Multimodal Large Language Models (MLLMs) trained on massive data may memorize sensitive personal information and photos, posing serious privacy risks. To mitigate this, MLLM unlearning methods are proposed, which fine-tune MLLMs to reduce the ``forget'' sensitive information. However, it remains unclear whether the knowledge has been truly forgotten or just hidden in the model. Therefore, we propose to study a novel problem of LLM unlearning attack, which aims to recover the unlearned knowledge of an unlearned LLM. To achieve the goal, we propose a novel framework Stealthy Unlearning Attack (SUA) framework that learns a universal noise pattern. When applied to input images, this noise can trigger the model to reveal unlearned content. While pixel-level perturbations may be visually subtle, they can be detected in the semantic embedding space, making such attacks vulnerable to potential defenses. To improve stealthiness, we introduce an embedding alignment loss that minimizes the difference between the perturbed and denoised image embeddings, ensuring the attack is semantically unnoticeable. Experimental results show that SUA can effectively recover unlearned information from MLLMs. Furthermore, the learned noise generalizes well: a single perturbation trained on a subset of samples can reveal forgotten content in unseen images. This indicates that knowledge reappearance is not an occasional failure, but a consistent behavior.

Parameter-efficient fine-tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA), offer an efficient way to adapt large language models with reduced computational costs. However, their performance is limited by the small number of trainable parameters. Recent work combines LoRA with the Mixture-of-Experts (MoE), i.e., LoRA-MoE, to enhance capacity, but two limitations remain in hindering the full exploitation of its potential: 1) the influence of downstream tasks when assigning expert numbers, and 2) the uniform rank assignment across all LoRA experts, which restricts representational diversity. To mitigate these gaps, we propose GuiLoMo, a fine-grained layer-wise expert numbers and ranks allocation strategy with GuidedSelection Vectors (GSVs). GSVs are learned via a prior bilevel optimization process to capture both model- and task-specific needs, and are then used to allocate optimal expert numbers and ranks. Experiments on three backbone models across diverse benchmarks show that GuiLoMo consistently achieves superior or comparable performance to all baselines. Further analysis offers key insights into how expert numbers and ranks vary across layers and tasks, highlighting the benefits of adaptive expert configuration. Our code is available at https://github.com/Liar406/Gui-LoMo.git.

Mosquito-borne diseases pose a major global health risk, requiring early detection and proactive control of breeding sites to prevent outbreaks. In this paper, we present VisT ext-Mosquito, a multimodal dataset that integrates visual and textual data to support automated detection, segmentation, and reasoning for mosquito breeding site analysis. The dataset includes 1,828 annotated images for object detection, 142 images for water surface segmentation, and natural language reasoning texts linked to each image. The YOLOv9s model achieves the highest precision of 0.92926 and mAP@50 of 0.92891 for object detection, while YOLOv11n-Seg reaches a segmentation precision of 0.91587 and mAP@50 of 0.79795. F or reasoning generation, we tested a range of large vision-language models (LVLMs) in both zero-shot and few-shot settings. Our fine-tuned Mosquito-LLaMA3-8B model achieved the best results, with a final loss of 0.0028, a BLEU score of 54.7, BERTScore of 0.91, and ROUGE-L of 0.85. This dataset and model framework emphasize the theme "Prevention is Better than Cure", showcasing how AI-based detection can proactively address mosquito-borne disease risks.

Large language models (LLMs) are increasingly used in psychological research and practice, yet traditional benchmarks reveal little about the values they express in real interaction. We introduce PAPERS, an output-based evaluation of the values LLMs prioritise in their text. Study 1 thematically analysed responses from eleven LLMs, identifying five recurring dimensions (Purposeful Contribution, Adaptive Growth, Positive Relationality, Ethical Integrity, and Robust Functionality) with Self-Actualised Autonomy appearing only under a hypothetical sentience prompt. These results suggest that LLMs are trained to prioritise humanistic and utility values as dual objectives of optimal functioning, a pattern supported by existing AI alignment and prioritisation frameworks. Study 2 operationalised PAPERS as a ranking instrument across the same eleven LLMs, yielding stable, non-random value priorities alongside systematic between-model differences. Hierarchical clustering distinguished "human-centric" models (e.g., ChatGPT-4o, Claude Sonnet 4) that prioritised relational/ethical values from "utility-driven" models (e.g., Llama 4, Gemini 2.5 Pro) that emphasised operational priorities. Study 3 benchmarked four LLMs against human judgements (N = 376) under matched prompts, finding near-perfect rank-order convergence (r = .97-.98) but moderate absolute agreement; among tested models, ChatGPT-4o showed the closest alignment with human ratings (ICC = .78). Humans also showed limited readiness to endorse sentient AI systems. Taken together, PAPERS enabled systematic value audits and revealed trade-offs with direct implications for deployment: human-centric models aligned more closely with human value judgments and appear better suited for humanistic psychological applications, whereas utility-driven models emphasised functional efficiency and may be more appropriate for instrumental or back-office tasks.

Recent studies have shown LLMs possess some ability to improve their responses when given external feedback. However, it remains unclear how effectively and thoroughly these models can incorporate extrinsic feedback. In an ideal scenario, if LLMs receive near-perfect and complete feedback, we would expect them to fully integrate the feedback and reach correct solutions. In this paper, we systematically investigate LLMs' ability to incorporate feedback by designing a controlled experimental environment. For each problem, a solver model attempts a solution, then a feedback generator with access to near-complete ground-truth answers produces targeted feedback, after which the solver tries again. We evaluate this pipeline across a diverse range of tasks, including math reasoning, knowledge reasoning, scientific reasoning, and general multi-domain evaluations with state-of-the-art language models including Claude 3.7 with extended thinking. Surprisingly, even under these near-ideal conditions, solver models consistently show resistance to feedback, a limitation that we term Feedback Friction. To mitigate this limitation, we experiment with sampling-based strategies like progressive temperature increases and explicit rejection of previously attempted incorrect answers, which yield improvements but still fail to help models achieve target performance. We analyze Feedback Friction and find that models' confidence on specific questions, measured by semantic entropy, predicts feedback resistance: high-confidence predictions remain resistant to external correction. We hope that highlighting this issue in LLMs will help future research in self-improvement.

Large language models have shown promise in clinical decision making, but current approaches struggle to localize and correct errors at specific steps of the reasoning process. This limitation is critical in medicine, where identifying and addressing reasoning errors is essential for accurate diagnosis and effective patient care. We introduce Med-PRM, a process reward modeling framework that leverages retrieval-augmented generation to verify each reasoning step against established medical knowledge bases. By verifying intermediate reasoning steps with evidence retrieved from clinical guidelines and literature, our model can precisely assess the reasoning quality in a fine-grained manner. Evaluations on five medical QA benchmarks and two open-ended diagnostic tasks demonstrate that Med-PRM achieves state-of-the-art performance, with improving the performance of base models by up to 13.50% using Med-PRM. Moreover, we demonstrate the generality of Med-PRM by integrating it in a plug-and-play fashion with strong policy models such as Meerkat, achieving over 80\% accuracy on MedQA for the first time using small-scale models of 8 billion parameters. Our code and data are available at: https://med-prm.github.io/

Though safety alignment has been applied to most large language models (LLMs), LLM service providers generally deploy a subsequent moderation as the external safety guardrail in real-world products. Existing moderators mainly practice a conventional full detection, which determines the harmfulness based on the complete LLM output, causing high service latency. Recent works pay more attention to partial detection where moderators oversee the generation midway and early stop the output if harmfulness is detected, but they directly apply moderators trained with the full detection paradigm to incomplete outputs, introducing a training-inference gap that lowers the performance. In this paper, we explore how to form a data-and-model solution that natively supports partial detection. For the data, we construct FineHarm, a dataset consisting of 29K prompt-response pairs with fine-grained annotations to provide reasonable supervision for token-level training. Then, we propose the streaming content monitor, which is trained with dual supervision of response- and token-level labels and can follow the output stream of LLM to make a timely judgment of harmfulness. Experiments show that SCM gains 0.95+ in macro F1 score that is comparable to full detection, by only seeing the first 18% of tokens in responses on average. Moreover, the SCM can serve as a pseudo-harmfulness annotator for improving safety alignment and lead to a higher harmlessness score than DPO.

Handwritten text recognition aims to convert visual input into machine-readable text, and it remains challenging due to the evolving and context-dependent nature of handwriting. Character sets change over time, and character frequency distributions shift across historical periods or regions, often causing models trained on broad, heterogeneous corpora to underperform on specific subsets. To tackle this, we propose a novel loss function that incorporates the Wasserstein distance between the character frequency distribution of the predicted text and a target distribution empirically derived from training data. By penalizing divergence from expected distributions, our approach enhances both accuracy and robustness under temporal and contextual intra-dataset shifts. Furthermore, we demonstrate that character distribution alignment can also improve existing models at inference time without requiring retraining by integrating it as a scoring function in a guided decoding scheme. Experimental results across multiple datasets and architectures confirm the effectiveness of our method in boosting generalization and performance. We open source our code at https://github.com/pkaliosis/fada.

Embodied Chain-of-Thought (ECoT) reasoning enhances vision-language-action (VLA) models by improving performance and interpretability through intermediate reasoning steps. However, its sequential autoregressive token generation introduces significant inference latency, limiting real-time deployment. We propose Fast ECoT, an inference-time acceleration method that exploits the structured and repetitive nature of ECoT to (1) cache and reuse high-level reasoning across timesteps and (2) parallelise the generation of modular reasoning steps. Additionally, we introduce an asynchronous scheduler that decouples reasoning from action decoding, further boosting responsiveness. Fast ECoT requires no model changes or additional training and integrates easily into existing VLA pipelines. Experiments in both simulation (LIBERO) and real-world robot tasks show up to a 7.5% reduction in latency with comparable or improved task success rate and reasoning faithfulness, bringing ECoT policies closer to practical real-time deployment.

Large Language Models (LLMs) have achieved impressive performance in text summarization and are increasingly deployed in real-world applications. However, these systems often inherit associative and framing biases from pre-training data, leading to inappropriate or unfair outputs in downstream tasks. In this work, we present AdvSumm (Adversarial Summarization), a domain-agnostic training framework designed to mitigate bias in text summarization through improved generalization. Inspired by adversarial robustness, AdvSumm introduces a novel Perturber component that applies gradient-guided perturbations at the embedding level of Sequence-to-Sequence models, enhancing the model's robustness to input variations. We empirically demonstrate that AdvSumm effectively reduces different types of bias in summarization-specifically, name-nationality bias and political framing bias-without compromising summarization quality. Compared to standard transformers and data augmentation techniques like back-translation, AdvSumm achieves stronger bias mitigation performance across benchmark datasets.