et al.

Unsupervised reinforcement learning aims to acquire skills without prior goal representations, where an agent automatically explores an open-ended environment to represent goals and learn the goal-conditioned policy. However, this procedure is often time-consuming, limiting the rollout in some potentially expensive target environments. The intuitive approach of training in another interaction-rich environment disrupts the reproducibility of trained skills in the target environment due to the dynamics shifts and thus inhibits direct transferring. Assuming free access to a source environment, we propose an unsupervised domain adaptation method to identify and acquire skills across dynamics. Particularly, we introduce a KL regularized objective to encourage emergence of skills, rewarding the agent for both discovering skills and aligning its behaviors respecting dynamics shifts. This suggests that both dynamics (source and target) shape the reward to facilitate the learning of adaptive skills. We also conduct empirical experiments to demonstrate that our method can effectively learn skills that can be smoothly deployed in target.

Yachen Kang, Jinxin Liu, Xin Cao et al.

We propose an approach for inverse reinforcement learning from hetero-domain which learns a reward function in the simulator, drawing on the demonstrations from the real world. The intuition behind the method is that the reward function should not only be oriented to imitate the experts, but should encourage actions adjusted for the dynamics difference between the simulator and the real world. To achieve this, the widely used GAN-inspired IRL method is adopted, and its discriminator, recognizing policy-generating trajectories, is modified with the quantification of dynamics difference. The training process of the discriminator can yield the transferable reward function suitable for simulator dynamics, which can be guaranteed by derivation. Effectively, our method assigns higher rewards for demonstration trajectories which do not exploit discrepancies between the two domains. With extensive experiments on continuous control tasks, our method shows its effectiveness and demonstrates its scalability to high-dimensional tasks.

Jinxin Liu, Murat Simsek, Burak Kantarci et al.

Access to resources by users may need to be granted only upon certain conditions and contexts, perhaps particularly in cyber-physical settings. Unfortunately, creating and modifying context-sensitive access control solutions in dynamic environments creates ongoing challenges to manage the authorization contexts. This paper proposes RASA, a context-sensitive access authorization approach and mechanism leveraging unsupervised machine learning to automatically infer risk-based authorization decision boundaries. We explore RASA in a healthcare usage environment, wherein cyber and physical conditions create context-specific risks for protecting private health information. The risk levels are associated with access control decisions recommended by a security policy. A coupling method is introduced to track coexistence of the objects within context using frequency and duration of coexistence, and these are clustered to reveal sets of actions with common risk levels; these are used to create authorization decision boundaries. In addition, we propose a method for assessing the risk level and labelling the clusters with respect to their corresponding risk levels. We evaluate the promise of RASA-generated policies against a heuristic rule-based policy. By employing three different coupling features (frequency-based, duration-based, and combined features), the decisions of the unsupervised method and that of the policy are more than 99% consistent.

Jinxin Liu, Donglin Wang, Qiangxing Tian et al.

It is of significance for an agent to learn a widely applicable and general-purpose policy that can achieve diverse goals including images and text descriptions. Considering such perceptually-specific goals, the frontier of deep reinforcement learning research is to learn a goal-conditioned policy without hand-crafted rewards. To learn this kind of policy, recent works usually take as the reward the non-parametric distance to a given goal in an explicit embedding space. From a different viewpoint, we propose a novel unsupervised learning approach named goal-conditioned policy with intrinsic motivation (GPIM), which jointly learns both an abstract-level policy and a goal-conditioned policy. The abstract-level policy is conditioned on a latent variable to optimize a discriminator and discovers diverse states that are further rendered into perceptually-specific goals for the goal-conditioned policy. The learned discriminator serves as an intrinsic reward function for the goal-conditioned policy to imitate the trajectory induced by the abstract-level policy. Experiments on various robotic tasks demonstrate the effectiveness and efficiency of our proposed GPIM method which substantially outperforms prior techniques.