CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models

Vision-Language Models (VLMs) excel across diverse tasks but suffer from high inference costs in time and memory. Token sparsity mitigates inefficiencies in token usage, while neuron sparsity reduces high-dimensional computations, both offering promising solutions to enhance efficiency. Recently, these two sparsity paradigms have evolved largely in parallel, fostering the prevailing assumption that they function independently. However, a fundamental yet underexplored question remains: Do they truly operate in isolation, or is there a deeper underlying interplay that has yet to be uncovered? In this paper, we conduct the first comprehensive investigation into this question. By introducing and analyzing the matching mechanism between Core Neurons and Core Tokens, we found that key neurons and tokens for inference mutually influence and reinforce each other. Building on this insight, we propose CoreMatching, a co-adaptive sparse inference framework, which leverages the synergy between token and neuron sparsity to enhance inference efficiency. Through theoretical analysis and efficiency evaluations, we demonstrate that the proposed method surpasses state-of-the-art baselines on ten image understanding tasks and three hardware devices. Notably, on the NVIDIA Titan Xp, it achieved 5x FLOPs reduction and a 10x overall speedup. Code is released at https://github.com/wangqinsi1/2025-ICML-CoreMatching/tree/main.

FL-PLAS: Federated Learning with Partial Layer Aggregation for Backdoor Defense Against High-Ratio Malicious Clients

Federated learning (FL) is gaining increasing attention as an emerging collaborative machine learning approach, particularly in the context of large-scale computing and data systems. However, the fundamental algorithm of FL, Federated Averaging (FedAvg), is susceptible to backdoor attacks. Although researchers have proposed numerous defense algorithms, two significant challenges remain. The attack is becoming more stealthy and harder to detect, and current defense methods are unable to handle 50\% or more malicious users or assume an auxiliary server dataset. To address these challenges, we propose a novel defense algorithm, FL-PLAS, \textbf{F}ederated \textbf{L}earning based on \textbf{P}artial\textbf{ L}ayer \textbf{A}ggregation \textbf{S}trategy. In particular, we divide the local model into a feature extractor and a classifier. In each iteration, the clients only upload the parameters of a feature extractor after local training. The server then aggregates these local parameters and returns the results to the clients. Each client retains its own classifier layer, ensuring that the backdoor labels do not impact other clients. We assess the effectiveness of FL-PLAS against state-of-the-art (SOTA) backdoor attacks on three image datasets and compare our approach to six defense strategies. The results of the experiment demonstrate that our methods can effectively protect local models from backdoor attacks. Without requiring any auxiliary dataset for the server, our method achieves a high main-task accuracy with a lower backdoor accuracy even under the condition of 90\% malicious users with the attacks of trigger, semantic and edge-case.

Keyframe-oriented Vision Token Pruning: Enhancing Efficiency of Large Vision Language Models on Long-Form Video Processing

Vision language models (VLMs) demonstrate strong capabilities in jointly processing visual and textual data. However, they often incur substantial computational overhead due to redundant visual information, particularly in long-form video scenarios. Existing approaches predominantly focus on either vision token pruning, which may overlook spatio-temporal dependencies, or keyframe selection, which identifies informative frames but discards others, thus disrupting contextual continuity. In this work, we propose KVTP (Keyframe-oriented Vision Token Pruning), a novel framework that overcomes the drawbacks of token pruning and keyframe selection. By adaptively assigning pruning rates based on frame relevance to the query, KVTP effectively retains essential contextual information while significantly reducing redundant computation. To thoroughly evaluate the long-form video understanding capacities of VLMs, we curated and reorganized subsets from VideoMME, EgoSchema, and NextQA into a unified benchmark named SparseKV-QA that highlights real-world scenarios with sparse but crucial events. Our experiments with VLMs of various scales show that KVTP can reduce token usage by 80% without compromising spatiotemporal and contextual consistency, significantly cutting computation while maintaining the performance. These results demonstrate our approach's effectiveness in efficient long-video processing, facilitating more scalable VLM deployment.

Proactive Privacy Amnesia for Large Language Models: Safeguarding PII with Negligible Impact on Model Utility

With the rise of large language models (LLMs), increasing research has recognized their risk of leaking personally identifiable information (PII) under malicious attacks. Although efforts have been made to protect PII in LLMs, existing methods struggle to balance privacy protection with maintaining model utility. In this paper, inspired by studies of amnesia in cognitive science, we propose a novel approach, Proactive Privacy Amnesia (PPA), to safeguard PII in LLMs while preserving their utility. This mechanism works by actively identifying and forgetting key memories most closely associated with PII in sequences, followed by a memory implanting using suitable substitute memories to maintain the LLM's functionality. We conduct evaluations across multiple models to protect common PII, such as phone numbers and physical addresses, against prevalent PII-targeted attacks, demonstrating the superiority of our method compared with other existing defensive techniques. The results show that our PPA method completely eliminates the risk of phone number exposure by 100% and significantly reduces the risk of physical address exposure by 9.8% - 87.6%, all while maintaining comparable model utility performance.

H-CoT: Hijacking the Chain-of-Thought Safety Reasoning Mechanism to Jailbreak Large Reasoning Models, Including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking

Large Reasoning Models (LRMs) have recently extended their powerful reasoning capabilities to safety checks-using chain-of-thought reasoning to decide whether a request should be answered. While this new approach offers a promising route for balancing model utility and safety, its robustness remains underexplored. To address this gap, we introduce Malicious-Educator, a benchmark that disguises extremely dangerous or malicious requests beneath seemingly legitimate educational prompts. Our experiments reveal severe security flaws in popular commercial-grade LRMs, including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking. For instance, although OpenAI's o1 model initially maintains a high refusal rate of about 98%, subsequent model updates significantly compromise its safety; and attackers can easily extract criminal strategies from DeepSeek-R1 and Gemini 2.0 Flash Thinking without any additional tricks. To further highlight these vulnerabilities, we propose Hijacking Chain-of-Thought (H-CoT), a universal and transferable attack method that leverages the model's own displayed intermediate reasoning to jailbreak its safety reasoning mechanism. Under H-CoT, refusal rates sharply decline-dropping from 98% to below 2%-and, in some instances, even transform initially cautious tones into ones that are willing to provide harmful content. We hope these findings underscore the urgent need for more robust safety mechanisms to preserve the benefits of advanced reasoning capabilities without compromising ethical standards.

SpeechPrune: Context-aware Token Pruning for Speech Information Retrieval

We introduce Speech Information Retrieval (SIR), a new long-context task for Speech Large Language Models (Speech LLMs), and present SPIRAL, a 1,012-sample benchmark testing models' ability to extract critical details from approximately 90-second spoken inputs. While current Speech LLMs excel at short-form tasks, they struggle with the computational and representational demands of longer audio sequences. To address this limitation, we propose SpeechPrune, a training-free token pruning strategy that uses speech-text similarity and approximated attention scores to efficiently discard irrelevant tokens. In SPIRAL, SpeechPrune achieves accuracy improvements of 29% and up to 47% over the original model and the random pruning model at a pruning rate of 20%, respectively. SpeechPrune can maintain network performance even at a pruning level of 80%. This approach highlights the potential of token-level pruning for efficient and scalable long-form speech understanding.

Sufficient Context: A New Lens on Retrieval Augmented Generation Systems

Augmenting LLMs with context leads to improved performance across many applications. Despite much research on Retrieval Augmented Generation (RAG) systems, an open question is whether errors arise because LLMs fail to utilize the context from retrieval or the context itself is insufficient to answer the query. To shed light on this, we develop a new notion of sufficient context, along with a way to classify instances that have enough information to answer the query. We then use sufficient context to analyze several models and datasets. By stratifying errors based on context sufficiency, we find that proprietary LLMs (Gemini, GPT, Claude) excel at answering queries when the context is sufficient, but often output incorrect answers instead of abstaining when the context is not. On the other hand, open-source LLMs (Llama, Mistral, Gemma) hallucinate or abstain often, even with sufficient context. We further categorize cases when the context is useful, and improves accuracy, even though it does not fully answer the query and the model errs without the context. Building on our findings, we explore ways to reduce hallucinations in RAG systems, including a new selective generation method that leverages sufficient context information for guided abstention. Our method improves the fraction of correct answers among times where the model responds by 2-10% for Gemini, GPT, and Gemma.

SLED: Self Logits Evolution Decoding for Improving Factuality in Large Language Models

Large language models (LLMs) have demonstrated remarkable capabilities, but their outputs can sometimes be unreliable or factually incorrect. To address this, we introduce Self Logits Evolution Decoding (SLED), a novel decoding framework that enhances the truthfulness of LLMs without relying on external knowledge bases or requiring further fine-tuning. From an optimization perspective, our SLED framework leverages the latent knowledge embedded within the LLM by contrasting the output logits from the final layer with those from early layers. It then utilizes an approximate gradient approach to enable latent knowledge to guide the self-refinement of outputs, thereby effectively improving factual accuracy. Extensive experiments have been conducted on established benchmarks across a diverse range of model families (LLaMA 2, LLaMA 3, Gemma) and scales (from 2B to 70B), including more advanced architectural configurations such as the mixture of experts (MoE). Our evaluation spans a wide variety of tasks, including multi-choice, open-generation, and adaptations to chain-of-thought reasoning tasks. The results demonstrate that SLED consistently improves factual accuracy by up to 20\% compared to existing decoding methods while maintaining natural language fluency and negligible latency overhead. Furthermore, it can be flexibly combined with other decoding methods to further enhance their performance.

CoreInfer: Accelerating Large Language Model Inference with Semantics-Inspired Adaptive Sparse Activation

Large language models (LLMs) with billions of parameters have sparked a new wave of exciting AI applications. However, their high computational costs and memory demands during inference pose significant challenges. Adaptive sparse activation inference, which activates only a small number of neurons for each token, offers a novel way to accelerate model inference without degrading performance, showing great potential for resource-constrained hardware devices. Nevertheless, existing methods predict activated neurons based on individual tokens with additional MLP, which involve frequent changes in activation maps and resource calls, limiting the acceleration benefits of sparse activation. In this paper, we introduce CoreInfer, an MLP-free adaptive sparse activation inference method based on sentence-level prediction. Specifically, we propose the concept of sentence-wise core neurons, which refers to the subset of neurons most critical for a given sentence, and empirically demonstrate its effectiveness. To determine the core neurons, we explore the correlation between core neurons and the sentence's semantics. Remarkably, we discovered that core neurons exhibit both stability and similarity in relation to the sentence's semantics -- an insight overlooked by previous studies. Building on this finding, we further design two semantic-based methods for predicting core neurons to fit different input scenarios. In CoreInfer, the core neurons are determined during the pre-filling stage and fixed during the encoding stage, enabling zero-cost sparse inference. We evaluated the model generalization and task generalization of CoreInfer across various models and tasks. Notably, on an NVIDIA TITAN XP GPU, CoreInfer achieved a 10.33 times and 2.72 times speedup compared to the Huggingface implementation and PowerInfer, respectively.

MLLM-FL: Multimodal Large Language Model Assisted Federated Learning on Heterogeneous and Long-tailed Data

Previous studies on federated learning (FL) often encounter performance degradation due to data heterogeneity among different clients. In light of the recent advances in multimodal large language models (MLLMs), such as GPT-4v and LLaVA, which demonstrate their exceptional proficiency in multimodal tasks, such as image captioning and multimodal question answering. We introduce a novel federated learning framework, named Multimodal Large Language Model Assisted Federated Learning (MLLM-FL), which which employs powerful MLLMs at the server end to address the heterogeneous and long-tailed challenges. Owing to the advanced cross-modality representation capabilities and the extensive open-vocabulary prior knowledge of MLLMs, our framework is adept at harnessing the extensive, yet previously underexploited, open-source data accessible from websites and powerful server-side computational resources. Hence, the MLLM-FL not only enhances the performance but also avoids increasing the risk of privacy leakage and the computational burden on local devices, distinguishing it from prior methodologies. Our framework has three key stages. Initially, prior to local training on local datasets of clients, we conduct global visual-text pretraining of the model. This pretraining is facilitated by utilizing the extensive open-source data available online, with the assistance of multimodal large language models. Subsequently, the pretrained model is distributed among various clients for local training. Finally, once the locally trained models are transmitted back to the server, a global alignment is carried out under the supervision of MLLMs to further enhance the performance. Experimental evaluations on established benchmarks, show that our framework delivers promising performance in the typical scenarios with data heterogeneity and long-tail distribution across different clients in FL.