Meng Fang

Yibin Lei, Yu Cao, Dianqi Li et al. · uw

Generating high-quality textual adversarial examples is critical for investigating the pitfalls of natural language processing (NLP) models and further promoting their robustness. Existing attacks are usually realized through word-level or sentence-level perturbations, which either limit the perturbation space or sacrifice fluency and textual quality, both affecting the attack effectiveness. In this paper, we propose Phrase-Level Textual Adversarial aTtack (PLAT) that generates adversarial samples through phrase-level perturbations. PLAT first extracts the vulnerable phrases as attack targets by a syntactic parser, and then perturbs them by a pre-trained blank-infilling model. Such flexible perturbation design substantially expands the search space for more effective attacks without introducing too many modifications, and meanwhile maintaining the textual fluency and grammaticality via contextualized generation using surrounding texts. Moreover, we develop a label-preservation filter leveraging the likelihoods of language models fine-tuned on each class, rather than textual similarity, to rule out those perturbations that potentially alter the original class label for humans. Extensive experiments and human evaluation demonstrate that PLAT has a superior attack effectiveness as well as a better label consistency than strong baselines.

Qin Zhang, Shangsi Chen, Dongkuan Xu et al. · uw

Open domain question answering (ODQA) is a longstanding task aimed at answering factual questions from a large knowledge corpus without any explicit evidence in natural language processing (NLP). Recent works have predominantly focused on improving the answering accuracy and achieved promising progress. However, higher accuracy often comes with more memory consumption and inference latency, which might not necessarily be efficient enough for direct deployment in the real world. Thus, a trade-off between accuracy, memory consumption and processing speed is pursued. In this paper, we provide a survey of recent advances in the efficiency of ODQA models. We walk through the ODQA models and conclude the core techniques on efficiency. Quantitative analysis on memory cost, processing speed, accuracy and overall comparison are given. We hope that this work would keep interested scholars informed of the advances and open challenges in ODQA efficiency research, and thus contribute to the further development of ODQA efficiency.

Yu Cao, Dianqi Li, Meng Fang et al. · uw

We present Twin Answer Sentences Attack (TASA), an adversarial attack method for question answering (QA) models that produces fluent and grammatical adversarial contexts while maintaining gold answers. Despite phenomenal progress on general adversarial attacks, few works have investigated the vulnerability and attack specifically for QA models. In this work, we first explore the biases in the existing models and discover that they mainly rely on keyword matching between the question and context, and ignore the relevant contextual relations for answer prediction. Based on two biases above, TASA attacks the target model in two folds: (1) lowering the model's confidence on the gold answer with a perturbed answer sentence; (2) misguiding the model towards a wrong answer with a distracting answer sentence. Equipped with designed beam search and filtering methods, TASA can generate more effective attacks than existing textual attack methods while sustaining the quality of contexts, in extensive experiments on five QA datasets and human evaluations.

Yuxuan Huang, Yihang Chen, Zhiyuan He et al.

Agentic web search increasingly faces two distinct demands: deep reasoning over a single target, and structured aggregation across many entities and heterogeneous sources. Current systems struggle on both fronts. Breadth-oriented tasks demand schema-aligned outputs with wide coverage and cross-entity consistency, while depth-oriented tasks require coherent reasoning over long, branching search trajectories. We introduce \textbf{Web2BigTable}, a multi-agent framework for web-to-table search that supports both regimes. Web2BigTable adopts a bi-level architecture in which an upper-level orchestrator decomposes the task into sub-problems and lower-level worker agents solve them in parallel. Through a closed-loop run--verify--reflect process, the framework jointly improves decomposition and execution over time via persistent, human-readable external memory, with self-evolving updates to each single-agent. During execution, workers coordinate through a shared workspace that makes partial findings visible, allowing them to reduce redundant exploration, reconcile conflicting evidence, and adapt to emerging coverage gaps. Web2BigTable sets a new state of the art on WideSearch, reaching an Avg@4 Success Rate of \textbf{38.50} ($7.5\times$ the second best at 5.10), Row F1 of \textbf{63.53} (+25.03 over the second best), and Item F1 of \textbf{80.12} (+14.42 over the second best). It also generalises to depth-oriented search on XBench-DeepSearch, achieving 73.0 accuracy. Code is available at https://github.com/web2bigtable/web2bigtable.

Tianjin Huang, Tianlong Chen, Meng Fang et al.

Recent works have impressively demonstrated that there exists a subnetwork in randomly initialized convolutional neural networks (CNNs) that can match the performance of the fully trained dense networks at initialization, without any optimization of the weights of the network (i.e., untrained networks). However, the presence of such untrained subnetworks in graph neural networks (GNNs) still remains mysterious. In this paper we carry out the first-of-its-kind exploration of discovering matching untrained GNNs. With sparsity as the core tool, we can find \textit{untrained sparse subnetworks} at the initialization, that can match the performance of \textit{fully trained dense} GNNs. Besides this already encouraging finding of comparable performance, we show that the found untrained subnetworks can substantially mitigate the GNN over-smoothing problem, hence becoming a powerful tool to enable deeper GNNs without bells and whistles. We also observe that such sparse untrained subnetworks have appealing performance in out-of-distribution detection and robustness of input perturbations. We evaluate our method across widely-used GNN architectures on various popular datasets including the Open Graph Benchmark (OGB).

Xin Zheng, Yixin Liu, Zhifeng Bao et al.

Data-centric AI, with its primary focus on the collection, management, and utilization of data to drive AI models and applications, has attracted increasing attention in recent years. In this article, we conduct an in-depth and comprehensive review, offering a forward-looking outlook on the current efforts in data-centric AI pertaining to graph data-the fundamental data structure for representing and capturing intricate dependencies among massive and diverse real-life entities. We introduce a systematic framework, Data-centric Graph Machine Learning (DC-GML), that encompasses all stages of the graph data lifecycle, including graph data collection, exploration, improvement, exploitation, and maintenance. A thorough taxonomy of each stage is presented to answer three critical graph-centric questions: (1) how to enhance graph data availability and quality; (2) how to learn from graph data with limited-availability and low-quality; (3) how to build graph MLOps systems from the graph data-centric view. Lastly, we pinpoint the future prospects of the DC-GML domain, providing insights to navigate its advancements and applications.

Zihan Zhang, Meng Fang, Ling Chen et al.

Recent work incorporates pre-trained word embeddings such as BERT embeddings into Neural Topic Models (NTMs), generating highly coherent topics. However, with high-quality contextualized document representations, do we really need sophisticated neural models to obtain coherent and interpretable topics? In this paper, we conduct thorough experiments showing that directly clustering high-quality sentence embeddings with an appropriate word selecting method can generate more coherent and diverse topics than NTMs, achieving also higher efficiency and simplicity.

Yunqiu Xu, Meng Fang, Ling Chen et al.

Text-based games provide an interactive way to study natural language processing. While deep reinforcement learning has shown effectiveness in developing the game playing agent, the low sample efficiency and the large action space remain to be the two major challenges that hinder the DRL from being applied in the real world. In this paper, we address the challenges by introducing world-perceiving modules, which automatically decompose tasks and prune actions by answering questions about the environment. We then propose a two-phase training framework to decouple language learning from reinforcement learning, which further improves the sample efficiency. The experimental results show that the proposed method significantly improves the performance and sample efficiency. Besides, it shows robustness against compound error and limited pre-training data.

Lu Yin, Shiwei Liu, Meng Fang et al.

Lottery tickets (LTs) is able to discover accurate and sparse subnetworks that could be trained in isolation to match the performance of dense networks. Ensemble, in parallel, is one of the oldest time-proven tricks in machine learning to improve performance by combining the output of multiple independent models. However, the benefits of ensemble in the context of LTs will be diluted since ensemble does not directly lead to stronger sparse subnetworks, but leverages their predictions for a better decision. In this work, we first observe that directly averaging the weights of the adjacent learned subnetworks significantly boosts the performance of LTs. Encouraged by this observation, we further propose an alternative way to perform an 'ensemble' over the subnetworks identified by iterative magnitude pruning via a simple interpolating strategy. We call our method Lottery Pools. In contrast to the naive ensemble which brings no performance gains to each single subnetwork, Lottery Pools yields much stronger sparse subnetworks than the original LTs without requiring any extra training or inference cost. Across various modern architectures on CIFAR-10/100 and ImageNet, we show that our method achieves significant performance gains in both, in-distribution and out-of-distribution scenarios. Impressively, evaluated with VGG-16 and ResNet-18, the produced sparse subnetworks outperform the original LTs by up to 1.88% on CIFAR-100 and 2.36% on CIFAR-100-C; the resulting dense network surpasses the pre-trained dense-model up to 2.22% on CIFAR-100 and 2.38% on CIFAR-100-C.

Yu Cao, Wei Bi, Meng Fang et al.

Towards building intelligent dialogue agents, there has been a growing interest in introducing explicit personas in generation models. However, with limited persona-based dialogue data at hand, it may be difficult to train a dialogue generation model well. We point out that the data challenges of this generation task lie in two aspects: first, it is expensive to scale up current persona-based dialogue datasets; second, each data sample in this task is more complex to learn with than conventional dialogue data. To alleviate the above data issues, we propose a data manipulation method, which is model-agnostic to be packed with any persona-based dialogue generation model to improve its performance. The original training samples will first be distilled and thus expected to be fitted more easily. Next, we show various effective ways that can diversify such easier distilled data. A given base model will then be trained via the constructed data curricula, i.e. first on augmented distilled samples and then on original ones. Experiments illustrate the superiority of our method with two strong base dialogue models (Transformer encoder-decoder and GPT2).

Yinya Huang, Lemao Liu, Kun Xu et al.

Textual logical reasoning, especially question-answering (QA) tasks with logical reasoning, requires awareness of particular logical structures. The passage-level logical relations represent entailment or contradiction between propositional units (e.g., a concluding sentence). However, such structures are unexplored as current QA systems focus on entity-based relations. In this work, we propose logic structural-constraint modeling to solve the logical reasoning QA and introduce discourse-aware graph networks (DAGNs). The networks first construct logic graphs leveraging in-line discourse connectives and generic logic theories, then learn logic representations by end-to-end evolving the logic relations with an edge-reasoning mechanism and updating the graph features. This pipeline is applied to a general encoder, whose fundamental features are joined with the high-level logic features for answer prediction. Experiments on three textual logical reasoning datasets demonstrate the reasonability of the logical structures built in DAGNs and the effectiveness of the learned logic features. Moreover, zero-shot transfer results show the features' generality to unseen logical texts.

Lu Yin, Vlado Menkovski, Meng Fang et al.

Recent works on sparse neural network training (sparse training) have shown that a compelling trade-off between performance and efficiency can be achieved by training intrinsically sparse neural networks from scratch. Existing sparse training methods usually strive to find the best sparse subnetwork possible in one single run, without involving any expensive dense or pre-training steps. For instance, dynamic sparse training (DST), is capable of reaching a competitive performance of dense training by iteratively evolving the sparse topology during the course of training. In this paper, we argue that it is better to allocate the limited resources to create multiple low-loss sparse subnetworks and superpose them into a stronger one, instead of allocating all resources entirely to find an individual subnetwork. To achieve this, two desiderata are required: (1) efficiently producing many low-loss subnetworks, the so-called cheap tickets, within one training process limited to the standard training time used in dense training; (2) effectively superposing these cheap tickets into one stronger subnetwork. To corroborate our conjecture, we present a novel sparse training approach, termed Sup-tickets, which can satisfy the above two desiderata concurrently in a single sparse-to-sparse training process. Across various modern architectures on CIFAR-10/100 and ImageNet, we show that Sup-tickets integrates seamlessly with the existing sparse training methods and demonstrates consistent performance improvement.

Tianjin Huang, Shiwei Liu, Tianlong Chen et al.

Despite the fact that adversarial training has become the de facto method for improving the robustness of deep neural networks, it is well-known that vanilla adversarial training suffers from daunting robust overfitting, resulting in unsatisfactory robust generalization. A number of approaches have been proposed to address these drawbacks such as extra regularization, adversarial weights perturbation, and training with more data over the last few years. However, the robust generalization improvement is yet far from satisfactory. In this paper, we approach this challenge with a brand new perspective -- refining historical optimization trajectories. We propose a new method named \textbf{Weighted Optimization Trajectories (WOT)} that leverages the optimization trajectories of adversarial training in time. We have conducted extensive experiments to demonstrate the effectiveness of WOT under various state-of-the-art adversarial attacks. Our results show that WOT integrates seamlessly with the existing adversarial training methods and consistently overcomes the robust overfitting issue, resulting in better adversarial robustness. For example, WOT boosts the robust accuracy of AT-PGD under AA-$L_{\infty}$ attack by 1.53\% $\sim$ 6.11\% and meanwhile increases the clean accuracy by 0.55\%$\sim$5.47\% across SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets.

Fanghua Ye, Meng Fang, Shenghui Li et al.

Query rewriting plays a vital role in enhancing conversational search by transforming context-dependent user queries into standalone forms. Existing approaches primarily leverage human-rewritten queries as labels to train query rewriting models. However, human rewrites may lack sufficient information for optimal retrieval performance. To overcome this limitation, we propose utilizing large language models (LLMs) as query rewriters, enabling the generation of informative query rewrites through well-designed instructions. We define four essential properties for well-formed rewrites and incorporate all of them into the instruction. In addition, we introduce the role of rewrite editors for LLMs when initial query rewrites are available, forming a "rewrite-then-edit" process. Furthermore, we propose distilling the rewriting capabilities of LLMs into smaller models to reduce rewriting latency. Our experimental evaluation on the QReCC dataset demonstrates that informative query rewrites can yield substantially improved retrieval performance compared to human rewrites, especially with sparse retrievers.

Zihan Zhang, Meng Fang, Ling Chen et al.

Continual learning (CL) is a paradigm that aims to replicate the human ability to learn and accumulate knowledge continually without forgetting previous knowledge and transferring it to new tasks. Recent instruction tuning (IT) involves fine-tuning models to make them more adaptable to solving NLP tasks in general. However, it is still uncertain how instruction tuning works in the context of CL tasks. This challenging yet practical problem is formulated as Continual Instruction Tuning (CIT). In this work, we establish a CIT benchmark consisting of learning and evaluation protocols. We curate two long dialogue task streams of different types, InstrDialog and InstrDialog++, to study various CL methods systematically. Our experiments show that existing CL methods do not effectively leverage the rich natural language instructions, and fine-tuning an instruction-tuned model sequentially can yield similar or better results. We further explore different aspects that might affect the learning of CIT. We hope this benchmark will facilitate more research in this direction.

Jun Wang, Meng Fang, Ziyu Wan et al.

In this technical report, we introduce OpenR, an open-source framework designed to integrate key components for enhancing the reasoning capabilities of large language models (LLMs). OpenR unifies data acquisition, reinforcement learning training (both online and offline), and non-autoregressive decoding into a cohesive software platform. Our goal is to establish an open-source platform and community to accelerate the development of LLM reasoning. Inspired by the success of OpenAI's o1 model, which demonstrated improved reasoning abilities through step-by-step reasoning and reinforcement learning, OpenR integrates test-time compute, reinforcement learning, and process supervision to improve reasoning in LLMs. Our work is the first to provide an open-source framework that explores the core techniques of OpenAI's o1 model with reinforcement learning, achieving advanced reasoning capabilities beyond traditional autoregressive methods. We demonstrate the efficacy of OpenR by evaluating it on the MATH dataset, utilising publicly available data and search methods. Our initial experiments confirm substantial gains, with relative improvements in reasoning and performance driven by test-time computation and reinforcement learning through process reward models. The OpenR framework, including code, models, and datasets, is accessible at https://openreasoner.github.io.

Zihan Zhang, Meng Fang, Ling Chen

Adaptive retrieval-augmented generation (ARAG) aims to dynamically determine the necessity of retrieval for queries instead of retrieving indiscriminately to enhance the efficiency and relevance of the sourced information. However, previous works largely overlook the evaluation of ARAG approaches, leading to their effectiveness being understudied. This work presents a benchmark, RetrievalQA, comprising 1,271 short-form questions covering new world and long-tail knowledge. The knowledge necessary to answer the questions is absent from LLMs; therefore, external information must be retrieved to answer correctly. This makes RetrievalQA a suitable testbed to evaluate existing ARAG methods. We observe that calibration-based methods heavily rely on threshold tuning, while vanilla prompting is inadequate for guiding LLMs to make reliable retrieval decisions. Based on our findings, we propose Time-Aware Adaptive Retrieval (TA-ARE), a simple yet effective method that helps LLMs assess the necessity of retrieval without calibration or additional training. The dataset and code will be available at https://github.com/hyintell/RetrievalQA

Zihan Zhang, Meng Fang, Fanghua Ye et al.

Dialogue state tracking (DST) plays an important role in task-oriented dialogue systems. However, collecting a large amount of turn-by-turn annotated dialogue data is costly and inefficient. In this paper, we propose a novel turn-level active learning framework for DST to actively select turns in dialogues to annotate. Given the limited labelling budget, experimental results demonstrate the effectiveness of selective annotation of dialogue turns. Additionally, our approach can effectively achieve comparable DST performance to traditional training approaches with significantly less annotated data, which provides a more efficient way to annotate new dialogue data.

Xiangyu Yin, Sihao Wu, Jiaxu Liu et al.

While Goal-Conditioned Reinforcement Learning (GCRL) has gained attention, its algorithmic robustness against adversarial perturbations remains unexplored. The attacks and robust representation training methods that are designed for traditional RL become less effective when applied to GCRL. To address this challenge, we first propose the Semi-Contrastive Representation attack, a novel approach inspired by the adversarial contrastive attack. Unlike existing attacks in RL, it only necessitates information from the policy function and can be seamlessly implemented during deployment. Then, to mitigate the vulnerability of existing GCRL algorithms, we introduce Adversarial Representation Tactics, which combines Semi-Contrastive Adversarial Augmentation with Sensitivity-Aware Regularizer to improve the adversarial robustness of the underlying RL agent against various types of perturbations. Extensive experiments validate the superior performance of our attack and defence methods across multiple state-of-the-art GCRL algorithms. Our tool ReRoGCRL is available at https://github.com/TrustAI/ReRoGCRL.

Qiang He, Tianyi Zhou, Meng Fang et al.

Representation rank is an important concept for understanding the role of Neural Networks (NNs) in Deep Reinforcement learning (DRL), which measures the expressive capacity of value networks. Existing studies focus on unboundedly maximizing this rank; nevertheless, that approach would introduce overly complex models in the learning, thus undermining performance. Hence, fine-tuning representation rank presents a challenging and crucial optimization problem. To address this issue, we find a guiding principle for adaptive control of the representation rank. We employ the Bellman equation as a theoretical foundation and derive an upper bound on the cosine similarity of consecutive state-action pairs representations of value networks. We then leverage this upper bound to propose a novel regularizer, namely BEllman Equation-based automatic rank Regularizer (BEER). This regularizer adaptively regularizes the representation rank, thus improving the DRL agent's performance. We first validate the effectiveness of automatic control of rank on illustrative experiments. Then, we scale up BEER to complex continuous control tasks by combining it with the deterministic policy gradient method. Among 12 challenging DeepMind control tasks, BEER outperforms the baselines by a large margin. Besides, BEER demonstrates significant advantages in Q-value approximation. Our code is available at https://github.com/sweetice/BEER-ICLR2024.

Jiaxu Zhao, Lu Yin, Shiwei Liu et al.

The deep neural network (DNN) has been proven effective in various domains. However, they often struggle to perform well on certain minority groups during inference, despite showing strong performance on the majority of data groups. This is because over-parameterized models learned \textit{bias attributes} from a large number of \textit{bias-aligned} training samples. These bias attributes are strongly spuriously correlated with the target variable, causing the models to be biased towards spurious correlations (i.e., \textit{bias-conflicting}). To tackle this issue, we propose a novel \textbf{re}weighted \textbf{s}parse \textbf{t}raining framework, dubbed as \textit{\textbf{REST}}, which aims to enhance the performance of biased data while improving computation and memory efficiency. Our proposed REST framework has been experimentally validated on three datasets, demonstrating its effectiveness in exploring unbiased subnetworks. We found that REST reduces the reliance on spuriously correlated features, leading to better performance across a wider range of data groups with fewer training and inference resources. We highlight that the \textit{REST} framework represents a promising approach for improving the performance of DNNs on biased data, while simultaneously improving computation and memory efficiency. By reducing the reliance on spurious correlations, REST has the potential to enhance the robustness of DNNs and improve their generalization capabilities. Code is released at \url{https://github.com/zhao1402072392/REST}

Tristan Tomilin, Meng Fang, Mykola Pechenizkiy

Advancing safe autonomous systems through reinforcement learning (RL) requires robust benchmarks to evaluate performance, analyze methods, and assess agent competencies. Humans primarily rely on embodied visual perception to safely navigate and interact with their surroundings, making it a valuable capability for RL agents. However, existing vision-based 3D benchmarks only consider simple navigation tasks. To address this shortcoming, we introduce \textbf{HASARD}, a suite of diverse and complex tasks to $\textbf{HA}$rness $\textbf{SA}$fe $\textbf{R}$L with $\textbf{D}$oom, requiring strategic decision-making, comprehending spatial relationships, and predicting the short-term future. HASARD features three difficulty levels and two action spaces. An empirical evaluation of popular baseline methods demonstrates the benchmark's complexity, unique challenges, and reward-cost trade-offs. Visualizing agent navigation during training with top-down heatmaps provides insight into a method's learning process. Incrementally training across difficulty levels offers an implicit learning curriculum. HASARD is the first safe RL benchmark to exclusively target egocentric vision-based learning, offering a cost-effective and insightful way to explore the potential and boundaries of current and future safe RL methods. The environments and baseline implementations are open-sourced at https://sites.google.com/view/hasard-bench/.

Jinchao Ge, Tengfei Cheng, Biao Wu et al.

Analyzing cultural-heritage artifacts remains challenging for MLLMs: general models lack domain expertise, and SFT often overfits superficial patterns, yielding brittle reasoning for authentication and historical attribution. This raises the question of how to equip MLLMs with robust, expert-level reasoning for ancient Greek pottery. We present VaseVL, an SFT-then-RL system that turns evaluation into supervision: we construct a taxonomy of question types, probe the SFT model to localize type-specific performance gaps, and optimize with type-conditioned, compositionality-oriented rewards targeting those gaps. We also release VaseVQA, a comprehensive benchmark of 31,773 images designed to probe deep understanding. Experiments show state-of-the-art results on style classification and historical attribution with marked gains in compositional robustness over SFT-only baselines, validating diagnosis-guided, taxonomy-conditioned reward engineering and providing a reusable resource for future research. Code and dataset will be available at https://github.com/AIGeeksGroup/VaseVQA.

Keqin Peng, Liang Ding, Yuanxin Ouyang et al.

Large language models (LLMs) excel at a range of tasks through in-context learning (ICL), where only a few task examples guide their predictions. However, prior research highlights that LLMs often overlook input-label mapping information in ICL, relying more on their pre-trained knowledge. To address this issue, we introduce In-Context Contrastive Decoding (ICCD), a novel method that emphasizes input-label mapping by contrasting the output distributions between positive and negative in-context examples. Experiments on 7 natural language understanding (NLU) tasks show that our ICCD method brings consistent and significant improvement (up to +1.8 improvement on average) upon 6 different scales of LLMs without requiring additional training. Our approach is versatile, enhancing performance with various demonstration selection methods, demonstrating its broad applicability and effectiveness. The code and scripts are released at https://github.com/Romainpkq/CD_ICL.

Tianjin Huang, Lu Yin, Zhenyu Zhang et al.

This paper reveals a new appeal of the recently emerged large-kernel Convolutional Neural Networks (ConvNets): as the teacher in Knowledge Distillation (KD) for small-kernel ConvNets. While Transformers have led state-of-the-art (SOTA) performance in various fields with ever-larger models and labeled data, small-kernel ConvNets are considered more suitable for resource-limited applications due to the efficient convolution operation and compact weight sharing. KD is widely used to boost the performance of small-kernel ConvNets. However, previous research shows that it is not quite effective to distill knowledge (e.g., global information) from Transformers to small-kernel ConvNets, presumably due to their disparate architectures. We hereby carry out a first-of-its-kind study unveiling that modern large-kernel ConvNets, a compelling competitor to Vision Transformers, are remarkably more effective teachers for small-kernel ConvNets, due to more similar architectures. Our findings are backed up by extensive experiments on both logit-level and feature-level KD ``out of the box", with no dedicated architectural nor training recipe modifications. Notably, we obtain the \textbf{best-ever pure ConvNet} under 30M parameters with \textbf{83.1\%} top-1 accuracy on ImageNet, outperforming current SOTA methods including ConvNeXt V2 and Swin V2. We also find that beneficial characteristics of large-kernel ConvNets, e.g., larger effective receptive fields, can be seamlessly transferred to students through this large-to-small kernel distillation. Code is available at: \url{https://github.com/VITA-Group/SLaK}.

Yinya Huang, Meng Fang, Yu Cao et al.

Recent QA with logical reasoning questions requires passage-level relations among the sentences. However, current approaches still focus on sentence-level relations interacting among tokens. In this work, we explore aggregating passage-level clues for solving logical reasoning QA by using discourse-based information. We propose a discourse-aware graph network (DAGN) that reasons relying on the discourse structure of the texts. The model encodes discourse information as a graph with elementary discourse units (EDUs) and discourse relations, and learns the discourse-aware features via a graph network for downstream QA tasks. Experiments are conducted on two logical reasoning QA datasets, ReClor and LogiQA, and our proposed DAGN achieves competitive results. The source code is available at https://github.com/Eleanor-H/DAGN.

Lei Han, Jiechao Xiong, Peng Sun et al.

StarCraft, one of the most difficult esport games with long-standing history of professional tournaments, has attracted generations of players and fans, and also, intense attentions in artificial intelligence research. Recently, Google's DeepMind announced AlphaStar, a grandmaster level AI in StarCraft II that can play with humans using comparable action space and operations. In this paper, we introduce a new AI agent, named TStarBot-X, that is trained under orders of less computations and can play competitively with expert human players. TStarBot-X takes advantage of important techniques introduced in AlphaStar, and also benefits from substantial innovations including new league training methods, novel multi-agent roles, rule-guided policy search, stabilized policy improvement, lightweight neural network architecture, and importance sampling in imitation learning, etc. We show that with orders of less computation scale, a faithful reimplementation of AlphaStar's methods can not succeed and the proposed techniques are necessary to ensure TStarBot-X's competitive performance. We reveal all technical details that are complementary to those mentioned in AlphaStar, showing the most sensitive parts in league training, reinforcement learning and imitation learning that affect the performance of the agents. Most importantly, this is an open-sourced study that all codes and resources (including the trained model parameters) are publicly accessible via \url{https://github.com/tencent-ailab/tleague_projpage}. We expect this study could be beneficial for both academic and industrial future research in solving complex problems like StarCraft, and also, might provide a sparring partner for all StarCraft II players and other AI agents.

Peng Sun, Jiechao Xiong, Lei Han et al.

Competitive Self-Play (CSP) based Multi-Agent Reinforcement Learning (MARL) has shown phenomenal breakthroughs recently. Strong AIs are achieved for several benchmarks, including Dota 2, Glory of Kings, Quake III, StarCraft II, to name a few. Despite the success, the MARL training is extremely data thirsty, requiring typically billions of (if not trillions of) frames be seen from the environment during training in order for learning a high performance agent. This poses non-trivial difficulties for researchers or engineers and prevents the application of MARL to a broader range of real-world problems. To address this issue, in this manuscript we describe a framework, referred to as TLeague, that aims at large-scale training and implements several main-stream CSP-MARL algorithms. The training can be deployed in either a single machine or a cluster of hybrid machines (CPUs and GPUs), where the standard Kubernetes is supported in a cloud native manner. TLeague achieves a high throughput and a reasonable scale-up when performing distributed training. Thanks to the modular design, it is also easy to extend for solving other multi-agent problems or implementing and verifying MARL algorithms. We present experiments over StarCraft II, ViZDoom and Pommerman to show the efficiency and effectiveness of TLeague. The code is open-sourced and available at https://github.com/tencent-ailab/tleague_projpage

Meng Fang, Shilong Deng, Yudi Zhang et al.

A wide range of real-world applications is characterized by their symbolic nature, necessitating a strong capability for symbolic reasoning. This paper investigates the potential application of Large Language Models (LLMs) as symbolic reasoners. We focus on text-based games, significant benchmarks for agents with natural language capabilities, particularly in symbolic tasks like math, map reading, sorting, and applying common sense in text-based worlds. To facilitate these agents, we propose an LLM agent designed to tackle symbolic challenges and achieve in-game objectives. We begin by initializing the LLM agent and informing it of its role. The agent then receives observations and a set of valid actions from the text-based games, along with a specific symbolic module. With these inputs, the LLM agent chooses an action and interacts with the game environments. Our experimental results demonstrate that our method significantly enhances the capability of LLMs as automated agents for symbolic reasoning, and our LLM agent is effective in text-based games involving symbolic tasks, achieving an average performance of 88% across all tasks.

Jiaxu Zhao, Meng Fang, Shirui Pan et al.

Warning: This paper contains content that may be offensive or upsetting. There has been a significant increase in the usage of large language models (LLMs) in various applications, both in their original form and through fine-tuned adaptations. As a result, LLMs have gained popularity and are being widely adopted by a large user community. However, one of the concerns with LLMs is the potential generation of socially biased content. The existing evaluation methods have many constraints, and their results exhibit a limited degree of interpretability. In this work, we propose a bias evaluation framework named GPTBIAS that leverages the high performance of LLMs (e.g., GPT-4 \cite{openai2023gpt4}) to assess bias in models. We also introduce prompts called Bias Attack Instructions, which are specifically designed for evaluating model bias. To enhance the credibility and interpretability of bias evaluation, our framework not only provides a bias score but also offers detailed information, including bias types, affected demographics, keywords, reasons behind the biases, and suggestions for improvement. We conduct extensive experiments to demonstrate the effectiveness and usability of our bias evaluation framework.

Yucheng Yang, Tianyi Zhou, Qiang He et al.

Unsupervised reinforcement learning (URL) aims to learn general skills for unseen downstream tasks. Mutual Information Skill Learning (MISL) addresses URL by maximizing the mutual information between states and skills but lacks sufficient theoretical analysis, e.g., how well its learned skills can initialize a downstream task's policy. Our new theoretical analysis in this paper shows that the diversity and separability of learned skills are fundamentally critical to downstream task adaptation but MISL does not necessarily guarantee these properties. To complement MISL, we propose a novel disentanglement metric LSEPIN. Moreover, we build an information-geometric connection between LSEPIN and downstream task adaptation cost. For better geometric properties, we investigate a new strategy that replaces the KL divergence in information geometry with Wasserstein distance. We extend the geometric analysis to it, which leads to a novel skill-learning objective WSEP. It is theoretically justified to be helpful to downstream task adaptation and it is capable of discovering more initial policies for downstream tasks than MISL. We finally propose another Wasserstein distance-based algorithm PWSEP that can theoretically discover all optimal initial policies.

Zirui Song, Jingpu Yang, Yuan Huang et al.

Geolocation, the task of identifying an image's location, requires complex reasoning and is crucial for navigation, monitoring, and cultural preservation. However, current methods often produce coarse, imprecise, and non-interpretable localization. A major challenge lies in the quality and scale of existing geolocation datasets. These datasets are typically small-scale and automatically constructed, leading to noisy data and inconsistent task difficulty, with images that either reveal answers too easily or lack sufficient clues for reliable inference. To address these challenges, we introduce a comprehensive geolocation framework with three key components: GeoComp, a large-scale dataset; GeoCoT, a novel reasoning method; and GeoEval, an evaluation metric, collectively designed to address critical challenges and drive advancements in geolocation research. At the core of this framework is GeoComp (Geolocation Competition Dataset), a large-scale dataset collected from a geolocation game platform involving 740K users over two years. It comprises 25 million entries of metadata and 3 million geo-tagged locations spanning much of the globe, with each location annotated thousands to tens of thousands of times by human users. The dataset offers diverse difficulty levels for detailed analysis and highlights key gaps in current models. Building on this dataset, we propose Geographical Chain-of-Thought (GeoCoT), a novel multi-step reasoning framework designed to enhance the reasoning capabilities of Large Vision Models (LVMs) in geolocation tasks. GeoCoT improves performance by integrating contextual and spatial cues through a multi-step process that mimics human geolocation reasoning. Finally, using the GeoEval metric, we demonstrate that GeoCoT significantly boosts geolocation accuracy by up to 25% while enhancing interpretability.

Zijing Shi, Meng Fang, Ling Chen

Text-based games provide valuable environments for language-based autonomous agents. However, planning-then-learning paradigms, such as those combining Monte Carlo Tree Search (MCTS) and reinforcement learning (RL), are notably time-consuming due to extensive iterations. Additionally, these algorithms perform uncertainty-driven exploration but lack language understanding and reasoning abilities. In this paper, we introduce the Monte Carlo planning with Dynamic Memory-guided Large language model (MC-DML) algorithm. MC-DML leverages the language understanding and reasoning capabilities of Large Language Models (LLMs) alongside the exploratory advantages of tree search algorithms. Specifically, we enhance LLMs with in-trial and cross-trial memory mechanisms, enabling them to learn from past experiences and dynamically adjust action evaluations during planning. We conduct experiments on a series of text-based games from the Jericho benchmark. Our results demonstrate that the MC-DML algorithm significantly enhances performance across various games at the initial planning phase, outperforming strong contemporary methods that require multiple iterations. This demonstrates the effectiveness of our algorithm, paving the way for more efficient language-grounded planning in complex environments.

Keqin Peng, Liang Ding, Yuanxin Ouyang et al.

Reasoning Large Language Models (RLLMs) have demonstrated impressive performance on complex tasks, largely due to the adoption of Long Chain-of-Thought (Long CoT) reasoning. However, they often exhibit overthinking -- performing unnecessary reasoning steps even after arriving at the correct answer. Prior work has largely focused on qualitative analyses of overthinking through sample-based observations of long CoTs. In contrast, we present a quantitative analysis of overthinking from the perspective of self-doubt, characterized by excessive token usage devoted to re-verifying already-correct answer. We find that self-doubt significantly contributes to overthinking. In response, we introduce a simple and effective prompting method to reduce the model's over-reliance on input questions, thereby avoiding self-doubt. Specifically, we first prompt the model to question the validity of the input question, and then respond concisely based on the outcome of that evaluation. Experiments on three mathematical reasoning tasks and four datasets with missing premises demonstrate that our method substantially reduces answer length and yields significant improvements across nearly all datasets upon 4 widely-used RLLMs. Further analysis demonstrates that our method effectively minimizes the number of reasoning steps and reduces self-doubt.

Wanqi Yang, Yanda Li, Yunchao Wei et al.

Large audio-language models (LALMs) have achieved near-human performance in sentence-level transcription and emotion recognition. However, existing evaluations focus mainly on surface-level perception, leaving the capacity of models for contextual and inference-driven reasoning in speech-based scenarios insufficiently examined. To address this gap, we introduce SpeechR, a unified benchmark for evaluating reasoning over speech in large audio-language models. SpeechR evaluates models along three key dimensions: factual retrieval, procedural inference, and normative judgment. It includes three distinct evaluation formats. The multiple-choice version measures answer selection accuracy. The generative version assesses the coherence and logical consistency of reasoning chains. The acoustic-feature version investigates whether variations in stress and emotion affect reasoning performance. Evaluations on eleven state-of-the-art LALMs reveal that high transcription accuracy does not translate into strong reasoning capabilities. SpeechR establishes a structured benchmark for evaluating reasoning in spoken language, enabling more targeted analysis of model capabilities across diverse dialogue-based tasks.

Wenhan Han, Meng Fang, Zihan Zhang et al.

The integration of large language model (LLM) techniques in the field of medical analysis has brought about significant advancements, yet the scarcity of large, diverse, and well-annotated datasets remains a major challenge. Medical data and tasks, which vary in format, size, and other parameters, require extensive preprocessing and standardization for effective use in training LLMs. To address these challenges, we introduce MedINST, the Meta Dataset of Biomedical Instructions, a novel multi-domain, multi-task instructional meta-dataset. MedINST comprises 133 biomedical NLP tasks and over 7 million training samples, making it the most comprehensive biomedical instruction dataset to date. Using MedINST as the meta dataset, we curate MedINST32, a challenging benchmark with different task difficulties aiming to evaluate LLMs' generalization ability. We fine-tune several LLMs on MedINST and evaluate on MedINST32, showcasing enhanced cross-task generalization.

Qin Zhang, Hao Ge, Xiaojun Chen et al.

Unsupervised question answering is a promising yet challenging task, which alleviates the burden of building large-scale annotated data in a new domain. It motivates us to study the unsupervised multiple-choice question answering (MCQA) problem. In this paper, we propose a novel framework designed to generate synthetic MCQA data barely based on contexts from the universal domain without relying on any form of manual annotation. Possible answers are extracted and used to produce related questions, then we leverage both named entities (NE) and knowledge graphs to discover plausible distractors to form complete synthetic samples. Experiments on multiple MCQA datasets demonstrate the effectiveness of our method.

Zi-Yuan Hu, Shuo Liang, Duo Zheng et al.

Recent advances in video large language models (VideoLLMs) call for new evaluation protocols and benchmarks for complex temporal reasoning in video-language understanding. Inspired by the needle in a haystack test widely used by LLMs, we introduce a novel task of Needle in a Montage (NeMo), designed to assess VideoLLMs' critical reasoning capabilities, including long-context recall and temporal grounding. To generate video question answering data for our task, we develop a scalable automated data generation pipeline that facilitates high-quality data synthesis. Built upon the proposed pipeline, we present NeMoBench, a video-language benchmark centered on our task. Specifically, our full set of NeMoBench features 31,378 automatically generated question-answer (QA) pairs from 13,486 videos with various durations ranging from seconds to hours. Experiments demonstrate that our pipeline can reliably and automatically generate high-quality evaluation data, enabling NeMoBench to be continuously updated with the latest videos. We evaluate 20 state-of-the-art models on our benchmark, providing extensive results and key insights into their capabilities and limitations. Our project page is available at: https://lavi-lab.github.io/NeMoBench.

Wenhan Han, Yifan Zhang, Zhixun Chen et al.

Multilingual large language models (LLMs) are advancing rapidly, with new models frequently claiming support for an increasing number of languages. However, existing evaluation datasets are limited and lack cross-lingual alignment, leaving assessments of multilingual capabilities fragmented in both language and skill coverage. To address this, we introduce MuBench, a benchmark covering 61 languages and evaluating a broad range of capabilities. We evaluate several state-of-the-art multilingual LLMs and find notable gaps between claimed and actual language coverage, particularly a persistent performance disparity between English and low-resource languages. Leveraging MuBench's alignment, we propose Multilingual Consistency (MLC) as a complementary metric to accuracy for analyzing performance bottlenecks and guiding model improvement. Finally, we pretrain a suite of 1.2B-parameter models on English and Chinese with 500B tokens, varying language ratios and parallel data proportions to investigate cross-lingual transfer dynamics.

Yudi Zhang, Pei Xiao, Lu Wang et al.

In-context learning (ICL) and Retrieval-Augmented Generation (RAG) have gained attention for their ability to enhance LLMs' reasoning by incorporating external knowledge but suffer from limited contextual window size, leading to insufficient information injection. To this end, we propose a novel framework, RuAG, to automatically distill large volumes of offline data into interpretable first-order logic rules, which are injected into LLMs to boost their reasoning capabilities. Our method begins by formulating the search process relying on LLMs' commonsense, where LLMs automatically define head and body predicates. Then, RuAG applies Monte Carlo Tree Search (MCTS) to address the combinational searching space and efficiently discover logic rules from data. The resulting logic rules are translated into natural language, allowing targeted knowledge injection and seamless integration into LLM prompts for LLM's downstream task reasoning. We evaluate our framework on public and private industrial tasks, including natural language processing, time-series, decision-making, and industrial tasks, demonstrating its effectiveness in enhancing LLM's capability over diverse tasks.

Tristan Tomilin, Luka van den Boogaard, Samuel Garcin et al.

Benchmarks play a crucial role in the development and analysis of reinforcement learning (RL) algorithms, with environment availability strongly impacting research. One particularly underexplored intersection is continual learning (CL) in cooperative multi-agent settings. To remedy this, we introduce MEAL (Multi-agent Environments for Adaptive Learning), the first benchmark tailored for continual multi-agent reinforcement learning (CMARL). Existing CL benchmarks run environments on the CPU, leading to computational bottlenecks and limiting the length of task sequences. MEAL leverages JAX for GPU acceleration, enabling continual learning across sequences of 100 tasks on a standard desktop PC in a few hours. We show that naively combining popular CL and MARL methods yields strong performance on simple environments, but fails to scale to more complex settings requiring sustained coordination and adaptation. Our ablation study identifies architectural and algorithmic features critical for CMARL on MEAL.

Jiangshu Du, Yibo Wang, Wenting Zhao et al.

This work is motivated by two key trends. On one hand, large language models (LLMs) have shown remarkable versatility in various generative tasks such as writing, drawing, and question answering, significantly reducing the time required for many routine tasks. On the other hand, researchers, whose work is not only time-consuming but also highly expertise-demanding, face increasing challenges as they have to spend more time reading, writing, and reviewing papers. This raises the question: how can LLMs potentially assist researchers in alleviating their heavy workload? This study focuses on the topic of LLMs assist NLP Researchers, particularly examining the effectiveness of LLM in assisting paper (meta-)reviewing and its recognizability. To address this, we constructed the ReviewCritique dataset, which includes two types of information: (i) NLP papers (initial submissions rather than camera-ready) with both human-written and LLM-generated reviews, and (ii) each review comes with "deficiency" labels and corresponding explanations for individual segments, annotated by experts. Using ReviewCritique, this study explores two threads of research questions: (i) "LLMs as Reviewers", how do reviews generated by LLMs compare with those written by humans in terms of quality and distinguishability? (ii) "LLMs as Metareviewers", how effectively can LLMs identify potential issues, such as Deficient or unprofessional review segments, within individual paper reviews? To our knowledge, this is the first work to provide such a comprehensive analysis.

Bram Grooten, Tristan Tomilin, Gautham Vasan et al.

The visual world provides an abundance of information, but many input pixels received by agents often contain distracting stimuli. Autonomous agents need the ability to distinguish useful information from task-irrelevant perceptions, enabling them to generalize to unseen environments with new distractions. Existing works approach this problem using data augmentation or large auxiliary networks with additional loss functions. We introduce MaDi, a novel algorithm that learns to mask distractions by the reward signal only. In MaDi, the conventional actor-critic structure of deep reinforcement learning agents is complemented by a small third sibling, the Masker. This lightweight neural network generates a mask to determine what the actor and critic will receive, such that they can focus on learning the task. The masks are created dynamically, depending on the current input. We run experiments on the DeepMind Control Generalization Benchmark, the Distracting Control Suite, and a real UR5 Robotic Arm. Our algorithm improves the agent's focus with useful masks, while its efficient Masker network only adds 0.2% more parameters to the original structure, in contrast to previous work. MaDi consistently achieves generalization results better than or competitive to state-of-the-art methods.

Yudi Zhang, Yali Du, Biwei Huang et al.

A major challenge in reinforcement learning is to determine which state-action pairs are responsible for future rewards that are delayed. Reward redistribution serves as a solution to re-assign credits for each time step from observed sequences. While the majority of current approaches construct the reward redistribution in an uninterpretable manner, we propose to explicitly model the contributions of state and action from a causal perspective, resulting in an interpretable reward redistribution and preserving policy invariance. In this paper, we start by studying the role of causal generative models in reward redistribution by characterizing the generation of Markovian rewards and trajectory-wise long-term return and further propose a framework, called Generative Return Decomposition (GRD), for policy optimization in delayed reward scenarios. Specifically, GRD first identifies the unobservable Markovian rewards and causal relations in the generative process. Then, GRD makes use of the identified causal generative model to form a compact representation to train policy over the most favorable subspace of the state space of the agent. Theoretically, we show that the unobservable Markovian reward function is identifiable, as well as the underlying causal structure and causal models. Experimental results show that our method outperforms state-of-the-art methods and the provided visualization further demonstrates the interpretability of our method. The project page is located at https://reedzyd.github.io/GenerativeReturnDecomposition/.

Jiaxu Zhao, Meng Fang, Zijing Shi et al.

\textit{\textbf{\textcolor{red}{Warning}:} This paper contains content that may be offensive or upsetting.} Pretrained conversational agents have been exposed to safety issues, exhibiting a range of stereotypical human biases such as gender bias. However, there are still limited bias categories in current research, and most of them only focus on English. In this paper, we introduce a new Chinese dataset, CHBias, for bias evaluation and mitigation of Chinese conversational language models. Apart from those previous well-explored bias categories, CHBias includes under-explored bias categories, such as ageism and appearance biases, which received less attention. We evaluate two popular pretrained Chinese conversational models, CDial-GPT and EVA2.0, using CHBias. Furthermore, to mitigate different biases, we apply several debiasing methods to the Chinese pretrained models. Experimental results show that these Chinese pretrained models are potentially risky for generating texts that contain social biases, and debiasing methods using the proposed dataset can make response generation less biased while preserving the models' conversational capabilities.

Iftitahu Ni'mah, Meng Fang, Vlado Menkovski et al.

In this study, we analyze automatic evaluation metrics for Natural Language Generation (NLG), specifically task-agnostic metrics and human-aligned metrics. Task-agnostic metrics, such as Perplexity, BLEU, BERTScore, are cost-effective and highly adaptable to diverse NLG tasks, yet they have a weak correlation with human. Human-aligned metrics (CTC, CtrlEval, UniEval) improves correlation level by incorporating desirable human-like qualities as training objective. However, their effectiveness at discerning system-level performance and quality of system outputs remain unclear. We present metric preference checklist as a framework to assess the effectiveness of automatic metrics in three NLG tasks: Text Summarization, Dialogue Response Generation, and Controlled Generation. Our proposed framework provides access: (i) for verifying whether automatic metrics are faithful to human preference, regardless of their correlation level to human; and (ii) for inspecting the strengths and limitations of NLG systems via pairwise evaluation. We show that automatic metrics provide a better guidance than human on discriminating system-level performance in Text Summarization and Controlled Generation tasks. We also show that multi-aspect human-aligned metric (UniEval) is not necessarily dominant over single-aspect human-aligned metrics (CTC, CtrlEval) and task-agnostic metrics (BLEU, BERTScore), particularly in Controlled Generation tasks.

Rui Yang, Yiming Lu, Wenzhe Li et al.

Solving goal-conditioned tasks with sparse rewards using self-supervised learning is promising because of its simplicity and stability over current reinforcement learning (RL) algorithms. A recent work, called Goal-Conditioned Supervised Learning (GCSL), provides a new learning framework by iteratively relabeling and imitating self-generated experiences. In this paper, we revisit the theoretical property of GCSL -- optimizing a lower bound of the goal reaching objective, and extend GCSL as a novel offline goal-conditioned RL algorithm. The proposed method is named Weighted GCSL (WGCSL), in which we introduce an advanced compound weight consisting of three parts (1) discounted weight for goal relabeling, (2) goal-conditioned exponential advantage weight, and (3) best-advantage weight. Theoretically, WGCSL is proved to optimize an equivalent lower bound of the goal-conditioned RL objective and generates monotonically improved policies via an iterated scheme. The monotonic property holds for any behavior policies, and therefore WGCSL can be applied to both online and offline settings. To evaluate algorithms in the offline goal-conditioned RL setting, we provide a benchmark including a range of point and simulated robot domains. Experiments in the introduced benchmark demonstrate that WGCSL can consistently outperform GCSL and existing state-of-the-art offline methods in the fully offline goal-conditioned setting.

Yunqiu Xu, Meng Fang, Ling Chen et al.

Deep reinforcement learning provides a promising approach for text-based games in studying natural language communication between humans and artificial agents. However, the generalization still remains a big challenge as the agents depend critically on the complexity and variety of training tasks. In this paper, we address this problem by introducing a hierarchical framework built upon the knowledge graph-based RL agent. In the high level, a meta-policy is executed to decompose the whole game into a set of subtasks specified by textual goals, and select one of them based on the KG. Then a sub-policy in the low level is executed to conduct goal-conditioned reinforcement learning. We carry out experiments on games with various difficulty levels and show that the proposed method enjoys favorable generalizability.

Iftitahu Ni'mah, Meng Fang, Vlado Menkovski et al.

The ability to detect Out-of-Domain (OOD) inputs has been a critical requirement in many real-world NLP applications. For example, intent classification in dialogue systems. The reason is that the inclusion of unsupported OOD inputs may lead to catastrophic failure of systems. However, it remains an empirical question whether current methods can tackle such problems reliably in a realistic scenario where zero OOD training data is available. In this study, we propose ProtoInfoMax, a new architecture that extends Prototypical Networks to simultaneously process in-domain and OOD sentences via Mutual Information Maximization (InfoMax) objective. Experimental results show that our proposed method can substantially improve performance up to 20% for OOD detection in low resource settings of text classification. We also show that ProtoInfoMax is less prone to typical overconfidence errors of Neural Networks, leading to more reliable prediction results.

Hao Sun, Ziping Xu, Meng Fang et al.

Safe exploration is crucial for the real-world application of reinforcement learning (RL). Previous works consider the safe exploration problem as Constrained Markov Decision Process (CMDP), where the policies are being optimized under constraints. However, when encountering any potential dangers, human tends to stop immediately and rarely learns to behave safely in danger. Motivated by human learning, we introduce a new approach to address safe RL problems under the framework of Early Terminated MDP (ET-MDP). We first define the ET-MDP as an unconstrained MDP with the same optimal value function as its corresponding CMDP. An off-policy algorithm based on context models is then proposed to solve the ET-MDP, which thereby solves the corresponding CMDP with better asymptotic performance and improved learning efficiency. Experiments on various CMDP tasks show a substantial improvement over previous methods that directly solve CMDP.