Zhuoran Jin

Lu Wang, Zhuoran Jin, Yupu Hao et al.

Multimodal large language models (MLLMs) have shown strong performance on offline video understanding, but most are limited to offline inference or have weak online reasoning, making multi-turn interaction over continuously arriving video streams difficult. Existing streaming methods typically use an interleaved perception-generation paradigm, which prevents concurrent perception and generation and leads to early memory decay as streams grow, hurting long-range dependency modeling. We propose Think While Watching, a memory-anchored streaming video reasoning framework that preserves continuous segment-level memory during multi-turn interaction. We build a three-stage, multi-round chain-of-thought dataset and adopt a stage-matched training strategy, while enforcing strict causality through a segment-level streaming causal mask and streaming positional encoding. During inference, we introduce an efficient pipeline that overlaps watching and thinking and adaptively selects the best attention backend. Under both single-round and multi-round streaming input protocols, our method achieves strong results. Built on Qwen3-VL, it improves single-round accuracy by 2.6% on StreamingBench and by 3.79% on OVO-Bench. In the multi-round setting, it maintains performance while reducing output tokens by 56%. Code is available at: https://github.com/wl666hhh/Think_While_Watching/

Hongbang Yuan, Zhuoran Jin, Pengfei Cao et al.

LLM have achieved success in many fields but still troubled by problematic content in the training corpora. LLM unlearning aims at reducing their influence and avoid undesirable behaviours. However, existing unlearning methods remain vulnerable to adversarial queries and the unlearned knowledge resurfaces after the manually designed attack queries. As part of a red-team effort to proactively assess the vulnerabilities of unlearned models, we design Dynamic Unlearning Attack (DUA), a dynamic and automated framework to attack these models and evaluate their robustness. It optimizes adversarial suffixes to reintroduce the unlearned knowledge in various scenarios. We find that unlearned knowledge can be recovered in $55.2\%$ of the questions, even without revealing the unlearned model's parameters. In response to this vulnerability, we propose Latent Adversarial Unlearning (LAU), a universal framework that effectively enhances the robustness of the unlearned process. It formulates the unlearning process as a min-max optimization problem and resolves it through two stages: an attack stage, where perturbation vectors are trained and added to the latent space of LLMs to recover the unlearned knowledge, and a defense stage, where previously trained perturbation vectors are used to enhance unlearned model's robustness. With our LAU framework, we obtain two robust unlearning methods, AdvGA and AdvNPO. We conduct extensive experiments across multiple unlearning benchmarks and various models, and demonstrate that they improve the unlearning effectiveness by over $53.5\%$, cause only less than a $11.6\%$ reduction in neighboring knowledge, and have almost no impact on the model's general capabilities.

Jiachun Li, Shaoping Huang, Zhuoran Jin et al.

Recent progress in the reasoning capabilities of multimodal large language models (MLLMs) has empowered them to address more complex tasks such as scientific analysis and mathematical reasoning. Despite their promise, MLLMs' reasoning abilities across different scenarios in real life remain largely unexplored and lack standardized benchmarks for evaluation. To address this gap, we introduce MMR-Life, a comprehensive benchmark designed to evaluate the diverse multimodal multi-image reasoning capabilities of MLLMs across real-life scenarios. MMR-Life consists of 2,646 multiple-choice questions based on 19,108 images primarily sourced from real-world contexts, comprehensively covering seven reasoning types: abductive, analogical, causal, deductive, inductive, spatial, and temporal. Unlike existing reasoning benchmarks, MMR-Life does not rely on domain-specific expertise but instead requires models to integrate information across multiple images and apply diverse reasoning abilities. The evaluation of 37 advanced models highlights the substantial challenge posed by MMR-Life. Even top models like GPT-5 achieve only 58% accuracy and display considerable variance in performance across reasoning types. Moreover, we analyze the reasoning paradigms of existing MLLMs, exploring how factors such as thinking length, reasoning method, and reasoning type affect their performance. In summary, MMR-Life establishes a comprehensive foundation for evaluating, analyzing, and improving the next generation of multimodal reasoning systems.

Yupu Hao, Pengfei Cao, Zhuoran Jin et al.

Tool learning enables the Large Language Models (LLMs) to interact with the external environment by invoking tools, enriching the accuracy and capability scope of LLMs. However, previous works predominantly focus on improving model's tool-utilizing accuracy and the ability to generalize to new, unseen tools, excessively forcing LLMs to adjust specific tool-invoking pattern without considering the harm to model's general performance. This deviates from the actual applications and original intention of integrating tools to enhance model. To tackle this problem, we dissect the capability trade-offs by examining the hidden representation changes and the gradient-based importance score of model's components. Based on the analysis result, we propose a Component Importance-based Tool-utilizing ability Injection method (CITI). According to the gradient-based importance score of different components, it alleviates the capability conflicts caused by fine-tuning process by applying distinct training strategies to different components. CITI applies Mixture-Of-LoRA (MOLoRA) for important components. Meanwhile, it fine-tunes the parameters of few components deemed less important in the backbone of the LLM, while keeping other parameters frozen. CITI can effectively enhance the model's tool-utilizing capability without excessively compromising its general performance. Experimental results demonstrate that our approach achieves outstanding performance across a range of evaluation metrics.

Zhuoran Jin, Hongbang Yuan, Tianyi Men et al.

Despite the significant progress made by existing retrieval augmented language models (RALMs) in providing trustworthy responses and grounding in reliable sources, they often overlook effective alignment with human preferences. In the alignment process, reward models (RMs) act as a crucial proxy for human values to guide optimization. However, it remains unclear how to evaluate and select a reliable RM for preference alignment in RALMs. To this end, we propose RAG-RewardBench, the first benchmark for evaluating RMs in RAG settings. First, we design four crucial and challenging RAG-specific scenarios to assess RMs, including multi-hop reasoning, fine-grained citation, appropriate abstain, and conflict robustness. Then, we incorporate 18 RAG subsets, six retrievers, and 24 RALMs to increase the diversity of data sources. Finally, we adopt an LLM-as-a-judge approach to improve preference annotation efficiency and effectiveness, exhibiting a strong correlation with human annotations. Based on the RAG-RewardBench, we conduct a comprehensive evaluation of 45 RMs and uncover their limitations in RAG scenarios. Additionally, we also reveal that existing trained RALMs show almost no improvement in preference alignment, highlighting the need for a shift towards preference-aligned training.We release our benchmark and code publicly at https://huggingface.co/datasets/jinzhuoran/RAG-RewardBench/ for future work.

Tianyi Men, Zhuoran Jin, Pengfei Cao et al.

As Multimodal Large Language Models (MLLMs) advance, multimodal agents show promise in real-world tasks like web navigation and embodied intelligence. However, due to limitations in a lack of external feedback, these agents struggle with self-correction and generalization. A promising approach is to use reward models as external feedback, but there is no clear on how to select reward models for agents. Thus, there is an urgent need to build a reward bench targeted at agents. To address these challenges, we propose Agent-RewardBench, a benchmark designed to evaluate reward modeling ability in MLLMs. The benchmark is characterized by three key features: (1) Multiple dimensions and real-world agent scenarios evaluation. It covers perception, planning, and safety with 7 scenarios; (2) Step-level reward evaluation. It allows for the assessment of agent capabilities at the individual steps of a task, providing a more granular view of performance during the planning process; and (3) Appropriately difficulty and high-quality. We carefully sample from 10 diverse models, difficulty control to maintain task challenges, and manual verification to ensure the integrity of the data. Experiments demonstrate that even state-of-the-art multimodal models show limited performance, highlighting the need for specialized training in agent reward modeling. Code is available at github.

Yupu Hao, Pengfei Cao, Zhuoran Jin et al.

Personalized tool utilization is essential for aligning large language models (LLMs) with user preference in interaction scenarios with various tools. However, most of the current benchmarks primarily focus on either personalization of text generation or direct tool-utilizing, without considering both. In this work, we introduce a novel benchmark ETAPP for evaluating personalized tool invocation, establishing a sandbox environment, and a comprehensive dataset of 800 testing cases covering diverse user profiles. To improve the accuracy of our evaluation, we propose a key-point-based LLM evaluation method, mitigating biases in the LLM-as-a-judge system by manually annotating key points for each test case and providing them to LLM as the reference. Additionally, we evaluate the excellent LLMs and provide an in-depth analysis. Furthermore, we investigate the impact of different tool-invoking strategies on LLMs' personalization performance and the effects of fine-tuning in our task. The effectiveness of our preference-setting and key-point-based evaluation method is also validated. Our findings offer insights into improving personalized LLM agents. Our Code is available at https://github.com/hypasd-art/ETAPP.

Kejian Zhu, Shangqing Tu, Zhuoran Jin et al.

The development of large language models (LLMs) depends on trustworthy evaluation. However, most current evaluations rely on public benchmarks, which are prone to data contamination issues that significantly compromise fairness. Previous researches have focused on constructing dynamic benchmarks to address contamination. However, continuously building new benchmarks is costly and cyclical. In this work, we aim to tackle contamination by analyzing the mechanisms of contaminated models themselves. Through our experiments, we discover that the overestimation of contaminated models is likely due to parameters acquiring shortcut solutions in training. We further propose a novel method for identifying shortcut neurons through comparative and causal analysis. Building on this, we introduce an evaluation method called shortcut neuron patching to suppress shortcut neurons. Experiments validate the effectiveness of our approach in mitigating contamination. Additionally, our evaluation results exhibit a strong linear correlation with MixEval, a recently released trustworthy benchmark, achieving a Spearman coefficient ($ρ$) exceeding 0.95. This high correlation indicates that our method closely reveals true capabilities of the models and is trustworthy. We conduct further experiments to demonstrate the generalizability of our method across various benchmarks and hyperparameter settings. Code: https://github.com/GaryStack/Trustworthy-Evaluation

Zhuoran Jin, Pengfei Cao, Yubo Chen et al.

Retrieval-augmented language models (RALMs) have demonstrated significant potential in refining and expanding their internal memory by retrieving evidence from external sources. However, RALMs will inevitably encounter knowledge conflicts when integrating their internal memory with external sources. Knowledge conflicts can ensnare RALMs in a tug-of-war between knowledge, limiting their practical applicability. In this paper, we focus on exploring and resolving knowledge conflicts in RALMs. First, we present an evaluation framework for assessing knowledge conflicts across various dimensions. Then, we investigate the behavior and preference of RALMs from the following two perspectives: (1) Conflicts between internal memory and external sources: We find that stronger RALMs emerge with the Dunning-Kruger effect, persistently favoring their faulty internal memory even when correct evidence is provided. Besides, RALMs exhibit an availability bias towards common knowledge; (2) Conflicts between truthful, irrelevant and misleading evidence: We reveal that RALMs follow the principle of majority rule, leaning towards placing trust in evidence that appears more frequently. Moreover, we find that RALMs exhibit confirmation bias, and are more willing to choose evidence that is consistent with their internal memory. To solve the challenge of knowledge conflicts, we propose a method called Conflict-Disentangle Contrastive Decoding (CD2) to better calibrate the model's confidence. Experimental results demonstrate that our CD2 can effectively resolve knowledge conflicts in RALMs.

Zhuoran Jin, Pengfei Cao, Hongbang Yuan et al.

Recently, retrieval augmentation and tool augmentation have demonstrated a remarkable capability to expand the internal memory boundaries of language models (LMs) by providing external context. However, internal memory and external context inevitably clash, leading to knowledge conflicts within LMs. In this paper, we aim to interpret the mechanism of knowledge conflicts through the lens of information flow, and then mitigate conflicts by precise interventions at the pivotal point. We find there are some attention heads with opposite effects in the later layers, where memory heads can recall knowledge from internal memory, and context heads can retrieve knowledge from external context. Moreover, we reveal that the pivotal point at which knowledge conflicts emerge in LMs is the integration of inconsistent information flows by memory heads and context heads. Inspired by the insights, we propose a novel method called Pruning Head via PatH PatcHing (PH3), which can efficiently mitigate knowledge conflicts by pruning conflicting attention heads without updating model parameters. PH3 can flexibly control eight LMs to use internal memory ($\uparrow$ 44.0%) or external context ($\uparrow$ 38.5%). Moreover, PH3 can also improve the performance of LMs on open-domain QA tasks. We also conduct extensive experiments to demonstrate the cross-model, cross-relation, and cross-format generalization of our method.

Jiachun Li, Pengfei Cao, Chenhao Wang et al.

Large language models exhibit high-level commonsense reasoning abilities, especially with enhancement methods like Chain-of-Thought (CoT). However, we find these CoT-like methods lead to a considerable number of originally correct answers turning wrong, which we define as the Toxic CoT problem. To interpret and mitigate this problem, we first utilize attribution tracing and causal tracing methods to probe the internal working mechanism of the LLM during CoT reasoning. Through comparisons, we prove that the model exhibits information loss from the question over the shallow attention layers when generating rationales or answers. Based on the probing findings, we design a novel method called RIDERS (Residual decodIng and sERial-position Swap), which compensates for the information deficit in the model from both decoding and serial-position perspectives. Through extensive experiments on multiple commonsense reasoning benchmarks, we validate that this method not only significantly eliminates Toxic CoT problems (decreased by 23.6%), but also effectively improves the model's overall commonsense reasoning performance (increased by 5.5%).

Zhitao He, Pengfei Cao, Chenhao Wang et al.

With the development of deep learning, natural language processing technology has effectively improved the efficiency of various aspects of the traditional judicial industry. However, most current efforts focus on tasks within individual judicial stages, making it difficult to handle complex tasks that span multiple stages. As the autonomous agents powered by large language models are becoming increasingly smart and able to make complex decisions in real-world settings, offering new insights for judicial intelligence. In this paper, (1) we propose a novel multi-agent framework, AgentsCourt, for judicial decision-making. Our framework follows the classic court trial process, consisting of court debate simulation, legal resources retrieval and decision-making refinement to simulate the decision-making of judge. (2) we introduce SimuCourt, a judicial benchmark that encompasses 420 Chinese judgment documents, spanning the three most common types of judicial cases. Furthermore, to support this task, we construct a large-scale legal knowledge base, Legal-KB, with multi-resource legal knowledge. (3) Extensive experiments show that our framework outperforms the existing advanced methods in various aspects, especially in generating legal articles, where our model achieves significant improvements of 8.6% and 9.1% F1 score in the first and second instance settings, respectively.

Jiachun Li, Pengfei Cao, Zhuoran Jin et al.

Inductive reasoning is an essential capability for large language models (LLMs) to achieve higher intelligence, which requires the model to generalize rules from observed facts and then apply them to unseen examples. We present MIRAGE, a synthetic dataset that addresses the limitations of previous work, specifically the lack of comprehensive evaluation and flexible test data. In it, we evaluate LLMs' capabilities in both the inductive and deductive stages, allowing for flexible variation in input distribution, task scenario, and task difficulty to analyze the factors influencing LLMs' inductive reasoning. Based on these multi-faceted evaluations, we demonstrate that the LLM is a poor rule-based reasoner. In many cases, when conducting inductive reasoning, they do not rely on a correct rule to answer the unseen case. From the perspectives of different prompting methods, observation numbers, and task forms, models tend to consistently conduct correct deduction without correct inductive rules. Besides, we find that LLMs are good neighbor-based reasoners. In the inductive reasoning process, the model tends to focus on observed facts that are close to the current test example in feature space. By leveraging these similar examples, the model maintains strong inductive capabilities within a localized region, significantly improving its deductive performance.

Hongbang Yuan, Pengfei Cao, Zhuoran Jin et al.

Large Language Models (LLMs) have shown impressive capabilities but still suffer from the issue of hallucinations. A significant type of this issue is the false premise hallucination, which we define as the phenomenon when LLMs generate hallucinated text when confronted with false premise questions. In this paper, we perform a comprehensive analysis of the false premise hallucination and elucidate its internal working mechanism: a small subset of attention heads (which we designate as false premise heads) disturb the knowledge extraction process, leading to the occurrence of false premise hallucination. Based on our analysis, we propose \textbf{FAITH} (\textbf{F}alse premise \textbf{A}ttention head constra\textbf{I}ining for mi\textbf{T}igating \textbf{H}allucinations), a novel and effective method to mitigate false premise hallucinations. It constrains the false premise attention heads during the model inference process. Impressively, extensive experiments demonstrate that constraining only approximately $1\%$ of the attention heads in the model yields a notable increase of nearly $20\%$ of model performance.

Kejian Zhu, Zhuoran Jin, Hongbang Yuan et al.

The sequential structure of videos poses a challenge to the ability of multimodal large language models (MLLMs) to locate multi-frame evidence and conduct multimodal reasoning. However, existing video benchmarks mainly focus on understanding tasks, which only require models to match frames mentioned in the question (hereafter referred to as "question frame") and perceive a few adjacent frames. To address this gap, we propose MMR-V: A Benchmark for Multimodal Deep Reasoning in Videos. The benchmark is characterized by the following features. (1) Long-range, multi-frame reasoning: Models are required to infer and analyze evidence frames that may be far from the question frame. (2) Beyond perception: Questions cannot be answered through direct perception alone but require reasoning over hidden information. (3) Reliability: All tasks are manually annotated, referencing extensive real-world user understanding to align with common perceptions. (4) Confusability: Carefully designed distractor annotation strategies to reduce model shortcuts. MMR-V consists of 317 videos and 1,257 tasks. Our experiments reveal that current models still struggle with multi-modal reasoning; even the best-performing model, o4-mini, achieves only 52.5% accuracy. Additionally, current reasoning enhancement strategies (Chain-of-Thought and scaling test-time compute) bring limited gains. Further analysis indicates that the CoT demanded for multi-modal reasoning differs from it in textual reasoning, which partly explains the limited performance gains. We hope that MMR-V can inspire further research into enhancing multi-modal reasoning capabilities.

Zhitao He, Pengfei Cao, Zhuoran Jin et al.

Event Causality Identification (ECI) refers to the detection of causal relations between events in texts. However, most existing studies focus on sentence-level ECI with high-resource languages, leaving more challenging document-level ECI (DECI) with low-resource languages under-explored. In this paper, we propose a Heterogeneous Graph Interaction Model with Multi-granularity Contrastive Transfer Learning (GIMC) for zero-shot cross-lingual document-level ECI. Specifically, we introduce a heterogeneous graph interaction network to model the long-distance dependencies between events that are scattered over a document. Then, to improve cross-lingual transferability of causal knowledge learned from the source language, we propose a multi-granularity contrastive transfer learning module to align the causal representations across languages. Extensive experiments show our framework outperforms the previous state-of-the-art model by 9.4% and 8.2% of average F1 score on monolingual and multilingual scenarios respectively. Notably, in the multilingual scenario, our zero-shot framework even exceeds GPT-3.5 with few-shot learning by 24.3% in overall performance.

Chenlong Zhang, Zhuoran Jin, Hongbang Yuan et al.

The widespread deployment of Large Language Models (LLMs) trained on massive, uncurated corpora has raised growing concerns about the inclusion of sensitive, copyrighted, or illegal content. This has led to increasing interest in LLM unlearning: the task of selectively removing specific information from a model without retraining from scratch or degrading overall utility. However, existing methods often rely on large-scale forget and retain datasets, and suffer from unnatural responses, poor generalization, or catastrophic utility loss. In this work, we propose Reinforcement UnLearning (RULE), an efficient framework that formulates unlearning as a refusal boundary optimization problem. RULE is trained with a small portion of the forget set and synthesized boundary queries, using a verifiable reward function that encourages safe refusal on forget--related queries while preserving helpful responses on permissible inputs. We provide both theoretical and empirical evidence demonstrating the effectiveness of RULE in achieving targeted unlearning without compromising model utility. Experimental results show that, with only $12%$ forget set and $8%$ synthesized boundary data, RULE outperforms existing baselines by up to $17.5%$ forget quality and $16.3%$ naturalness response while maintaining general utility, achieving forget--retain Pareto optimality. Remarkably, we further observe that RULE improves the naturalness of model outputs, enhances training efficiency, and exhibits strong generalization ability, generalizing refusal behavior to semantically related but unseen queries.

Suifeng Zhao, Zhuoran Jin, Sujian Li et al.

Retrieval-Augmented Generation (RAG) plays a vital role in the financial domain, powering applications such as real-time market analysis, trend forecasting, and interest rate computation. However, most existing RAG research in finance focuses predominantly on textual data, overlooking the rich visual content in financial documents, resulting in the loss of key analytical insights. To bridge this gap, we present FinRAGBench-V, a comprehensive visual RAG benchmark tailored for finance which effectively integrates multimodal data and provides visual citation to ensure traceability. It includes a bilingual retrieval corpus with 60,780 Chinese and 51,219 English pages, along with a high-quality, human-annotated question-answering (QA) dataset spanning heterogeneous data types and seven question categories. Moreover, we introduce RGenCite, an RAG baseline that seamlessly integrates visual citation with generation. Furthermore, we propose an automatic citation evaluation method to systematically assess the visual citation capabilities of Multimodal Large Language Models (MLLMs). Extensive experiments on RGenCite underscore the challenging nature of FinRAGBench-V, providing valuable insights for the development of multimodal RAG systems in finance.

Chenlong Zhang, Tong Zhou, Pengfei Cao et al.

The rapid proliferation of online news has posed significant challenges in tracking the continuous development of news topics. Traditional timeline summarization constructs a chronological summary of the events but often lacks the flexibility to meet the diverse granularity needs. To overcome this limitation, we introduce a new paradigm, Dynamic-granularity TimELine Summarization, (DTELS), which aims to construct adaptive timelines based on user instructions or requirements. This paper establishes a comprehensive benchmark for DTLES that includes: (1) an evaluation framework grounded in journalistic standards to assess the timeline quality across four dimensions: Informativeness, Granular Consistency, Factuality, and Coherence; (2) a large-scale, multi-source dataset with multiple granularity timeline annotations based on a consensus process to facilitate authority; (3) extensive experiments and analysis with two proposed solutions based on Large Language Models (LLMs) and existing state-of-the-art TLS methods. The experimental results demonstrate the effectiveness of LLM-based solutions. However, even the most advanced LLMs struggle to consistently generate timelines that are both informative and granularly consistent, highlighting the challenges of the DTELS task.

Jiachun Li, Pengfei Cao, Chenhao Wang et al.

Large language models (LLMs) sometimes demonstrate poor performance on knowledge-intensive tasks, commonsense reasoning is one of them. Researchers typically address these issues by retrieving related knowledge from knowledge graphs or employing self-enhancement methods to elicit knowledge in LLMs. However, noisy knowledge and invalid reasoning issues hamper their ability to answer questions accurately. To this end, we propose a novel method named eliciting, filtering and integrating knowledge in large language model (LINKED). In it, we design a reward model to filter out the noisy knowledge and take the marginal consistent reasoning module to reduce invalid reasoning. With our comprehensive experiments on two complex commonsense reasoning benchmarks, our method outperforms SOTA baselines (up to 9.0% improvement of accuracy). Besides, to measure the positive and negative impact of the injected knowledge, we propose a new metric called effectiveness-preservation score for the knowledge enhancement works. Finally, through extensive experiments, we conduct an in-depth analysis and find many meaningful conclusions about LLMs in commonsense reasoning tasks.

Zhuoran Jin, Hongbang Yuan, Kejian Zhu et al.

Reward models (RMs) play a critical role in aligning AI behaviors with human preferences, yet they face two fundamental challenges: (1) Modality Imbalance, where most RMs are mainly focused on text and image modalities, offering limited support for video, audio, and other modalities; and (2) Preference Rigidity, where training on fixed binary preference pairs fails to capture the complexity and diversity of personalized preferences. To address the above challenges, we propose Omni-Reward, a step toward generalist omni-modal reward modeling with support for free-form preferences, consisting of: (1) Evaluation: We introduce Omni-RewardBench, the first omni-modal RM benchmark with free-form preferences, covering nine tasks across five modalities including text, image, video, audio, and 3D; (2) Data: We construct Omni-RewardData, a multimodal preference dataset comprising 248K general preference pairs and 69K instruction-tuning pairs for training generalist omni-modal RMs; (3) Model: We propose Omni-RewardModel, which includes both discriminative and generative RMs, and achieves strong performance on Omni-RewardBench as well as other widely used reward modeling benchmarks.

Pengfei Cao, Yuheng Chen, Zhuoran Jin et al.

Large language models (LLMs) have learned vast amounts of factual knowledge through self-supervised pre-training on large-scale corpora. Meanwhile, LLMs have also demonstrated excellent multilingual capabilities, which can express the learned knowledge in multiple languages. However, the knowledge storage mechanism in LLMs still remains mysterious. Some researchers attempt to demystify the factual knowledge in LLMs from the perspective of knowledge neurons, and subsequently discover language-agnostic knowledge neurons that store factual knowledge in a form that transcends language barriers. However, the preliminary finding suffers from two limitations: 1) High Uncertainty in Localization Results. Existing study only uses a prompt-based probe to localize knowledge neurons for each fact, while LLMs cannot provide consistent answers for semantically equivalent queries. Thus, it leads to inaccurate localization results with high uncertainty. 2) Lack of Analysis in More Languages. The study only analyzes language-agnostic knowledge neurons on English and Chinese data, without exploring more language families and languages. Naturally, it limits the generalizability of the findings. To address aforementioned problems, we first construct a new benchmark called Rephrased Multilingual LAMA (RML-LAMA), which contains high-quality cloze-style multilingual parallel queries for each fact. Then, we propose a novel method named Multilingual Integrated Gradients with Uncertainty Estimation (MATRICE), which quantifies the uncertainty across queries and languages during knowledge localization. Extensive experiments show that our method can accurately localize language-agnostic knowledge neurons. We also further investigate the role of language-agnostic knowledge neurons in cross-lingual knowledge editing, knowledge enhancement and new knowledge injection.

Tianyi Men, Pengfei Cao, Zhuoran Jin et al.

With the development of large language models, they are widely used as agents in various fields. A key component of agents is memory, which stores vital information but is susceptible to jailbreak attacks. Existing research mainly focuses on single-agent attacks and shared memory attacks. However, real-world scenarios often involve independent memory. In this paper, we propose the Troublemaker Makes Chaos in Honest Town (TMCHT) task, a large-scale, multi-agent, multi-topology text-based attack evaluation framework. TMCHT involves one attacker agent attempting to mislead an entire society of agents. We identify two major challenges in multi-agent attacks: (1) Non-complete graph structure, (2) Large-scale systems. We attribute these challenges to a phenomenon we term toxicity disappearing. To address these issues, we propose an Adversarial Replication Contagious Jailbreak (ARCJ) method, which optimizes the retrieval suffix to make poisoned samples more easily retrieved and optimizes the replication suffix to make poisoned samples have contagious ability. We demonstrate the superiority of our approach in TMCHT, with 23.51%, 18.95%, and 52.93% improvements in line topology, star topology, and 100-agent settings. Encourage community attention to the security of multi-agent systems.

Tianyi Men, Pengfei Cao, Zhuoran Jin et al.

Planning, as the core module of agents, is crucial in various fields such as embodied agents, web navigation, and tool using. With the development of large language models (LLMs), some researchers treat large language models as intelligent agents to stimulate and evaluate their planning capabilities. However, the planning mechanism is still unclear. In this work, we focus on exploring the look-ahead planning mechanism in large language models from the perspectives of information flow and internal representations. First, we study how planning is done internally by analyzing the multi-layer perception (MLP) and multi-head self-attention (MHSA) components at the last token. We find that the output of MHSA in the middle layers at the last token can directly decode the decision to some extent. Based on this discovery, we further trace the source of MHSA by information flow, and we reveal that MHSA mainly extracts information from spans of the goal states and recent steps. According to information flow, we continue to study what information is encoded within it. Specifically, we explore whether future decisions have been encoded in advance in the representation of flow. We demonstrate that the middle and upper layers encode a few short-term future decisions to some extent when planning is successful. Overall, our research analyzes the look-ahead planning mechanisms of LLMs, facilitating future research on LLMs performing planning tasks.

Hongbang Yuan, Yubo Chen, Pengfei Cao et al.

Large language models (LLMs) have achieved remarkable success but still tend to generate factually erroneous responses, a phenomenon known as hallucination. A recent trend is to use preference learning to fine-tune models to align with factuality. However, existing work primarily evaluates fine-tuned models on in-domain (ID) datasets and the factuality on out-of-domain (OOD) datasets remains underexplored. In this paper, we conduct a comprehensive evaluation of the factuality of different models tuned by various preference learning algorithms and demonstrate that their performance on OOD datasets either increases minimally or decreases. Subsequently, we reveal that the main cause of model's failure to uphold factuality under a distribution shift is \textbf{under-alignment}, rather than \textbf{over-alignment}, by analyzing the token distribution shift of the models before and after tuning. Finally, we propose \textbf{APEFT} (\textbf{A}tomic \textbf{P}reference \textbf{E}nhanced \textbf{F}actuality \textbf{T}uning), a framework that enhances model's awareness of factuality at the granularity of individual facts. Extensive experiments demonstrate that APEFT improves model performance by an average of $\boldsymbol{3.45\%}$ on both ID and OOD datasets, which is highly effective.

Zhuoran Jin, Pengfei Cao, Chenhao Wang et al.

Large language models (LLMs) inevitably memorize sensitive, copyrighted, and harmful knowledge from the training corpus; therefore, it is crucial to erase this knowledge from the models. Machine unlearning is a promising solution for efficiently removing specific knowledge by post hoc modifying models. In this paper, we propose a Real-World Knowledge Unlearning benchmark (RWKU) for LLM unlearning. RWKU is designed based on the following three key factors: (1) For the task setting, we consider a more practical and challenging unlearning setting, where neither the forget corpus nor the retain corpus is accessible. (2) For the knowledge source, we choose 200 real-world famous people as the unlearning targets and show that such popular knowledge is widely present in various LLMs. (3) For the evaluation framework, we design the forget set and the retain set to evaluate the model's capabilities across various real-world applications. Regarding the forget set, we provide four four membership inference attack (MIA) methods and nine kinds of adversarial attack probes to rigorously test unlearning efficacy. Regarding the retain set, we assess locality and utility in terms of neighbor perturbation, general ability, reasoning ability, truthfulness, factuality, and fluency. We conduct extensive experiments across two unlearning scenarios, two models and six baseline methods and obtain some meaningful findings. We release our benchmark and code publicly at http://rwku-bench.github.io for future work.