Hongning Wang

Jiale Cheng, Xiao Liu, Kehan Zheng et al. · tsinghua

Large language models (LLMs) have shown impressive success in various applications. However, these models are often not well aligned with human intents, which calls for additional treatments on them; that is, the alignment problem. To make LLMs better follow user instructions, existing alignment methods primarily focus on further training them. However, the extra training of LLMs is usually expensive in terms of GPU computing; even worse, some LLMs are not accessible for user-demanded training, such as GPTs. In this work, we take a different perspective -- Black-Box Prompt Optimization (BPO) -- to perform alignments. The idea is to optimize user prompts to suit LLMs' input understanding, so as to best realize users' intents without updating LLMs' parameters. BPO leverages human preferences to optimize prompts, thus making it superior to LLM (e.g., ChatGPT) as a prompt engineer. Moreover, BPO is model-agnostic, and the empirical results demonstrate that the BPO-aligned ChatGPT yields a 22% increase in the win rate against its original version and 10% for GPT-4. Notably, the BPO-aligned LLMs can outperform the same models aligned by PPO and DPO, and it also brings additional performance gains when combining BPO with PPO or DPO. Code and datasets are released at https://github.com/thu-coai/BPO.

Xiao Liu, Xuanyu Lei, Shengyuan Wang et al. · tsinghua

Alignment has become a critical step for instruction-tuned Large Language Models (LLMs) to become helpful assistants. However, the effective evaluation of alignment for emerging Chinese LLMs is still largely unexplored. To fill in this gap, we introduce AlignBench, a comprehensive multi-dimensional benchmark for evaluating LLMs' alignment in Chinese. We design a human-in-the-loop data curation pipeline, containing eight main categories, 683 real-scenario rooted queries and corresponding human verified references. To ensure the correctness of references, each knowledge-intensive query is accompanied with evidences collected from reliable web sources (including URLs and quotations) by our annotators. For automatic evaluation, our benchmark employs a rule-calibrated multi-dimensional LLM-as-Judge~\cite{zheng2023judging} approach with Chain-of-Thought to generate explanations and final ratings, ensuring high reliability and interpretability. All evaluation code, data, and LLM generations are available at \url{https://github.com/THUDM/AlignBench}. Since its release, AlignBench has been adopted by top (Chinese) LLMs for evaluating their alignment capabilities in Chinese, including ChatGLM, Qwen, DeepSeek, Yi, Baichuan, and Abab.

Pei Ke, Bosi Wen, Zhuoer Feng et al. · tsinghua

Since the natural language processing (NLP) community started to make large language models (LLMs) act as a critic to evaluate the quality of generated texts, most of the existing works train a critique generation model on the evaluation data labeled by GPT-4's direct prompting. We observe that these models lack the ability to generate informative critiques in both pointwise grading and pairwise comparison especially without references. As a result, their generated critiques cannot provide fine-grained distinguishability on generated texts, causing unsatisfactory evaluation performance. In this paper, we propose a simple yet effective method called Eval-Instruct, which can first acquire pointwise grading critiques with pseudo references and then revise these critiques via multi-path prompting to obtain informative evaluation data in different tasks and settings, including pointwise grading and pairwise comparison with / without references. After fine-tuning on these data, the resulting model CritiqueLLM is empirically shown to outperform ChatGPT and all the open-source baselines and even achieve comparable evaluation performance to GPT-4 in system-level correlations of pointwise grading. We also demonstrate that our generated critiques can act as scalable feedback to further improve the generation quality of strong LLMs like ChatGPT.

Zhendong Chu, Hongning Wang, Yun Xiao et al.

Conversational recommender systems (CRS) explicitly solicit users' preferences for improved recommendations on the fly. Most existing CRS solutions count on a single policy trained by reinforcement learning for a population of users. However, for users new to the system, such a global policy becomes ineffective to satisfy them, i.e., the cold-start challenge. In this paper, we study CRS policy learning for cold-start users via meta-reinforcement learning. We propose to learn a meta policy and adapt it to new users with only a few trials of conversational recommendations. To facilitate fast policy adaptation, we design three synergetic components. Firstly, we design a meta-exploration policy dedicated to identifying user preferences via a few exploratory conversations, which accelerates personalized policy adaptation from the meta policy. Secondly, we adapt the item recommendation module for each user to maximize the recommendation quality based on the collected conversation states during conversations. Thirdly, we propose a Transformer-based state encoder as the backbone to connect the previous two components. It provides comprehensive state representations by modeling complicated relations between positive and negative feedback during the conversation. Extensive experiments on three datasets demonstrate the advantage of our solution in serving new users, compared with a rich set of state-of-the-art CRS solutions.

Nan Wang, Qifan Wang, Yi-Chia Wang et al.

As language models become increasingly integrated into our digital lives, Personalized Text Generation (PTG) has emerged as a pivotal component with a wide range of applications. However, the bias inherent in user written text, often used for PTG model training, can inadvertently associate different levels of linguistic quality with users' protected attributes. The model can inherit the bias and perpetuate inequality in generating text w.r.t. users' protected attributes, leading to unfair treatment when serving users. In this work, we investigate fairness of PTG in the context of personalized explanation generation for recommendations. We first discuss the biases in generated explanations and their fairness implications. To promote fairness, we introduce a general framework to achieve measure-specific counterfactual fairness in explanation generation. Extensive experiments and human evaluations demonstrate the effectiveness of our method.

Bosi Wen, Pei Ke, Xiaotao Gu et al.

Instruction following is one of the fundamental capabilities of large language models (LLMs). As the ability of LLMs is constantly improving, they have been increasingly applied to deal with complex human instructions in real-world scenarios. Therefore, how to evaluate the ability of complex instruction-following of LLMs has become a critical research problem. Existing benchmarks mainly focus on modeling different types of constraints in human instructions while neglecting the composition of different constraints, which is an indispensable constituent in complex instructions. To this end, we propose ComplexBench, a benchmark for comprehensively evaluating the ability of LLMs to follow complex instructions composed of multiple constraints. We propose a hierarchical taxonomy for complex instructions, including 4 constraint types, 19 constraint dimensions, and 4 composition types, and manually collect a high-quality dataset accordingly. To make the evaluation reliable, we augment LLM-based evaluators with rules to effectively verify whether generated texts can satisfy each constraint and composition. Furthermore, we obtain the final evaluation score based on the dependency structure determined by different composition types. ComplexBench identifies significant deficiencies in existing LLMs when dealing with complex instructions with multiple constraints composition.

Xiaoying Zhang, Junpu Chen, Hongning Wang et al.

Off-policy learning, referring to the procedure of policy optimization with access only to logged feedback data, has shown importance in various real-world applications, such as search engines, recommender systems, and etc. While the ground-truth logging policy, which generates the logged data, is usually unknown, previous work simply takes its estimated value in off-policy learning, ignoring both high bias and high variance resulted from such an estimator, especially on samples with small and inaccurately estimated logging probabilities. In this work, we explicitly model the uncertainty in the estimated logging policy and propose a Uncertainty-aware Inverse Propensity Score estimator (UIPS) for improved off-policy learning, with a theoretical convergence guarantee. Experiment results on synthetic and three real-world recommendation datasets demonstrate the advantageous sample efficiency of the proposed UIPS estimator against an extensive list of state-of-the-art baselines.

Chuanhao Li, Huazheng Wang, Mengdi Wang et al.

We tackle the communication efficiency challenge of learning kernelized contextual bandits in a distributed setting. Despite the recent advances in communication-efficient distributed bandit learning, existing solutions are restricted to simple models like multi-armed bandits and linear bandits, which hamper their practical utility. In this paper, instead of assuming the existence of a linear reward mapping from the features to the expected rewards, we consider non-linear reward mappings, by letting agents collaboratively search in a reproducing kernel Hilbert space (RKHS). This introduces significant challenges in communication efficiency as distributed kernel learning requires the transfer of raw data, leading to a communication cost that grows linearly w.r.t. time horizon $T$. We addresses this issue by equipping all agents to communicate via a common Nyström embedding that gets updated adaptively as more data points are collected. We rigorously proved that our algorithm can attain sub-linear rate in both regret and communication cost.

Qing Zhang, Xiaoying Zhang, Yang Liu et al.

Recommendation systems aim to predict users' feedback on items not exposed to them. Confounding bias arises due to the presence of unmeasured variables (e.g., the socio-economic status of a user) that can affect both a user's exposure and feedback. Existing methods either (1) make untenable assumptions about these unmeasured variables or (2) directly infer latent confounders from users' exposure. However, they cannot guarantee the identification of counterfactual feedback, which can lead to biased predictions. In this work, we propose a novel method, i.e., identifiable deconfounder (iDCF), which leverages a set of proxy variables (e.g., observed user features) to resolve the aforementioned non-identification issue. The proposed iDCF is a general deconfounded recommendation framework that applies proximal causal inference to infer the unmeasured confounders and identify the counterfactual feedback with theoretical guarantees. Extensive experiments on various real-world and synthetic datasets verify the proposed method's effectiveness and robustness.

Xiaoying Zhang, Hongning Wang, Hang Li

Accuracy and diversity have long been considered to be two conflicting goals for recommendations. We point out, however, that as the diversity is typically measured by certain pre-selected item attributes, e.g., category as the most popularly employed one, improved diversity can be achieved without sacrificing recommendation accuracy, as long as the diversification respects the user's preference about the pre-selected attributes. This calls for a fine-grained understanding of a user's preferences over items, where one needs to recognize the user's choice is driven by the quality of the item itself, or the pre-selected attributes of the item. In this work, we focus on diversity defined on item categories. We propose a general diversification framework agnostic to the choice of recommendation algorithms. Our solution disentangles the learnt user representation in the recommendation module into category-independent and category-dependent components to differentiate a user's preference over items from two orthogonal perspectives. Experimental results on three benchmark datasets and online A/B test demonstrate the effectiveness of our solution in improving both recommendation accuracy and diversity. In-depth analysis suggests that the improvement is due to our improved modeling of users' categorical preferences and refined ranking within item categories.

Jiayi Gui, Yiming Liu, Jiale Cheng et al.

Large Language Models (LLMs) have demonstrated notable capabilities across various tasks, showcasing complex problem-solving abilities. Understanding and executing complex rules, along with multi-step planning, are fundamental to logical reasoning and critical for practical LLM agents and decision-making systems. However, evaluating LLMs as effective rule-based executors and planners remains underexplored. In this paper, we introduce LogicGame, a novel benchmark designed to evaluate the comprehensive rule understanding, execution, and planning capabilities of LLMs. Unlike traditional benchmarks, LogicGame provides diverse games that contain a series of rules with an initial state, requiring models to comprehend and apply predefined regulations to solve problems. We create simulated scenarios in which models execute or plan operations to achieve specific outcomes. These game scenarios are specifically designed to distinguish logical reasoning from mere knowledge by relying exclusively on predefined rules. This separation allows for a pure assessment of rule-based reasoning capabilities. The evaluation considers not only final outcomes but also intermediate steps, providing a comprehensive assessment of model performance. Moreover, these intermediate steps are deterministic and can be automatically verified. LogicGame defines game scenarios with varying difficulty levels, from simple rule applications to complex reasoning chains, in order to offer a precise evaluation of model performance on rule understanding and multi-step execution. Utilizing LogicGame, we test various LLMs and identify notable shortcomings in their rule-based logical reasoning abilities.

Lu Lin, Jinghui Chen, Hongning Wang

Graph contrastive learning (GCL), as an emerging self-supervised learning technique on graphs, aims to learn representations via instance discrimination. Its performance heavily relies on graph augmentation to reflect invariant patterns that are robust to small perturbations; yet it still remains unclear about what graph invariance GCL should capture. Recent studies mainly perform topology augmentations in a uniformly random manner in the spatial domain, ignoring its influence on the intrinsic structural properties embedded in the spectral domain. In this work, we aim to find a principled way for topology augmentations by exploring the invariance of graphs from the spectral perspective. We develop spectral augmentation which guides topology augmentations by maximizing the spectral change. Extensive experiments on both graph and node classification tasks demonstrate the effectiveness of our method in self-supervised representation learning. The proposed method also brings promising generalization capability in transfer learning, and is equipped with intriguing robustness property under adversarial attacks. Our study sheds light on a general principle for graph topology augmentation.

Fan Yao, Chuanhao Li, Denis Nekipelov et al.

Content creators compete for exposure on recommendation platforms, and such strategic behavior leads to a dynamic shift over the content distribution. However, how the creators' competition impacts user welfare and how the relevance-driven recommendation influences the dynamics in the long run are still largely unknown. This work provides theoretical insights into these research questions. We model the creators' competition under the assumptions that: 1) the platform employs an innocuous top-$K$ recommendation policy; 2) user decisions follow the Random Utility model; 3) content creators compete for user engagement and, without knowing their utility function in hindsight, apply arbitrary no-regret learning algorithms to update their strategies. We study the user welfare guarantee through the lens of Price of Anarchy and show that the fraction of user welfare loss due to creator competition is always upper bounded by a small constant depending on $K$ and randomness in user decisions; we also prove the tightness of this bound. Our result discloses an intrinsic merit of the myopic approach to the recommendation, i.e., relevance-driven matching performs reasonably well in the long run, as long as users' decisions involve randomness and the platform provides reasonably many alternatives to its users.

Jiaxin Wen, Jian Guan, Hongning Wang et al.

Despite the remarkable success of large language models (LLMs) on traditional natural language processing tasks, their planning ability remains a critical bottleneck in tackling complex multi-step reasoning tasks. Existing approaches mainly rely on prompting or task-specific fine-tuning, often suffering from poor robustness and cross-task generalization. To address the limitation, we introduce CodePlan, a scalable framework that empowers LLMs to generate and follow \textit{code-form plans} -- pseudocode that outlines high-level, structured reasoning processes. By leveraging the structured and versatile nature of code, CodePlan effectively captures the rich semantics and control flows inherent to sophisticated reasoning tasks. Importantly, CodePlan allows automatic extraction of code-form plans from massive, wide-ranging text corpora without the need for curated, task-specific datasets. This enables it to scale up efficiently and improve LLM's reasoning capabilities across diverse scenarios. To train CodePlan, we construct a large-scale dataset of 2M examples that integrate code-form plans with standard prompt-response pairs from existing corpora. With minimal computation overhead during both training and inference, CodePlan achieves a 25.1\% relative improvement compared with directly generating responses, averaged across 13 challenging multi-step reasoning benchmarks, spanning mathematical reasoning, symbolic reasoning, instruction-following, multi-hop QA, and decision-making tasks. Further analysis reveals CodePlan's increasing performance gains on more complex reasoning tasks, as well as significant data efficiency thanks to its generalization ability.

Haozhe Ji, Pei Ke, Hongning Wang et al.

Despite the remarkable advances in language modeling, current mainstream decoding methods still struggle to generate texts that align with human texts across different aspects. In particular, sampling-based methods produce less-repetitive texts which are often disjunctive in discourse, while search-based methods maintain topic coherence at the cost of increased repetition. Overall, these methods fall short in achieving holistic alignment across a broad range of aspects. In this work, we frame decoding from a language model as an optimization problem with the goal of strictly matching the expected performance with human texts measured by multiple metrics of desired aspects simultaneously. The resulting decoding distribution enjoys an analytical solution that scales the input language model distribution via a sequence-level energy function defined by these metrics. And most importantly, we prove that this induced distribution is guaranteed to improve the perplexity on human texts, which suggests a better approximation to the underlying distribution of human texts. To facilitate tractable sampling from this globally normalized distribution, we adopt the Sampling-Importance-Resampling technique. Experiments on various domains and model scales demonstrate the superiority of our method in metrics alignment with human texts and human evaluation over strong baselines.

Zhepei Wei, Chuanhao Li, Haifeng Xu et al.

Most existing works on federated bandits take it for granted that all clients are altruistic about sharing their data with the server for the collective good whenever needed. Despite their compelling theoretical guarantee on performance and communication efficiency, this assumption is overly idealistic and oftentimes violated in practice, especially when the algorithm is operated over self-interested clients, who are reluctant to share data without explicit benefits. Negligence of such self-interested behaviors can significantly affect the learning efficiency and even the practical operability of federated bandit learning. In light of this, we aim to spark new insights into this under-explored research area by formally introducing an incentivized communication problem for federated bandits, where the server shall motivate clients to share data by providing incentives. Without loss of generality, we instantiate this bandit problem with the contextual linear setting and propose the first incentivized communication protocol, namely, Inc-FedUCB, that achieves near-optimal regret with provable communication and incentive cost guarantees. Extensive empirical experiments on both synthetic and real-world datasets further validate the effectiveness of the proposed method across various environments.

Huazheng Wang, David Zhao, Hongning Wang

Bandit algorithms have become a reference solution for interactive recommendation. However, as such algorithms directly interact with users for improved recommendations, serious privacy concerns have been raised regarding its practical use. In this work, we propose a differentially private linear contextual bandit algorithm, via a tree-based mechanism to add Laplace or Gaussian noise to model parameters. Our key insight is that as the model converges during online update, the global sensitivity of its parameters shrinks over time (thus named dynamic global sensitivity). Compared with existing solutions, our dynamic global sensitivity analysis allows us to inject less noise to obtain $(ε, δ)$-differential privacy with added regret caused by noise injection in $\tilde O(\log{T}\sqrt{T}/ε)$. We provide a rigorous theoretical analysis over the amount of noise added via dynamic global sensitivity and the corresponding upper regret bound of our proposed algorithm. Experimental results on both synthetic and real-world datasets confirmed the algorithm's advantage against existing solutions.

Zhexin Zhang, Yida Lu, Jingyuan Ma et al.

The safety of Large Language Models (LLMs) has gained increasing attention in recent years, but there still lacks a comprehensive approach for detecting safety issues within LLMs' responses in an aligned, customizable and explainable manner. In this paper, we propose ShieldLM, an LLM-based safety detector, which aligns with common safety standards, supports customizable detection rules, and provides explanations for its decisions. To train ShieldLM, we compile a large bilingual dataset comprising 14,387 query-response pairs, annotating the safety of responses based on various safety standards. Through extensive experiments, we demonstrate that ShieldLM surpasses strong baselines across four test sets, showcasing remarkable customizability and explainability. Besides performing well on standard detection datasets, ShieldLM has also been shown to be effective as a safety evaluator for advanced LLMs. ShieldLM is released at \url{https://github.com/thu-coai/ShieldLM} to support accurate and explainable safety detection under various safety standards.

Bosi Wen, Yilin Niu, Cunxiang Wang et al.

Instruction following is a fundamental ability of Large Language Models (LLMs), requiring their generated outputs to follow multiple constraints imposed in input instructions. Numerous studies have attempted to enhance this ability through preference optimization or reinforcement learning based on reward signals from LLM-as-a-Judge. However, existing evaluation models for instruction following still possess many deficiencies, such as substantial costs and unreliable assessments. To this end, we propose IF-CRITIC, an LLM critic that can provide efficient and reliable assessments of constraint following in the instructions. We first develop a checklist generator to decompose instructions and generate constraint checklists. With the assistance of the checklists, we collect high-quality critique training data through a multi-stage critique filtering mechanism and employ a constraint-level preference optimization method to train IF-CRITIC. Extensive experiments demonstrate that the evaluation performance of IF-CRITIC can beat strong LLM-as-a-Judge baselines, including Deepseek-R1 and o4-mini. With the scalable reward signals provided by IF-CRITIC, LLMs can achieve substantial performance gains in instruction-following optimization under lower computational overhead compared to strong LLM critic baselines.

Haozhe Ji, Cheng Lu, Yilin Niu et al.

The alignment of language models with human preferences is vital for their application in real-world tasks. The problem is formulated as optimizing the model's policy to maximize the expected reward that reflects human preferences with minimal deviation from the initial policy. While considered as a straightforward solution, reinforcement learning (RL) suffers from high variance in policy updates, which impedes efficient policy improvement. Recently, direct preference optimization (DPO) was proposed to directly optimize the policy from preference data. However, we show that DPO derived based on the optimal solution of the problem leads to a compromised mean-seeking approximation of the optimal solution in practice. In this paper, we propose efficient exact optimization (EXO) of the alignment objective. EXO is guaranteed to optimize in the same direction as RL algorithms asymptotically for arbitrary policy parametrization. This leads to the same mode-seeking solution, while enables efficient optimization by circumventing the complexities of RL. We also compare our method to DPO with both theoretical and empirical analyses, and further demonstrate the advantages of our method over existing approaches on realistic human preference data. Code is available at https://github.com/haozheji/exact-optimization.

Jian Guan, Wei Wu, Zujie Wen et al.

The notable success of large language models (LLMs) has sparked an upsurge in building language agents to complete various complex tasks. We present AMOR, an agent framework based on open-source LLMs, which reasons with external knowledge bases and adapts to specific domains through human supervision to the reasoning process. AMOR builds reasoning logic over a finite state machine (FSM) that solves problems through autonomous executions and transitions over disentangled modules. This allows humans to provide direct feedback to the individual modules, and thus naturally forms process supervision. Based on this reasoning and feedback framework, we develop AMOR through two-stage fine-tuning: warm-up and adaptation. The former fine-tunes the LLM with examples automatically constructed from various public datasets, enabling AMOR to generalize across different knowledge environments, while the latter tailors AMOR to specific domains using process feedback. Extensive experiments across multiple domains demonstrate the advantage of AMOR to strong baselines, thanks to its FSM-based reasoning and process feedback mechanism. The code and data are publicly available at \url{https://github.com/JianGuanTHU/AMOR}.

Weihang Su, Yichen Tang, Qingyao Ai et al.

Retrieval-augmented generation (RAG) techniques have emerged as a promising solution to enhance the reliability of large language models (LLMs) by addressing issues like hallucinations, outdated knowledge, and domain adaptation. In particular, existing RAG methods append relevant documents retrieved from external corpus or databases to the input of LLMs to guide their generation process, which we refer to as the in-context knowledge injection method. While this approach is simple and often effective, it has inherent limitations. Firstly, increasing the context length and number of relevant documents can lead to higher computational overhead and degraded performance, especially in complex reasoning tasks. More importantly, in-context knowledge injection operates primarily at the input level, but LLMs store their internal knowledge in their parameters. This gap fundamentally limits the capacity of in-context methods. To this end, we introduce Parametric retrieval-augmented generation (Parametric RAG), a new RAG paradigm that integrates external knowledge directly into the parameters of feed-forward networks (FFN) of an LLM through document parameterization. This approach not only saves online computational costs by eliminating the need to inject multiple documents into the LLMs' input context, but also deepens the integration of external knowledge into the parametric knowledge space of the LLM. Experimental results demonstrate that Parametric RAG substantially enhances both the effectiveness and efficiency of knowledge augmentation in LLMs. Also, it can be combined with in-context RAG methods to achieve even better performance. We have open-sourced all the code, data, and models in the following anonymized GitHub link: https://github.com/oneal2000/PRAG

Zhexin Zhang, Xian Qi Loye, Victor Shea-Jay Huang et al.

Large Reasoning Models (LRMs) have achieved remarkable success on reasoning-intensive tasks such as mathematics and programming. However, their enhanced reasoning capabilities do not necessarily translate to improved safety performance-and in some cases, may even degrade it. This raises an important research question: how can we enhance the safety of LRMs? In this paper, we present a comprehensive empirical study on how to enhance the safety of LRMs through Supervised Fine-Tuning (SFT). Our investigation begins with an unexpected observation: directly distilling safe responses from DeepSeek-R1 fails to significantly enhance safety. We analyze this phenomenon and identify three key failure patterns that contribute to it. We then demonstrate that explicitly addressing these issues during the data distillation process can lead to substantial safety improvements. Next, we explore whether a long and complex reasoning process is necessary for achieving safety. Interestingly, we find that simply using short or template-based reasoning process can attain comparable safety performance-and are significantly easier for models to learn than more intricate reasoning chains. These findings prompt a deeper reflection on the role of reasoning in ensuring safety. Finally, we find that mixing math reasoning data during safety fine-tuning is helpful to balance safety and over-refusal. Overall, we hope our empirical study could provide a more holistic picture on enhancing the safety of LRMs. The code and data used in our experiments are released in https://github.com/thu-coai/LRM-Safety-Study.

Jiale Cheng, Xiao Liu, Cunxiang Wang et al.

Instruction-following is a fundamental capability of language models, requiring the model to recognize even the most subtle requirements in the instructions and accurately reflect them in its output. Such an ability is well-suited for and often optimized by preference learning. However, existing methods often directly sample multiple independent responses from the model when creating preference pairs. Such practice can introduce content variations irrelevant to whether the instruction is precisely followed (e.g., different expressions about the same semantic), interfering with the goal of teaching models to recognize the key differences that lead to improved instruction following. In light of this, we introduce SPaR, a self-play framework integrating tree-search self-refinement to yield valid and comparable preference pairs free from distractions. By playing against itself, an LLM employs a tree-search strategy to refine its previous responses with respect to the instruction while minimizing unnecessary variations. Our experiments show that a LLaMA3-8B model, trained over three iterations guided by SPaR, surpasses GPT-4-Turbo on the IFEval benchmark without losing general capabilities. Furthermore, SPaR demonstrates promising scalability, greatly enhancing models like GLM-4-9B and LLaMA3-70B. We also identify how inference scaling in tree search would impact model performance. Our code and data are publicly available at https://github.com/thu-coai/SPaR.

Jiale Cheng, Ruiliang Lyu, Xiaotao Gu et al. · tsinghua

Video generation models have achieved remarkable progress in text-to-video tasks. These models are typically trained on text-video pairs with highly detailed and carefully crafted descriptions, while real-world user inputs during inference are often concise, vague, or poorly structured. This gap makes prompt optimization crucial for generating high-quality videos. Current methods often rely on large language models (LLMs) to refine prompts through in-context learning, but suffer from several limitations: they may distort user intent, omit critical details, or introduce safety risks. Moreover, they optimize prompts without considering the impact on the final video quality, which can lead to suboptimal results. To address these issues, we introduce VPO, a principled framework that optimizes prompts based on three core principles: harmlessness, accuracy, and helpfulness. The generated prompts faithfully preserve user intents and, more importantly, enhance the safety and quality of generated videos. To achieve this, VPO employs a two-stage optimization approach. First, we construct and refine a supervised fine-tuning (SFT) dataset based on principles of safety and alignment. Second, we introduce both text-level and video-level feedback to further optimize the SFT model with preference learning. Our extensive experiments demonstrate that VPO significantly improves safety, alignment, and video quality compared to baseline methods. Moreover, VPO shows strong generalization across video generation models. Furthermore, we demonstrate that VPO could outperform and be combined with RLHF methods on video generation models, underscoring the effectiveness of VPO in aligning video generation models. Our code and data are publicly available at https://github.com/thu-coai/VPO.

Bosi Wen, Pei Ke, Yufei Sun et al.

Since the adoption of large language models (LLMs) for text evaluation has become increasingly prevalent in the field of natural language processing (NLP), a series of existing works attempt to optimize the prompts for LLM evaluators to improve their alignment with human judgment. However, their efforts are limited to optimizing individual factors of evaluation prompts, such as evaluation criteria or output formats, neglecting the combinatorial impact of multiple factors, which leads to insufficient optimization of the evaluation pipeline. Nevertheless, identifying well-behaved prompting strategies for adjusting multiple factors requires extensive enumeration. To this end, we comprehensively integrate 8 key factors for evaluation prompts and propose a novel automatic prompting strategy optimization method called Heuristic Prompting Strategy Search (HPSS). Inspired by the genetic algorithm, HPSS conducts an iterative search to find well-behaved prompting strategies for LLM evaluators. A heuristic function is employed to guide the search process, enhancing the performance of our algorithm. Extensive experiments across four evaluation tasks demonstrate the effectiveness of HPSS, consistently outperforming both human-designed evaluation prompts and existing automatic prompt optimization methods. Our code is available at https://github.com/thu-coai/HPSS.

Jiale Cheng, Yusen Liu, Xinyu Zhang et al.

Large language models (LLMs) increasingly rely on long-context modeling for tasks such as document understanding, code analysis, and multi-step reasoning. However, scaling context windows to the million-token level brings prohibitive computational and memory costs, limiting the practicality of long-context LLMs. In this work, we take a different perspective-visual context scaling-to tackle this challenge. Instead of extending token-based sequences, we propose Glyph, a framework that renders long texts into images and processes them with vision-language models (VLMs). This approach substantially compresses textual input while preserving semantic information, and we further design an LLM-driven genetic search to identify optimal visual rendering configurations for balancing accuracy and compression. Through extensive experiments, we demonstrate that our method achieves 3-4x token compression while maintaining accuracy comparable to leading LLMs such as Qwen3-8B on various long-context benchmarks. This compression also leads to around 4x faster prefilling and decoding, and approximately 2x faster SFT training. Furthermore, under extreme compression, a 128K-context VLM could scale to handle 1M-token-level text tasks. In addition, the rendered text data benefits real-world multimodal tasks, such as document understanding. Our code and model are released at https://github.com/thu-coai/Glyph.

Renmiao Chen, Shiyao Cui, Xuancheng Huang et al.

Jailbreak attacks against multimodal large language Models (MLLMs) are a significant research focus. Current research predominantly focuses on maximizing attack success rate (ASR), often overlooking whether the generated responses actually fulfill the attacker's malicious intent. This oversight frequently leads to low-quality outputs that bypass safety filters but lack substantial harmful content. To address this gap, we propose JPS, \underline{J}ailbreak MLLMs with collaborative visual \underline{P}erturbation and textual \underline{S}teering, which achieves jailbreaks via corporation of visual image and textually steering prompt. Specifically, JPS utilizes target-guided adversarial image perturbations for effective safety bypass, complemented by "steering prompt" optimized via a multi-agent system to specifically guide LLM responses fulfilling the attackers' intent. These visual and textual components undergo iterative co-optimization for enhanced performance. To evaluate the quality of attack outcomes, we propose the Malicious Intent Fulfillment Rate (MIFR) metric, assessed using a Reasoning-LLM-based evaluator. Our experiments show JPS sets a new state-of-the-art in both ASR and MIFR across various MLLMs and benchmarks, with analyses confirming its efficacy. Codes are available at \href{https://github.com/thu-coai/JPS}{https://github.com/thu-coai/JPS}. \color{warningcolor}{Warning: This paper contains potentially sensitive contents.}

Jiale Cheng, Yida Lu, Xiaotao Gu et al.

Although Large Language Models (LLMs) are becoming increasingly powerful, they still exhibit significant but subtle weaknesses, such as mistakes in instruction-following or coding tasks. As these unexpected errors could lead to severe consequences in practical deployments, it is crucial to investigate the limitations within LLMs systematically. Traditional benchmarking approaches cannot thoroughly pinpoint specific model deficiencies, while manual inspections are costly and not scalable. In this paper, we introduce a unified framework, AutoDetect, to automatically expose weaknesses in LLMs across various tasks. Inspired by the educational assessment process that measures students' learning outcomes, AutoDetect consists of three LLM-powered agents: Examiner, Questioner, and Assessor. The collaboration among these three agents is designed to realize comprehensive and in-depth weakness identification. Our framework demonstrates significant success in uncovering flaws, with an identification success rate exceeding 30% in prominent models such as ChatGPT and Claude. More importantly, these identified weaknesses can guide specific model improvements, proving more effective than untargeted data augmentation methods like Self-Instruct. Our approach has led to substantial enhancements in popular LLMs, including the Llama series and Mistral-7b, boosting their performance by over 10% across several benchmarks. Code and data are publicly available at https://github.com/thu-coai/AutoDetect.

Team GLM, Aohan Zeng, Bin Xu et al.

We introduce ChatGLM, an evolving family of large language models that we have been developing over time. This report primarily focuses on the GLM-4 language series, which includes GLM-4, GLM-4-Air, and GLM-4-9B. They represent our most capable models that are trained with all the insights and lessons gained from the preceding three generations of ChatGLM. To date, the GLM-4 models are pre-trained on ten trillions of tokens mostly in Chinese and English, along with a small set of corpus from 24 languages, and aligned primarily for Chinese and English usage. The high-quality alignment is achieved via a multi-stage post-training process, which involves supervised fine-tuning and learning from human feedback. Evaluations show that GLM-4 1) closely rivals or outperforms GPT-4 in terms of general metrics such as MMLU, GSM8K, MATH, BBH, GPQA, and HumanEval, 2) gets close to GPT-4-Turbo in instruction following as measured by IFEval, 3) matches GPT-4 Turbo (128K) and Claude 3 for long context tasks, and 4) outperforms GPT-4 in Chinese alignments as measured by AlignBench. The GLM-4 All Tools model is further aligned to understand user intent and autonomously decide when and which tool(s) touse -- including web browser, Python interpreter, text-to-image model, and user-defined functions -- to effectively complete complex tasks. In practical applications, it matches and even surpasses GPT-4 All Tools in tasks like accessing online information via web browsing and solving math problems using Python interpreter. Over the course, we have open-sourced a series of models, including ChatGLM-6B (three generations), GLM-4-9B (128K, 1M), GLM-4V-9B, WebGLM, and CodeGeeX, attracting over 10 million downloads on Hugging face in the year 2023 alone. The open models can be accessed through https://github.com/THUDM and https://huggingface.co/THUDM.

Jinfeng Zhou, Yongkang Huang, Bosi Wen et al.

Character-based dialogue (aka role-playing) enables users to freely customize characters for interaction, which often relies on LLMs, raising the need to evaluate LLMs' character customization capability. However, existing benchmarks fail to ensure a robust evaluation as they often only involve a single character category or evaluate limited dimensions. Moreover, the sparsity of character features in responses makes feature-focused generative evaluation both ineffective and inefficient. To address these issues, we propose CharacterBench, the largest bilingual generative benchmark, with 22,859 human-annotated samples covering 3,956 characters from 25 detailed character categories. We define 11 dimensions of 6 aspects, classified as sparse and dense dimensions based on whether character features evaluated by specific dimensions manifest in each response. We enable effective and efficient evaluation by crafting tailored queries for each dimension to induce characters' responses related to specific dimensions. Further, we develop CharacterJudge model for cost-effective and stable evaluations. Experiments show its superiority over SOTA automatic judges (e.g., GPT-4) and our benchmark's potential to optimize LLMs' character customization. Our repository is at https://github.com/thu-coai/CharacterBench.

Zhenyu Hou, Yilin Niu, Zhengxiao Du et al. · tsinghua

ChatGLM is a free-to-use AI service powered by the ChatGLM family of large language models (LLMs). In this paper, we present the ChatGLM-RLHF pipeline -- a reinforcement learning from human feedback (RLHF) system -- designed to enhance ChatGLM's alignment with human preferences. ChatGLM-RLHF encompasses three major components: the collection of human preference data, the training of the reward model, and the optimization of policies. Throughout the process of integrating ChatGLM-RLHF into production, we encountered and addressed several unprecedented challenges. We introduce the strategies to mitigate reward variance for stabilized large-scale training, implement model parallelism with fused gradient-descent, and design regularization constraints to avoid catastrophic forgetting in LLMs. Experiments show that ChatGLM-RLHF brings significant improvements in alignment tasks compared to the supervised fine-tuned (SFT) version of ChatGLM. For instance, it achieves on average 15\% more wins against ChatGLM-SFT in Chinese alignment tasks. The work presents our practices of aligning LLMs with human preferences, offering insights into the challenges and solutions in RLHF implementations.

Xiaoying Zhang, Jean-Francois Ton, Wei Shen et al.

We introduce Adversarial Policy Optimization (AdvPO), a novel solution to the pervasive issue of reward over-optimization in Reinforcement Learning from Human Feedback (RLHF) for Large Language Models (LLMs). Over-optimization occurs when a reward model serves as an imperfect proxy for human preference, and RL-driven policy optimization erroneously exploits reward inaccuracies. In this paper, we begin by introducing a lightweight way to quantify uncertainties in rewards, relying solely on the last layer embeddings of the reward model, without the need for computationally expensive reward ensembles. AdvPO then addresses a distributionally robust optimization problem centred around the confidence interval of the reward model's predictions for policy improvement. Through comprehensive experiments on the Anthropic HH and TL;DR summarization datasets, we illustrate the efficacy of AdvPO in mitigating the overoptimization issue, consequently resulting in enhanced performance as evaluated through human-assisted evaluation.

Zhenyu Hou, Pengfan Du, Yilin Niu et al.

This study explores the scaling properties of Reinforcement Learning from Human Feedback (RLHF) in Large Language Models (LLMs). Although RLHF is considered an important step in post-training of LLMs, its scaling potential is still largely unknown. We systematically analyze key components in the RLHF framework--model size, data composition, and inference budget--and their impacts on performance. Our findings show that increasing data diversity and volume improves reward model performance, helping process-supervision models scale better. For policy training, more response samples per prompt boost performance initially but quickly plateau. And larger reward models offer modest gains in policy training. In addition, larger policy models benefit less from RLHF with a fixed reward model. Overall, RLHF scales less efficiently than pretraining, with diminishing returns from additional computational resources. Based on these observations, we propose strategies to optimize RLHF performance within computational limits.

Jinfeng Zhou, Yuxuan Chen, Yihan Shi et al.

LLMs exhibit promising Social Intelligence (SI) in modeling human behavior, raising the need to evaluate LLMs' SI and their discrepancy with humans. SI equips humans with interpersonal abilities to behave wisely in navigating social interactions to achieve social goals. This presents an operational evaluation paradigm: outcome-oriented goal achievement evaluation and process-oriented interpersonal ability evaluation, which existing work fails to address. To this end, we propose SocialEval, a script-based bilingual SI benchmark, integrating outcome- and process-oriented evaluation by manually crafting narrative scripts. Each script is structured as a world tree that contains plot lines driven by interpersonal ability, providing a comprehensive view of how LLMs navigate social interactions. Experiments show that LLMs fall behind humans on both SI evaluations, exhibit prosociality, and prefer more positive social behaviors, even if they lead to goal failure. Analysis of LLMs' formed representation space and neuronal activations reveals that LLMs have developed ability-specific functional partitions akin to the human brain.

Zhuang Chen, Yaru Cao, Guanqun Bi et al.

Emotional support conversation (ESC) helps reduce people's psychological stress and provide emotional value through interactive dialogues. Due to the high cost of crowdsourcing a large ESC corpus, recent attempts use large language models for dialogue augmentation. However, existing approaches largely overlook the social dynamics inherent in ESC, leading to less effective simulations. In this paper, we introduce SocialSim, a novel framework that simulates ESC by integrating key aspects of social interactions: social disclosure and social awareness. On the seeker side, we facilitate social disclosure by constructing a comprehensive persona bank that captures diverse and authentic help-seeking scenarios. On the supporter side, we enhance social awareness by eliciting cognitive reasoning to generate logical and supportive responses. Building upon SocialSim, we construct SSConv, a large-scale synthetic ESC corpus of which quality can even surpass crowdsourced ESC data. We further train a chatbot on SSConv and demonstrate its state-of-the-art performance in both automatic and human evaluations. We believe SocialSim offers a scalable way to synthesize ESC, making emotional care more accessible and practical.

Kehan Zheng, Jinfeng Zhou, Hongning Wang

Large language models are increasingly used in strategic decision-making settings, yet evidence shows that, like humans, they often deviate from full rationality. In this study, we compare LLMs and humans using experimental paradigms directly adapted from behavioral game-theory research. We focus on two well-studied strategic games, Rock-Paper-Scissors and the Prisoner's Dilemma, which are well known for revealing systematic departures from rational play in human subjects. By placing LLMs in identical experimental conditions, we evaluate whether their behaviors exhibit the bounded rationality characteristic of humans. Our findings show that LLMs reproduce familiar human heuristics, such as outcome-based strategy switching and increased cooperation when future interaction is possible, but they apply these rules more rigidly and demonstrate weaker sensitivity to the dynamic changes in the game environment. Model-level analyses reveal distinctive architectural signatures in strategic behavior, and even reasoning models sometimes struggle to find effective strategies in adaptive situations. These results indicate that current LLMs capture only a partial form of human-like bounded rationality and highlight the need for training methods that encourage flexible opponent modeling and stronger context awareness.

Ethan Blaser, Chuanhao Li, Hongning Wang

The demand for collaborative and private bandit learning across multiple agents is surging due to the growing quantity of data generated from distributed systems. Federated bandit learning has emerged as a promising framework for private, efficient, and decentralized online learning. However, almost all previous works rely on strong assumptions of client homogeneity, i.e., all participating clients shall share the same bandit model; otherwise, they all would suffer linear regret. This greatly restricts the application of federated bandit learning in practice. In this work, we introduce a new approach for federated bandits for heterogeneous clients, which clusters clients for collaborative bandit learning under the federated learning setting. Our proposed algorithm achieves non-trivial sub-linear regret and communication cost for all clients, subject to the communication protocol under federated learning that at anytime only one model can be shared by the server.

Shiyao Cui, Qinglin Zhang, Xuan Ouyang et al.

Toxicity detection in multimodal text-image content faces growing challenges, especially with multimodal implicit toxicity, where each modality appears benign on its own but conveys hazard when combined. Multimodal implicit toxicity appears not only as formal statements in social platforms but also prompts that can lead to toxic dialogs from Large Vision-Language Models (LVLMs). Despite the success in unimodal text or image moderation, toxicity detection for multimodal content, particularly the multimodal implicit toxicity, remains underexplored. To fill this gap, we comprehensively build a taxonomy for multimodal implicit toxicity (MMIT) and introduce an MMIT-dataset, comprising 2,100 multimodal statements and prompts across 7 risk categories (31 sub-categories) and 5 typical cross-modal correlation modes. To advance the detection of multimodal implicit toxicity, we build ShieldVLM, a model which identifies implicit toxicity in multimodal statements, prompts and dialogs via deliberative cross-modal reasoning. Experiments show that ShieldVLM outperforms existing strong baselines in detecting both implicit and explicit toxicity. The model and dataset will be publicly available to support future researches. Warning: This paper contains potentially sensitive contents.

Zhepei Wei, Chuanhao Li, Tianze Ren et al.

To enhance the efficiency and practicality of federated bandit learning, recent advances have introduced incentives to motivate communication among clients, where a client participates only when the incentive offered by the server outweighs its participation cost. However, existing incentive mechanisms naively assume the clients are truthful: they all report their true cost and thus the higher cost one participating client claims, the more the server has to pay. Therefore, such mechanisms are vulnerable to strategic clients aiming to optimize their own utility by misreporting. To address this issue, we propose an incentive compatible (i.e., truthful) communication protocol, named Truth-FedBan, where the incentive for each participant is independent of its self-reported cost, and reporting the true cost is the only way to achieve the best utility. More importantly, Truth-FedBan still guarantees the sub-linear regret and communication cost without any overheads. In other words, the core conceptual contribution of this paper is, for the first time, demonstrating the possibility of simultaneously achieving incentive compatibility and nearly optimal regret in federated bandit learning. Extensive numerical studies further validate the effectiveness of our proposed solution.

Zhiwei Wang, Huazheng Wang, Hongning Wang

Adversarial attacks against stochastic multi-armed bandit (MAB) algorithms have been extensively studied in the literature. In this work, we focus on reward poisoning attacks and find most existing attacks can be easily detected by our proposed detection method based on the test of homogeneity, due to their aggressive nature in reward manipulations. This motivates us to study the notion of stealthy attack against stochastic MABs and investigate the resulting attackability. Our analysis shows that against two popularly employed MAB algorithms, UCB1 and $ε$-greedy, the success of a stealthy attack depends on the environmental conditions and the realized reward of the arm pulled in the first round. We also analyze the situation for general MAB algorithms equipped with our attack detection method and find that it is possible to have a stealthy attack that almost always succeeds. This brings new insights into the security risks of MAB algorithms.

Jiaxin Wen, Ruiqi Zhong, Pei Ke et al.

When using language models (LMs) to solve complex problems, humans might struggle to understand the LM-generated solutions and repair the flawed ones. To assist humans in repairing them, we propose to automatically decompose complex solutions into multiple simpler pieces that correspond to specific subtasks. We introduce a novel objective for learning task decomposition, termed assistive value (AssistV), which measures the feasibility and speed for humans to repair the decomposed solution. We collect a dataset of human repair experiences on different decomposed solutions. Utilizing the collected data as in-context examples, we then learn to critique, refine, and rank decomposed solutions to improve AssistV. We validate our method under competitive programming problems: under 177 hours of human study, our method enables non-experts to solve 33.3\% more problems, speeds them up by 3.3x, and empowers them to match unassisted experts.

Peng Wang, Renqin Cai, Hongning Wang

Explanations in a recommender system assist users in making informed decisions among a set of recommended items. Great research attention has been devoted to generating natural language explanations to depict how the recommendations are generated and why the users should pay attention to them. However, due to different limitations of those solutions, e.g., template-based or generation-based, it is hard to make the explanations easily perceivable, reliable and personalized at the same time. In this work, we develop a graph attentive neural network model that seamlessly integrates user, item, attributes, and sentences for extraction-based explanation. The attributes of items are selected as the intermediary to facilitate message passing for user-item specific evaluation of sentence relevance. And to balance individual sentence relevance, overall attribute coverage, and content redundancy, we solve an integer linear programming problem to make the final selection of sentences. Extensive empirical evaluations against a set of state-of-the-art baseline methods on two benchmark review datasets demonstrated the generation quality of the proposed solution.

Fan Yao, Chuanhao Li, Denis Nekipelov et al.

In real-world recommendation problems, especially those with a formidably large item space, users have to gradually learn to estimate the utility of any fresh recommendations from their experience about previously consumed items. This in turn affects their interaction dynamics with the system and can invalidate previous algorithms built on the omniscient user assumption. In this paper, we formalize a model to capture such "learning users" and design an efficient system-side learning solution, coined Noise-Robust Active Ellipsoid Search (RAES), to confront the challenges brought by the non-stationary feedback from such a learning user. Interestingly, we prove that the regret of RAES deteriorates gracefully as the convergence rate of user learning becomes worse, until reaching linear regret when the user's learning fails to converge. Experiments on synthetic datasets demonstrate the strength of RAES for such a contemporaneous system-user learning problem. Our study provides a novel perspective on modeling the feedback loop in recommendation problems.

Chuanhao Li, Hongning Wang

Contextual bandit algorithms have been recently studied under the federated learning setting to satisfy the demand of keeping data decentralized and pushing the learning of bandit models to the client side. But limited by the required communication efficiency, existing solutions are restricted to linear models to exploit their closed-form solutions for parameter estimation. Such a restricted model choice greatly hampers these algorithms' practical utility. In this paper, we take the first step to addressing this challenge by studying generalized linear bandit models under the federated learning setting. We propose a communication-efficient solution framework that employs online regression for local update and offline regression for global update. We rigorously proved, though the setting is more general and challenging, our algorithm can attain sub-linear rate in both regret and communication cost, which is also validated by our extensive empirical evaluations.

Yiling Jia, Weitong Zhang, Dongruo Zhou et al.

Thanks to the power of representation learning, neural contextual bandit algorithms demonstrate remarkable performance improvement against their classical counterparts. But because their exploration has to be performed in the entire neural network parameter space to obtain nearly optimal regret, the resulting computational cost is prohibitively high. We perturb the rewards when updating the neural network to eliminate the need of explicit exploration and the corresponding computational overhead. We prove that a $\tilde{O}(\tilde{d}\sqrt{T})$ regret upper bound is still achievable under standard regularity conditions, where $T$ is the number of rounds of interactions and $\tilde{d}$ is the effective dimension of a neural tangent kernel matrix. Extensive comparisons with several benchmark contextual bandit algorithms, including two recent neural contextual bandit models, demonstrate the effectiveness and computational efficiency of our proposed neural bandit algorithm.

Yiling Jia, Hongning Wang

Existing online learning to rank (OL2R) solutions are limited to linear models, which are incompetent to capture possible non-linear relations between queries and documents. In this work, to unleash the power of representation learning in OL2R, we propose to directly learn a neural ranking model from users' implicit feedback (e.g., clicks) collected on the fly. We focus on RankNet and LambdaRank, due to their great empirical success and wide adoption in offline settings, and control the notorious explore-exploit trade-off based on the convergence analysis of neural networks using neural tangent kernel. Specifically, in each round of result serving, exploration is only performed on document pairs where the predicted rank order between the two documents is uncertain; otherwise, the ranker's predicted order will be followed in result ranking. We prove that under standard assumptions our OL2R solution achieves a gap-dependent upper regret bound of $O(\log^2(T))$, in which the regret is defined on the total number of mis-ordered pairs over $T$ rounds. Comparisons against an extensive set of state-of-the-art OL2R baselines on two public learning to rank benchmark datasets demonstrate the effectiveness of the proposed solution.

Yiling Jia, Hongning Wang

Online learning to rank (OL2R) has attracted great research interests in recent years, thanks to its advantages in avoiding expensive relevance labeling as required in offline supervised ranking model learning. Such a solution explores the unknowns (e.g., intentionally present selected results on top positions) to improve its relevance estimation. This however triggers concerns on its ranking fairness: different groups of items might receive differential treatments during the course of OL2R. But existing fair ranking solutions usually require the knowledge of result relevance or a performing ranker beforehand, which contradicts with the setting of OL2R and thus cannot be directly applied to guarantee fairness. In this work, we propose a general framework to achieve fairness defined by group exposure in OL2R. The key idea is to calibrate exploration and exploitation for fairness control, relevance learning and online ranking quality. In particular, when the model is exploring a set of results for relevance feedback, we confine the exploration within a subset of random permutations, where fairness across groups is maintained while the feedback is still unbiased. Theoretically we prove such a strategy introduces minimum distortion in OL2R's regret to obtain fairness. Extensive empirical analysis is performed on two public learning to rank benchmark datasets to demonstrate the effectiveness of the proposed solution compared to existing fair OL2R solutions.

Lu Lin, Ethan Blaser, Hongning Wang

Graph Convolutional Networks (GCNs) have fueled a surge of research interest due to their encouraging performance on graph learning tasks, but they are also shown vulnerability to adversarial attacks. In this paper, an effective graph structural attack is investigated to disrupt graph spectral filters in the Fourier domain, which are the theoretical foundation of GCNs. We define the notion of spectral distance based on the eigenvalues of graph Laplacian to measure the disruption of spectral filters. We realize the attack by maximizing the spectral distance and propose an efficient approximation to reduce the time complexity brought by eigen-decomposition. The experiments demonstrate the remarkable effectiveness of the proposed attack in both black-box and white-box settings for both test-time evasion attacks and training-time poisoning attacks. Our qualitative analysis suggests the connection between the imposed spectral changes in the Fourier domain and the attack behavior in the spatial domain, which provides empirical evidence that maximizing spectral distance is an effective way to change the graph structural property and thus disturb the frequency components for graph filters to affect the learning of GCNs.

Aobo Yang, Nan Wang, Renqin Cai et al.

As recommendation is essentially a comparative (or ranking) process, a good explanation should illustrate to users why an item is believed to be better than another, i.e., comparative explanations about the recommended items. Ideally, after reading the explanations, a user should reach the same ranking of items as the system's. Unfortunately, little research attention has yet been paid on such comparative explanations. In this work, we develop an extract-and-refine architecture to explain the relative comparisons among a set of ranked items from a recommender system. For each recommended item, we first extract one sentence from its associated reviews that best suits the desired comparison against a set of reference items. Then this extracted sentence is further articulated with respect to the target user through a generative model to better explain why the item is recommended. We design a new explanation quality metric based on BLEU to guide the end-to-end training of the extraction and refinement components, which avoids generation of generic content. Extensive offline evaluations on two large recommendation benchmark datasets and serious user studies against an array of state-of-the-art explainable recommendation algorithms demonstrate the necessity of comparative explanations and the effectiveness of our solution.