Video-KTR: Reinforcing Video Reasoning via Key Token Attribution

Reinforcement learning (RL) has shown strong potential for enhancing reasoning in multimodal large language models, yet existing video reasoning methods often rely on coarse sequence-level rewards or single-factor token selection, neglecting fine-grained links among visual inputs, temporal dynamics, and linguistic outputs, limiting both accuracy and interpretability. We propose Video-KTR, a modality-aware policy shaping framework that performs selective, token-level RL by combining three attribution signals: (1) visual-aware tokens identified via counterfactual masking to reveal perceptual dependence; (2) temporal-aware tokens detected through frame shuffling to expose temporal sensitivity; and (3) high-entropy tokens signaling predictive uncertainty. By reinforcing only these key tokens, Video-KTR focuses learning on semantically informative, modality-sensitive content while filtering out low-value tokens. Across five challenging benchmarks, Video-KTR achieves state-of-the-art or highly competitive results, achieving 42.7\% on Video-Holmes (surpassing GPT-4o) with consistent gains on both reasoning and general video understanding tasks. Ablation studies verify the complementary roles of the attribution signals and the robustness of targeted token-level updates. Overall, Video-KTR improves accuracy and interpretability, offering a simple, drop-in extension to RL for complex video reasoning. Our code and models are available at https://github.com/zywang0104/Video-KTR.

The Side Effects of Being Smart: Safety Risks in MLLMs' Multi-Image Reasoning

As Multimodal Large Language Models (MLLMs) acquire stronger reasoning capabilities to handle complex, multi-image instructions, this advancement may pose new safety risks. We study this problem by introducing MIR-SafetyBench, the first benchmark focused on multi-image reasoning safety, which consists of 2,676 instances across a taxonomy of 9 multi-image relations. Our extensive evaluations on 19 MLLMs reveal a troubling trend: models with more advanced multi-image reasoning can be more vulnerable on MIR-SafetyBench. Beyond attack success rates, we find that many responses labeled as safe are superficial, often driven by misunderstanding or evasive, non-committal replies. We further observe that unsafe generations exhibit lower attention entropy than safe ones on average. This internal signature suggests a possible risk that models may over-focus on task solving while neglecting safety constraints. Our code and data are available at https://github.com/thu-coai/MIR-SafetyBench.

SafeRedir: Prompt Embedding Redirection for Robust Unlearning in Image Generation Models

Image generation models (IGMs), while capable of producing impressive and creative content, often memorize a wide range of undesirable concepts from their training data, leading to the reproduction of unsafe content such as NSFW imagery and copyrighted artistic styles. Such behaviors pose persistent safety and compliance risks in real-world deployments and cannot be reliably mitigated by post-hoc filtering, owing to the limited robustness of such mechanisms and a lack of fine-grained semantic control. Recent unlearning methods seek to erase harmful concepts at the model level, which exhibit the limitations of requiring costly retraining, degrading the quality of benign generations, or failing to withstand prompt paraphrasing and adversarial attacks. To address these challenges, we introduce SafeRedir, a lightweight inference-time framework for robust unlearning via prompt embedding redirection. Without modifying the underlying IGMs, SafeRedir adaptively routes unsafe prompts toward safe semantic regions through token-level interventions in the embedding space. The framework comprises two core components: a latent-aware multi-modal safety classifier for identifying unsafe generation trajectories, and a token-level delta generator for precise semantic redirection, equipped with auxiliary predictors for token masking and adaptive scaling to localize and regulate the intervention. Empirical results across multiple representative unlearning tasks demonstrate that SafeRedir achieves effective unlearning capability, high semantic and perceptual preservation, robust image quality, and enhanced resistance to adversarial attacks. Furthermore, SafeRedir generalizes effectively across a variety of diffusion backbones and existing unlearned models, validating its plug-and-play compatibility and broad applicability. Code and data are available at https://github.com/ryliu68/SafeRedir.

Revisiting the Reliability of Language Models in Instruction-Following

Advanced LLMs have achieved near-ceiling instruction-following accuracy on benchmarks such as IFEval. However, these impressive scores do not necessarily translate to reliable services in real-world use, where users often vary their phrasing, contextual framing, and task formulations. In this paper, we study nuance-oriented reliability: whether models exhibit consistent competence across cousin prompts that convey analogous user intents but with subtle nuances. To quantify this, we introduce a new metric, reliable@k, and develop an automated pipeline that generates high-quality cousin prompts via data augmentation. Building upon this, we construct IFEval++ for systematic evaluation. Across 20 proprietary and 26 open-source LLMs, we find that current models exhibit substantial insufficiency in nuance-oriented reliability -- their performance can drop by up to 61.8% with nuanced prompt modifications. What's more, we characterize it and explore three potential improvement recipes. Our findings highlight nuance-oriented reliability as a crucial yet underexplored next step toward more dependable and trustworthy LLM behavior. Our code and benchmark are accessible: https://github.com/jianshuod/IFEval-pp.

Spatial Preference Rewarding for MLLMs Spatial Understanding

Multimodal large language models~(MLLMs) have demonstrated promising spatial understanding capabilities, such as referencing and grounding object descriptions. Despite their successes, MLLMs still fall short in fine-grained spatial perception abilities, such as generating detailed region descriptions or accurately localizing objects. Additionally, they often fail to respond to the user's requirements for desired fine-grained spatial understanding. This issue might arise because existing approaches primarily focus on tuning MLLMs to model pre-annotated instruction data to inject spatial knowledge, without direct supervision of MLLMs' actual responses. We address this issue by SPR, a Spatial Preference Rewarding~(SPR) approach that enhances MLLMs' spatial capabilities by rewarding MLLMs' detailed responses with precise object localization over vague or inaccurate responses. With randomly selected image regions and region descriptions from MLLMs, SPR introduces semantic and localization scores to comprehensively evaluate the text quality and localization quality in MLLM-generated descriptions. We also refine the MLLM descriptions with better localization accuracy and pair the best-scored refinement with the initial descriptions of the lowest score for direct preference optimization, thereby enhancing fine-grained alignment with visual input. Extensive experiments over standard referring and grounding benchmarks show that SPR improves MLLM spatial understanding capabilities effectively with minimal overhead in training. Data and code will be released at https://github.com/hanqiu-hq/SPR

Speculating LLMs' Chinese Training Data Pollution from Their Tokens

Tokens are basic elements in the datasets for LLM training. It is well-known that many tokens representing Chinese phrases in the vocabulary of GPT (4o/4o-mini/o1/o3/4.5/4.1/o4-mini) are indicating contents like pornography or online gambling. Based on this observation, our goal is to locate Polluted Chinese (PoC) tokens in LLMs and study the relationship between PoC tokens' existence and training data. (1) We give a formal definition and taxonomy of PoC tokens based on the GPT's vocabulary. (2) We build a PoC token detector via fine-tuning an LLM to label PoC tokens in vocabularies by considering each token's both semantics and related contents from the search engines. (3) We study the speculation on the training data pollution via PoC tokens' appearances (token ID). Experiments on GPT and other 23 LLMs indicate that tokens widely exist while GPT's vocabulary behaves the worst: more than 23% long Chinese tokens (i.e., a token with more than two Chinese characters) are either porn or online gambling. We validate the accuracy of our speculation method on famous pre-training datasets like C4 and Pile. Then, considering GPT-4o, we speculate that the ratio of "Yui Hatano" related webpages in GPT-4o's training data is around 0.5%.

Exploring Multimodal Challenges in Toxic Chinese Detection: Taxonomy, Benchmark, and Findings

Detecting toxic content using language models is important but challenging. While large language models (LLMs) have demonstrated strong performance in understanding Chinese, recent studies show that simple character substitutions in toxic Chinese text can easily confuse the state-of-the-art (SOTA) LLMs. In this paper, we highlight the multimodal nature of Chinese language as a key challenge for deploying LLMs in toxic Chinese detection. First, we propose a taxonomy of 3 perturbation strategies and 8 specific approaches in toxic Chinese content. Then, we curate a dataset based on this taxonomy, and benchmark 9 SOTA LLMs (from both the US and China) to assess if they can detect perturbed toxic Chinese text. Additionally, we explore cost-effective enhancement solutions like in-context learning (ICL) and supervised fine-tuning (SFT). Our results reveal two important findings. (1) LLMs are less capable of detecting perturbed multimodal Chinese toxic contents. (2) ICL or SFT with a small number of perturbed examples may cause the LLMs "overcorrect'': misidentify many normal Chinese contents as toxic.

BitHydra: Towards Bit-flip Inference Cost Attack against Large Language Models

Large language models (LLMs) have shown impressive capabilities across a wide range of applications, but their ever-increasing size and resource demands make them vulnerable to inference cost attacks, where attackers induce victim LLMs to generate the longest possible output content. In this paper, we revisit existing inference cost attacks and reveal that these methods can hardly produce large-scale malicious effects since they are self-targeting, where attackers are also the users and therefore have to execute attacks solely through the inputs, whose generated content will be charged by LLMs and can only directly influence themselves. Motivated by these findings, this paper introduces a new type of inference cost attacks (dubbed 'bit-flip inference cost attack') that target the victim model itself rather than its inputs. Specifically, we design a simple yet effective method (dubbed 'BitHydra') to effectively flip critical bits of model parameters. This process is guided by a loss function designed to suppress <EOS> token's probability with an efficient critical bit search algorithm, thus explicitly defining the attack objective and enabling effective optimization. We evaluate our method on 11 LLMs ranging from 1.5B to 14B parameters under both int8 and float16 settings. Experimental results demonstrate that with just 4 search samples and as few as 3 bit flips, BitHydra can force 100% of test prompts to reach the maximum generation length (e.g., 2048 tokens) on representative LLMs such as LLaMA3, highlighting its efficiency, scalability, and strong transferability across unseen inputs.

ShieldVLM: Safeguarding the Multimodal Implicit Toxicity via Deliberative Reasoning with LVLMs

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.

Holmes: Automated Fact Check with Large Language Models

The rise of Internet connectivity has accelerated the spread of disinformation, threatening societal trust, decision-making, and national security. Disinformation has evolved from simple text to complex multimodal forms combining images and text, challenging existing detection methods. Traditional deep learning models struggle to capture the complexity of multimodal disinformation. Inspired by advances in AI, this study explores using Large Language Models (LLMs) for automated disinformation detection. The empirical study shows that (1) LLMs alone cannot reliably assess the truthfulness of claims; (2) providing relevant evidence significantly improves their performance; (3) however, LLMs cannot autonomously search for accurate evidence. To address this, we propose Holmes, an end-to-end framework featuring a novel evidence retrieval method that assists LLMs in collecting high-quality evidence. Our approach uses (1) LLM-powered summarization to extract key information from open sources and (2) a new algorithm and metrics to evaluate evidence quality. Holmes enables LLMs to verify claims and generate justifications effectively. Experiments show Holmes achieves 88.3% accuracy on two open-source datasets and 90.2% in real-time verification tasks. Notably, our improved evidence retrieval boosts fact-checking accuracy by 30.8% over existing methods