We present PersonaConvBench, a large-scale benchmark for evaluating personalized reasoning and generation in multi-turn conversations with large language models (LLMs). Unlike existing work that focuses on either personalization or conversational structure in isolation, PersonaConvBench integrates both, offering three core tasks: sentence classification, impact regression, and user-centric text generation across ten diverse Reddit-based domains. This design enables systematic analysis of how personalized conversational context shapes LLM outputs in realistic multi-user scenarios. We benchmark several commercial and open-source LLMs under a unified prompting setup and observe that incorporating personalized history yields substantial performance improvements, including a 198 percent relative gain over the best non-conversational baseline in sentiment classification. By releasing PersonaConvBench with evaluations and code, we aim to support research on LLMs that adapt to individual styles, track long-term context, and produce contextually rich, engaging responses.

Ensuring the safety of large language models (LLMs) is critical for responsible deployment, yet existing evaluations often prioritize performance over identifying failure modes. We introduce Phare, a multilingual diagnostic framework to probe and evaluate LLM behavior across three critical dimensions: hallucination and reliability, social biases, and harmful content generation. Our evaluation of 17 state-of-the-art LLMs reveals patterns of systematic vulnerabilities across all safety dimensions, including sycophancy, prompt sensitivity, and stereotype reproduction. By highlighting these specific failure modes rather than simply ranking models, Phare provides researchers and practitioners with actionable insights to build more robust, aligned, and trustworthy language systems.

Concept erasing has recently emerged as an effective paradigm to prevent text-to-image diffusion models from generating visually undesirable or even harmful content. However, current removal methods heavily rely on manually crafted text prompts, making it challenging to achieve a high erasure (efficacy) while minimizing the impact on other benign concepts (usability). In this paper, we attribute the limitations to the inherent gap between the text and image modalities, which makes it hard to transfer the intricately entangled concept knowledge from text prompts to the image generation process. To address this, we propose a novel solution by directly integrating visual supervision into the erasure process, introducing the first text-image Collaborative Concept Erasing (Co-Erasing) framework. Specifically, Co-Erasing describes the concept jointly by text prompts and the corresponding undesirable images induced by the prompts, and then reduces the generating probability of the target concept through negative guidance. This approach effectively bypasses the knowledge gap between text and image, significantly enhancing erasure efficacy. Additionally, we design a text-guided image concept refinement strategy that directs the model to focus on visual features most relevant to the specified text concept, minimizing disruption to other benign concepts. Finally, comprehensive experiments suggest that Co-Erasing outperforms state-of-the-art erasure approaches significantly with a better trade-off between efficacy and usability. Codes are available at https://github.com/Ferry-Li/Co-Erasing.

Recent studies have uncovered a troubling vulnerability in the fine-tuning stage of large language models (LLMs): even fine-tuning on entirely benign datasets can lead to a significant increase in the harmfulness of LLM outputs. Building on this finding, our red teaming study takes this threat one step further by developing a more effective attack. Specifically, we analyze and identify samples within benign datasets that contribute most to safety degradation, then fine-tune LLMs exclusively on these samples. We approach this problem from an outlier detection perspective and propose Self-Inf-N, to detect and extract outliers for fine-tuning. Our findings reveal that fine-tuning LLMs on 100 outlier samples selected by Self-Inf-N in the benign datasets severely compromises LLM safety alignment. Extensive experiments across seven mainstream LLMs demonstrate that our attack exhibits high transferability across different architectures and remains effective in practical scenarios. Alarmingly, our results indicate that most existing mitigation strategies fail to defend against this attack, underscoring the urgent need for more robust alignment safeguards. Codes are available at https://github.com/GuanZihan/Benign-Samples-Matter.

Automated fact-checking (AFC) still falters on claims that are time-sensitive, entity-ambiguous, or buried beneath noisy search-engine results. We present PASS-FC, a Progressive and Adaptive Search Scheme for Fact Checking. Each atomic claim is first grounded with a precise time span and disambiguated entity descriptors. An adaptive search loop then issues structured queries, filters domains through credible-source selection, and expands queries cross-lingually; when necessary, a lightweight reflection routine restarts the loop. Experiments on six benchmark--covering general knowledge, scientific literature, real-world events, and ten languages--show that PASS-FC consistently outperforms prior systems, even those powered by larger backbone LLMs. On the multilingual X-FACT set, performance of different languages partially correlates with typological closeness to English, and forcing the model to reason in low-resource languages degrades accuracy. Ablations highlight the importance of temporal grounding and the adaptive search scheme, while detailed analysis shows that cross-lingual retrieval contributes genuinely new evidence. Code and full results will be released to facilitate further research.

Multimodal out-of-context (OOC) misinformation is misinformation that repurposes real images with unrelated or misleading captions. Detecting such misinformation is challenging because it requires resolving the context of the claim before checking for misinformation. Many current methods, including LLMs and LVLMs, do not perform this contextualization step. LLMs hallucinate in absence of context or parametric knowledge. In this work, we propose a graph-based method that evaluates the consistency between the image and the caption by constructing two graph representations: an evidence graph, derived from online textual evidence, and a claim graph, from the claim in the caption. Using graph neural networks (GNNs) to encode and compare these representations, our framework then evaluates the truthfulness of image-caption pairs. We create datasets for our graph-based method, evaluate and compare our baseline model against popular LLMs on the misinformation detection task. Our method scores $93.05\%$ detection accuracy on the evaluation set and outperforms the second-best performing method (an LLM) by $2.82\%$, making a case for smaller and task-specific methods.

LLMs have made impressive progress, but their growing capabilities also expose them to highly flexible jailbreaking attacks designed to bypass safety alignment. While many existing defenses focus on known types of attacks, it is more critical to prepare LLMs for unseen attacks that may arise during deployment. To address this, we propose a lifelong safety alignment framework that enables LLMs to continuously adapt to new and evolving jailbreaking strategies. Our framework introduces a competitive setup between two components: a Meta-Attacker, trained to actively discover novel jailbreaking strategies, and a Defender, trained to resist them. To effectively warm up the Meta-Attacker, we first leverage the GPT-4o API to extract key insights from a large collection of jailbreak-related research papers. Through iterative training, the first iteration Meta-Attacker achieves a 73% attack success rate (ASR) on RR and a 57% transfer ASR on LAT using only single-turn attacks. Meanwhile, the Defender progressively improves its robustness and ultimately reduces the Meta-Attacker's success rate to just 7%, enabling safer and more reliable deployment of LLMs in open-ended environments. The code is available at https://github.com/sail-sg/LifelongSafetyAlignment.

Reasoning models have recently attracted significant attention, especially for tasks that involve complex inference. Their strengths exemplify the System II paradigm (slow, structured thinking), contrasting with the System I (rapid, heuristic-driven). Yet, does slower reasoning necessarily lead to greater truthfulness? Our findings suggest otherwise. In this study, we present the first systematic investigation of distortions associated with System I and System II reasoning in multimodal contexts. We demonstrate that slower reasoning models, when presented with incomplete or misleading visual inputs, are more likely to fabricate plausible yet false details to support flawed reasoning -- a phenomenon we term the "Mirage of Multimodality". To examine this, we constructed a 5,000-sample hierarchical prompt dataset annotated by 50 human participants. These prompts gradually increase in complexity, revealing a consistent pattern: slower reasoning models tend to employ depth-first thinking (delving deeper into incorrect premises), whereas faster chat models favor breadth-first inference, exhibiting greater caution under uncertainty. Our results highlight a critical vulnerability of slower reasoning models: although highly effective in structured domains such as mathematics, it becomes brittle when confronted with ambiguous multimodal inputs.

Effective generalization in language models depends critically on the diversity of their training data. Yet existing diversity metrics often fall short of this goal, relying on surface-level heuristics that are decoupled from model behavior. This motivates us to ask: What kind of diversity in training data actually drives generalization in language models -- and how can we measure and amplify it? Through large-scale empirical analyses spanning over 300 training runs, carefully controlled for data scale and quality, we show that data diversity can be a strong predictor of generalization in LLM reasoning -- as measured by average model performance on unseen out-of-distribution benchmarks. We introduce G-Vendi, a metric that quantifies diversity via the entropy of model-induced gradients. Despite using a small off-the-shelf proxy model for gradients, G-Vendi consistently outperforms alternative measures, achieving strong correlation (Spearman's $ρ\approx 0.9$) with out-of-distribution (OOD) performance on both natural language inference (NLI) and math reasoning tasks. Building on this insight, we present Prismatic Synthesis, a framework for generating diverse synthetic data by targeting underrepresented regions in gradient space. Experimental results show that Prismatic Synthesis consistently improves model performance as we scale synthetic data -- not just on in-distribution test but across unseen, out-of-distribution benchmarks -- significantly outperforming state-of-the-art models that rely on 20 times larger data generator than ours. For example, PrismMath-7B, our model distilled from a 32B LLM, outperforms R1-Distill-Qwen-7B -- the same base model trained on proprietary data generated by 671B R1 -- on 6 out of 7 challenging benchmarks.

Large language models (LLMs) have been widely integrated into information retrieval to advance traditional techniques. However, effectively enabling LLMs to seek accurate knowledge in complex tasks remains a challenge due to the complexity of multi-hop queries as well as the irrelevant retrieved content. To address these limitations, we propose EXSEARCH, an agentic search framework, where the LLM learns to retrieve useful information as the reasoning unfolds through a self-incentivized process. At each step, the LLM decides what to retrieve (thinking), triggers an external retriever (search), and extracts fine-grained evidence (recording) to support next-step reasoning. To enable LLM with this capability, EXSEARCH adopts a Generalized Expectation-Maximization algorithm. In the E-step, the LLM generates multiple search trajectories and assigns an importance weight to each; the M-step trains the LLM on them with a re-weighted loss function. This creates a self-incentivized loop, where the LLM iteratively learns from its own generated data, progressively improving itself for search. We further theoretically analyze this training process, establishing convergence guarantees. Extensive experiments on four knowledge-intensive benchmarks show that EXSEARCH substantially outperforms baselines, e.g., +7.8% improvement on exact match score. Motivated by these promising results, we introduce EXSEARCH-Zoo, an extension that extends our method to broader scenarios, to facilitate future work.