High-quality chain-of-thought has demonstrated strong potential for unlocking the reasoning capabilities of large language models. However, current paradigms typically treat the reasoning process as an indivisible sequence, lacking an intrinsic mechanism to quantify step-wise information gain. This granularity gap manifests in two limitations: inference inefficiency from redundant exploration without explicit guidance, and optimization difficulty due to sparse outcome supervision or costly external verifiers. In this work, we propose CoT-Flow, a framework that reconceptualizes discrete reasoning steps as a continuous probabilistic flow, quantifying the contribution of each step toward the ground-truth answer. Built on this formulation, CoT-Flow enables two complementary methodologies: flow-guided decoding, which employs a greedy flow-based decoding strategy to extract information-efficient reasoning paths, and flow-based reinforcement learning, which constructs a verifier-free dense reward function. Experiments on challenging benchmarks demonstrate that CoT-Flow achieves a superior balance between inference efficiency and reasoning performance.

Knowledge distillation has emerged as a pivotal paradigm for transferring the superior reasoning capabilities of Large Reasoning Models (LRMs) to efficient student models. However, the raw Chain-of-Thought (CoT) trajectories are often verbose and redundant, which dilutes the underlying logic and hinders effective knowledge distillation for student models. Although recent work has focused on pruning CoT to streamline these reasoning paths, existing local heuristic methods often fail to capture global causal logic due to rigid rules and limited search spaces, while global heuristic approaches incur substantial computational costs. To address these issues, we propose Pru-CoT (Pruning Chain-of-Thought), a framework that aims to extract the essential logical structure from reasoning chains. Pru-CoT implements a step-level importance assessment via global optimization on a frozen student large language model (LLM), quantifying the gradient-based causal contribution of each component. Guided by these important signals, the framework performs fidelity-constrained pruning, utilizing an LLM-driven process to synthesize concise, logically coherent narratives. Extensive experiments on mathematical reasoning benchmarks demonstrate that models trained with Pru-CoT not only achieve superior accuracy but also generate significantly more compact reasoning paths compared to those trained on raw verbose data.

With the continuously expanding parameters, efficiently adapting large language models to downstream tasks is crucial in resource-limited conditions. Many parameter-efficient fine-tuning methods have emerged to address this challenge. However, they lack flexibility, like LoRA requires manually selecting trainable parameters and rank size, (IA)³ can only scale the activations along columns, yielding inferior results due to less precise fine-tuning. To address these issues, we propose a novel method named AdaDHP with fewer parameters and finer granularity, which can adaptively select important parameters for each task. Specifically, we introduce two trainable vectors for each parameter and fine-tune the parameters through Hadamard product along both rows and columns. This significantly reduces the number of trainable parameters, with our parameter count capped at the lower limit of LoRA. Moreover, we design an adaptive parameter selection strategy to select important parameters for downstream tasks dynamically. This allows our method to flexibly remove unimportant parameters for downstream tasks. Finally, we demonstrate the superiority of our method on the T5-base model across 17 NLU tasks and on complex mathematical tasks with the Llama series models.

This paper delves into a novel backdoor attack scenario, aiming to uncover potential security risks associated with Multimodal Large Language Models (MLLMs) during multi-round open-ended conversations with users. In the practical use of MLLMs, users have full control over the interaction process with the model, such as using their own collected photos and posing arbitrary open-ended questions. Traditional backdoor attacks that rely on adding external triggers are less applicable. To this end, we introduce a new shadow-activated backdoor attacking paradigm in this paper, wherein attacks implicitly inject malicious content into the responses of MLLMs when the responses explicitly relate to the shadowed object, i.e., without any triggers. To facilitate the shadow-activated backdoor attack, we present a novel framework named BadMLLM to achieve the desired behaviors by constructing a poisoned dataset using GPT-4 Vision and implementing an attention-regularized tuning strategy to address the semantic discontinuity between the original response and the inserted promotion. Extensive experimental results conducted on five MLLMs, three objects, and two types of promotion slogans have demonstrated impressive performance in achieving both efficacy and utility goals, thereby highlighting the significant potential risks concealed within MLLMs.

Generalized zero-shot intent detection (GZID) aims to recognize the labels of utterances from both seen and unseen intents by utilizing the knowledge learned from seen intents. Enhancing the generalization ability from seen intents to unseen intents is a key challenge in the GZID setting. Existing methods attempt to tackle this challenge by distinguishing unseen intents from seen intents or focusing on enhancing the model discriminability. However, the challenge is not solved substantially as they ignore to promote the representation learning ability of the model itself and neglect to strengthen the model adaptability to new tasks, resulting in overfitting on the seen intents. In this paper, we propose a pairwise prompt-based tuning model with parameter efficient fast adaptation which involves two training steps. In the first step, we leverage hybrid contrastive learning in discriminant space and masked language modeling to make predictions at both sentence and token levels, which can enhance the model discriminability and representation learning ability respectively. In the second step, we design a pipeline for generating and filtering unseen data by only providing unseen intent labels, and utilize parameter-efficient fine-tuning to quickly adapt to unseen intents. Experiments on four intent detection datasets demonstrate that our two-step training method has better comprehension and generalization capabilities.

Multi-label document classification (MLDC) aims to allocate more than one label to each document and attracts increasing attention in many practical applications. However, previous studies have failed to pay sufficient attention to the lack of semantic information on labels and the long-tail problem prevalent in the datasets. Additionally, most existing methods focus on optimizing document features, overlooking the potential of high-quality label features to enhance classification performance. In this paper, we propose a simple and effective paradigm for MLDC. Regarding the problem of insufficient label information and imbalance in the sample size of categories, we utilize large language models (LLMs) to semantically expand the label content and generate pseudo-samples for the tail categories. To optimize the features of both documents and labels, we design the contrastive learning boosted feature optimization module facilitated by the similarity matrices. Finally, we construct a label-guided feature selection module to incorporate the optimized label features into the input features to provide richer semantic information for the classifier. Extensive experiments have demonstrated that our proposed method significantly outperforms state-of-the-art baselines.

Parameter Efficient Fine-Tuning (PEFT) offers an efficient solution for fine-tuning large pretrained language models for downstream tasks. However, most PEFT strategies are manually designed, often resulting in suboptimal performance. Recent automatic PEFT approaches aim to address this but face challenges such as search space entanglement, inefficiency, and lack of integration between parameter budgets and search processes. To overcome these issues, we introduce a novel Budget-guided Iterative search strategy for automatic PEFT (BIPEFT), significantly enhancing search efficiency. BIPEFT employs a new iterative search strategy to disentangle the binary module and rank dimension search spaces. Additionally, we design early selection strategies based on parameter budgets, accelerating the learning process by gradually removing unimportant modules and fixing rank dimensions. Extensive experiments on public benchmarks demonstrate the superior performance of BIPEFT in achieving efficient and effective PEFT for downstream tasks with a low parameter budget.

Few-shot intent detection is a challenging task, particularly in scenarios involving multiple labels and diverse domains. This paper presents a novel prototype learning approach that combines the label synset augmentation and the coarse-to-fine prototype distillation for multi-label few-shot intent detection. To tackle the data scarcity issue and the lack of information for unseen domains, we propose to enhance the representations of utterances with label synset augmentation and refine the prototypes by distilling the coarse domain knowledge from a universal teacher model. To solve the multilingual intent detection in real-world dialogue systems, we fine-tune a cross-lingual teacher model to make our method fast adapt to different languages and re-annotate two non-English task-oriented dialogue datasets CrossWOZ and JMultiWOZ in multi-label form. Experimental results on one English and two non-English datasets demonstrate that our approach significantly outperforms existing methods in terms of accuracy and generalization across different domains.