CLIP is a seminal multimodal model that maps images and text into a shared representation space by contrastive learning on billions of image–caption pairs. Inspired by the rapid progress of large language models (LLMs), we investigate how the superior linguistic understanding and broad world knowledge of LLMs can further strengthen CLIP—particularly in handling long, complex captions. We introduce an efficient fine-tuning framework that embeds an LLM into a pretrained CLIP while incurring almost the same training cost as regular CLIP fine-tuning. Our method first “embedding-izes” the LLM for the CLIP setting, then couples it to the pretrained CLIP vision encoder through a lightweight adaptor trained on only a few million image–caption pairs. With this strategy we achieve large performance gains—without large-scale retraining—over state-of-the-art CLIP variants such as EVA02 and SigLIP-2. The LLM-enhanced CLIP delivers consistent improvements across a wide spectrum of downstream tasks, including linear-probe classification, zero-shot image–text retrieval with both short and long captions (in English and other languages), zero-shot/supervised image segmentation, object detection, and used as tokenizer for multimodal large-model benchmarks.

Hallucination remains a major challenge for Large Vision-Language Models (LVLMs). Direct Preference Optimization (DPO) has gained increasing attention as a simple solution to hallucination issues. It directly learns from constructed preference pairs that reflect the severity of hallucinations in responses to the same prompt and image. Nonetheless, different data construction methods in existing works bring notable performance variations. We identify a crucial factor here: outcomes are largely contingent on whether the constructed data aligns on-policy w.r.t the initial (reference) policy of DPO. Theoretical analysis suggests that learning from off-policy data is impeded by the presence of KL-divergence between the updated policy and the reference policy. From the perspective of dataset distribution, we systematically summarize the inherent flaws in existing algorithms that employ DPO to address hallucination issues. To alleviate the problems, we propose On-Policy Alignment (OPA)-DPO framework, which uniquely leverages expert feedback to correct hallucinated responses and aligns both the original and expert-revised responses in an on-policy manner. Notably, with only 4.8k data, OPADPO achieves an additional reduction in the hallucination rate of LLaVA-1.5-7B: 13.26% on the AMBER benchmark and 5.39% on the Object-Hal benchmark, compared to the previous SOTA algorithm trained with 16k samples.

Visual understanding is inherently intention-driven - humans selectively focus on different regions of a scene based on their goals. Recent advances in large multimodal models (LMMs) enable flexible expression of such intentions through natural language, allowing queries to guide visual reasoning processes. Frameworks like Visual Chain-of-Thought have demonstrated the benefit of incorporating explicit reasoning steps, where the model predicts a focus region before answering a query. However, existing approaches rely heavily on supervised training with annotated intermediate bounding boxes, which severely limits scalability due to the combinatorial explosion of intention-region pairs. To overcome this limitation, we propose VisRL, the first framework that applies reinforcement learning (RL) to the problem of intention-driven visual perception. VisRL optimizes the entire visual reasoning process using only reward signals. By treating intermediate focus selection as an internal decision optimized through trial-and-error, our method eliminates the need for costly region annotations while aligning more closely with how humans learn to perceive the world. Extensive experiments across multiple benchmarks show that VisRL consistently outperforms strong baselines, demonstrating both its effectiveness and its strong generalization across different LMMs. Our code is available at https://github.com/zhangquanchen/VisRL.

Long-context Large Language Models (LLMs) have enabled numerous downstream applications but also introduced significant challenges related to computational and memory efficiency. To address these challenges, optimizations for long-context inference have been developed, centered around the KV cache. However, existing benchmarks often evaluate in single-request, neglecting the full lifecycle of the KV cache in real-world use. This oversight is particularly critical, as KV cache reuse has become widely adopted in LLMs inference frameworks, such as vLLM and SGLang, as well as by LLM providers, including OpenAI, Microsoft, Google, and Anthropic. To address this gap, we introduce SCBENCH (SharedContextBENCH), a comprehensive benchmark for evaluating long-context methods from a KV cache centric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, and 4) KV cache loading. Specifically, SCBench uses test examples with shared context, ranging 12 tasks with two shared context modes, covering four categories of long-context capabilities: string retrieval, semantic retrieval, global information, and multi-task. With SCBench, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs (Codestal-Mamba), Mamba-Attention hybrids (Jamba-1.5-Mini), and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on six Transformer-based long-context LLMs: Llama-3.1-8B/70B, Qwen2.5-72B/32B, Llama-3-8B-262K, and GLM-4-9B. Our findings show that sub-O(n) memory methods suffer in multi-turn scenarios, while sparse encoding with O(n) memory and sub-O(n^2) pre-filling computation perform robustly. Dynamic sparsity yields more expressive KV caches than static patterns, and layer-level sparsity in hybrid architectures reduces memory usage with strong performance. Additionally, we identify attention distribution shift issues in long-generation scenarios.

Parameter-efficient fine-tuning has demonstrated promising results across various visual adaptation tasks, such as classification and segmentation. Typically, prompt tuning techniques have harnessed knowledge from a single pre-trained model, whether from a general or a specialized medical domain. However, this approach typically overlooks the potential synergies that could arise from integrating diverse domain knowledge within the same tuning process. In this work, we propose a novel Mixture-of-Experts prompt tuning method called pMoE, which leverages the strengths of multiple expert domains through expert-specialized prompt tokens and the learnable dispatcher, effectively combining their expertise in a unified model framework. Our pMoE introduces expert-specific prompt tokens and utilizes a dynamic token dispatching mechanism at various prompt layers to optimize the contribution of each domain expert during the adaptation phase. By incorporating both domain knowledge from diverse experts, the proposed pMoE significantly enhances the model's versatility and applicability to a broad spectrum of tasks. We conduct extensive experiments across 47 adaptation tasks, including both classification and segmentation in general and medical domains. The results demonstrate that our pMoE not only achieves superior performance with a large margin of improvements but also offers an optimal trade-off between computational efficiency and adaptation effectiveness compared to existing methods.

To deliver coherent and personalized experiences in long-term conversations, existing approaches typically perform retrieval augmented response generation by constructing memory banks from conversation history at either the turn-level, session-level, or through summarization techniques. In this paper, we explore the impact of different memory granularities and present two key findings: (1) Both turn-level and session-level memory units are suboptimal, affecting not only the quality of final responses, but also the accuracy of the retrieval process. (2) The redundancy in natural language introduces noise, hindering precise retrieval. We demonstrate that *LLMLingua-2*, originally designed for prompt compression to accelerate LLM inference, can serve as an effective denoising method to enhance memory retrieval accuracy. Building on these insights, we propose **SeCom**, a method that constructs a memory bank with topical segments by introducing a conversation **Se**gmentation model, while performing memory retrieval based on **Com**pressed memory units. Experimental results show that **SeCom** outperforms turn-level, session-level, and several summarization-based methods on long-term conversation benchmarks such as *LOCOMO* and *Long-MT-Bench+*. Additionally, the proposed conversation segmentation method demonstrates superior performance on dialogue segmentation datasets such as *DialSeg711*, *TIAGE*, and *SuperDialSeg*.

The integration of long-context capabilities with visual understanding unlocks unprecedented potential for Vision Language Models (VLMs). However, the quadratic attention complexity during the pre-filling phase remains a significant obstacle to real-world deployment. To overcome this limitation, we introduce MMInference (Multimodality Million tokens Inference), a dynamic sparse attention method that accelerates the prefilling stage for long-context multi-modal inputs. First, our analysis reveals that the temporal and spatial locality of video input leads to a unique sparse pattern, the Grid pattern. Simultaneously, VLMs exhibit markedly different sparse distributions across different modalities. We introduce a permutation-based method to leverage the unique Grid pattern and handle modality boundary issues. By offline search the optimal sparse patterns for each head, MMInference constructs the sparse distribution dynamically based on the input. We also provide optimized GPU kernels for efficient sparse computations. Notably, MMInference integrates seamlessly into existing VLM pipelines without any model modifications or fine-tuning. Experiments on multi-modal benchmarks-including Video QA, Captioning, VisionNIAH, and Mixed-Modality NIAH-with state-of-the-art long-context VLMs (LongVila, LlavaVideo, VideoChat-Flash, Qwen2.5-VL) show that MMInference accelerates the pre-filling stage by up to 8.3x at 1M tokens while maintaining accuracy. Our code is available at https://ama.ms/MMInference.

Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, Lili Qiu. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

We introduce Nada, the first framework to autonomously design network algorithms by leveraging the generative capabilities of large language models (LLMs). Starting with an existing algorithm implementation, Nada enables LLMs to create a wide variety of alternative designs in the form of code blocks. It then efficiently identifies the top-performing designs through a series of filtering techniques, minimizing the need for full-scale evaluations and significantly reducing computational costs. Using adaptive bitrate (ABR) streaming as a case study, we demonstrate that Nada produces novel ABR algorithms---previously unknown to human developers---that consistently outperform the original algorithm in diverse network environments, including broadband, satellite, 4G, and 5G.

Shapelets and CNN are two typical approaches to model time series. Shapelets aim at finding a set of sub-sequences that extract feature-based interpretable shapes, but may suffer from accuracy and efficiency issues. CNN performs well by encoding sequences with a series of hidden representations, but lacks interpretability. In this paper, we demonstrate that shapelets are essentially equivalent to a specific type of CNN kernel with a squared norm and pooling. Based on this finding, we propose ShapeConv, an interpretable CNN layer with its kernel serving as shapelets to conduct time-series modeling tasks in both supervised and unsupervised settings. By incorporating shaping regularization, we enforce the similarity for maximum interpretability. We also find human knowledge can be easily injected to ShapeConv by adjusting its initialization and model performance is boosted with it. Experiments show that ShapeConv can achieve state-of-the-art performance on time-series benchmarks without sacrificing interpretability and controllability.

Despite the notable advancements in deep reinforcement learning (DRL) in recent years, a prevalent issue that is often overlooked is the impact of signal delay. Signal delay occurs when there is a lag between an agent's perception of the environment and its corresponding actions. In this paper, we first formalize delayed-observation Markov decision processes (DOMDP) by extending the standard MDP framework to incorporate signal delays. Next, we elucidate the challenges posed by the presence of signal delay in DRL, showing that trivial DRL algorithms and generic methods for partially observable tasks suffer greatly from delays. Lastly, we propose effective strategies to overcome these challenges. Our methods achieve remarkable performance in continuous robotic control tasks with large delays, yielding results comparable to those in non-delayed cases. Overall, our work contributes to a deeper understanding of DRL in the presence of signal delays and introduces novel approaches to address the associated challenges.

The computational challenges of Large Language Model (LLM) inference remain a significant barrier to their widespread deployment, especially as prompt lengths continue to increase. Due to the quadratic complexity of the attention computation, it takes 30 minutes for an 8B LLM to process a prompt of 1M tokens (i.e., the pre-filling stage) on a single A100 GPU. Existing methods for speeding up prefilling often fail to maintain acceptable accuracy or efficiency when applied to long-context LLMs. To address this gap, we introduce MInference (Milliontokens Inference), a sparse calculation method designed to accelerate pre-filling of long-sequence processing. Specifically, we identify three unique patterns in long-context attention matrices-the A-shape, Vertical-Slash, and Block-Sparse-that can be leveraged for efficient sparse computation on GPUs. We determine the optimal pattern for each attention head offline and dynamically build sparse indices based on the assigned pattern during inference. With the pattern and sparse indices, we perform efficient sparse attention calculations via our optimized GPU kernels to significantly reduce the latency in the pre-filling stage of longcontext LLMs. Our proposed technique can be directly applied to existing LLMs without any modifications to the pre-training setup or additional fine-tuning. By evaluating on a wide range of downstream tasks, including InfiniteBench, RULER, PG-19, and Needle In A Haystack, and models including LLaMA-3-1M, GLM-4-1M, Yi-200K, Phi-3-128K, and Qwen2-128K, we demonstrate that MInference effectively reduces inference latency by up to 10x for pre-filling on an A100, while maintaining accuracy. Our code is available at https://aka.ms/MInference.

Conventional reinforcement learning (RL) needs an environment to collect fresh data, which is impractical when online interactions are costly. Offline RL provides an alternative solution by directly learning from the previously collected dataset. However, it will yield unsatisfactory performance if the quality of the offline datasets is poor. In this paper, we consider an offline-to-online setting where the agent is first learned from the offline dataset and then trained online, and propose a framework called Adaptive Policy Learning for effectively taking advantage of offline and online data. Specifically, we explicitly consider the difference between the online and offline data and apply an adaptive update scheme accordingly, that is, a pessimistic update strategy for the offline dataset and an optimistic/greedy update scheme for the online dataset. Such a simple and effective method provides a way to mix the offline and online RL and achieve the best of both worlds. We further provide two detailed algorithms for implementing the framework through embedding value or policy-based RL algorithms into it. Finally, we conduct extensive experiments on popular continuous control tasks, and results show that our algorithm can learn the expert policy with high sample efficiency even when the quality of offline dataset is poor, e.g., random dataset.

The successes of artificial neural networks (ANNs) are largely attributed to mimicking the human brain structures. Recent advances in neuroscience revealed that neurons interact with each other through various kinds of connectivity patterns to process information, in which the common connectivity patterns are also called circuit motifs. However, many existing ANNs can only model one or two circuit motifs in their architectures, so that their performance may drastically vary among different types of machine learning tasks. In this paper, we propose a new type of neural network inspired by the architectures of neuronal circuits, namely Circuit Neural Network (CircuitNet). In CircuitNet, a group of densely connected neurons, namely circuit motif unit (CMU), form the basic unit of the network, which is capable of modeling universal circuit motifs by adjusting the weights within the CMUs. Compared with traditional feed-forward networks, CircuitNet has the ability to model more types of neuron connections such as feed-back and lateral motifs. Inspired by the locally dense and globally sparse structure of the human brain, several iterations of signal transmission among different CMUs are achieved by sparse connections through the input ports and output ports of different CMUs. Experiments have demonstrated that CircuitNet can outperform popular neural network architectures in function approximation, reinforcement learning, image classification, and time series forecasting tasks.

A timely detection of seizures for newborn infants with electroencephalogram (EEG) has been a common yet lifesaving practice in the Neonatal Intensive Care Unit (NICU). However, it requires great human efforts for real-time monitoring, which calls for automated solutions to neonatal seizure detection. Moreover, the current automated methods focusing on adult epilepsy monitoring often fail due to (i) dynamic seizure onset location in human brains; (ii) different montages on neonates and (iii) huge distribution shift among different subjects. In this paper, we propose a deep learning framework, namely STATENet, to address the exclusive challenges with exquisite designs at the temporal, spatial and model levels. The experiments over the real-world large-scale neonatal EEG dataset illustrate that our framework achieves significantly better seizure detection performance.

It is challenging for reinforcement learning (RL) algorithms to succeed in real-world applications. Take financial trading as an example, the market information is noisy yet imperfect and the macroeconomic regulation or other factors may shift between training and evaluation, thus it requires both generalization and high sample efficiency for resolving the task. However, directly applying typical RL algorithms can lead to poor performance in such scenarios. To derive a robust and applicable RL algorithm, in this work, we design a simple but effective method named Ensemble Proximal Policy Optimization (EPPO), which learns ensemble policies in an end-to-end manner. Notably, EPPO combines each policy and the policy ensemble organically and optimizes both simultaneously. In addition, EPPO adopts a diversity enhancement regularization over the policy space which helps to generalize to unseen states and promotes exploration. We theoretically prove that EPPO can increase exploration efficacy, and through comprehensive experimental evaluations on various tasks, we demonstrate that EPPO achieves higher efficiency and is robust for real-world applications compared with vanilla policy optimization algorithms and other ensemble methods. Code and supplemental materials are available at https://seqml.github.io/eppo.

We propose VRL3, a powerful data-driven framework with a simple design for solving challenging visual deep reinforcement learning (DRL) tasks. We analyze a number of major obstacles in taking a data-driven approach, and present a suite of design principles, novel findings, and critical insights about data-driven visual DRL. Our framework has three stages: in stage 1, we leverage non-RL datasets (e.g. ImageNet) to learn task-agnostic visual representations; in stage 2, we use offline RL data (e.g. a limited number of expert demonstrations) to convert the task-agnostic representations into more powerful task-specific representations; in stage 3, we fine-tune the agent with online RL. On a set of challenging hand manipulation tasks with sparse reward and realistic visual inputs, compared to the previous SOTA, VRL3 achieves an average of 780% better sample efficiency. And on the hardest task, VRL3 is 1220% more sample efficient (2440% when using a wider encoder) and solves the task with only 10% of the computation. These significant results clearly demonstrate the great potential of data-driven deep reinforcement learning.

Offline reinforcement learning (RL) aims at learning policies from previously collected static trajectory data without interacting with the real environment. Recent works provide a novel perspective by viewing offline RL as a generic sequence generation problem, adopting sequence models such as Transformer architecture to model distributions over trajectories and repurposing beam search as a planning algorithm. However, the training datasets utilized in general offline RL tasks are quite limited and often suffering from insufficient distribution coverage, which could me harmful to training sequence generation models yet has not drawn enough attention in the previous works. In this paper, we propose a novel algorithm named Bootstrapped Transformer, which incorporates the idea of bootstrapping and leverages the learned model to self-generate more offline data to further boost the training of sequence model. We conduct extensive experiments on two offline RL benchmarks and demonstrate that our model can largely remedy the limitations of the existing offline RL training and beat other strong baseline methods. We also analyze the generated pseudo data and the revealed characteristics may shed some light on offline RL training.

Recommender system is playing an increasingly important role in online news platforms nowadays. Recently, there is a growing demand for applying reinforcement learning (RL) algorithms to news recommendation aiming to maximize long-term and/or non-differentiable objectives. However, without an interactive simulated environment, it is extremely costly to develop powerful RL agents for news recommendation. In this paper, we build a user simulator, namely MINDSim, for news recommendation. Targeting at new user generation and corresponding behavior simulation, we first construct a hidden space for users using a generative adversarial network, so that new users can be generated by sampling from this hidden space. To capture complex and fast user interest drifts over time, we adopt an encoder-decoder architecture, which takes the clicked news during the simulation as input and outputs the new user interests for the next period of time. Finally, we build the MINDSim simulator using MIcrosoft News Dataset (MIND), and extensive experimental results on this large-scale real-world dataset demonstrate that MINDSim can simulate the behaviors of real users with high quality.

Auxiliary information, such as reviews, have been widely adopted to improve collaborative filtering (CF) algorithms, e.g., to boost the accuracy and provide explanations. However, most of the existing methods cannot distinguish between co-appearance and causality when learning from the reviews, so that they may rely on spurious correlations rather than causal relations in the recommendation — leading to poor generalization performance and unconvincing explanations. In this paper, we propose a Recommendation via Review Rationalization (R3) method including 1) a rationale generator to extract rationales from reviews to alleviate the effects of spurious correlations; 2) a rationale predictor to predict user ratings on items only from generated rationales; and 3) a correlation predictor upon both rationales and correlational features to ensure conditional independence between spurious correlations and rating predictions given causal rationales. Extensive experiments on real-world datasets show that the proposed method can achieve better generalization performance than state-of-the-art CF methods and provide causal-aware explanations even when the test data distribution changes.

Neural networks with ReLU activation functions (abbrev. ReLU Networks), have demonstrated their success in many applications. Recently, researchers noticed that ReLU networks are positively scale-invariant (PSI) while the weights are not. This mismatch may lead to undesirable behaviors in the optimization process. Hence, some new algorithms that conduct optimization directly in the path space (the path space is proven to be PSI) were developed, such as Stochastic Gradient Descent (SGD) in the path space. %nd it was shown that, SGD in the path space is superior to that in the weight space. However, it is still unknown that whether other deep learning techniques such as batch normalization (BN), could also have their counterparts in the path space. In this paper, we conduct a formal study on the design of BN in the path space. First, we propose path-reparameterization of ReLU networks, in which the weights in the networks are reparameterized by path-values. Then, the feedforward and backward propagation of the path-reparameterized networks can calculate the values of the hidden nodes and the gradients in the path space, respectively. Next, we design the a novel way to do batch normalization for the path-reparameterized ReLU networks, called Path-BN. Specifically, we notice that, path-reparameterized ReLU NNs have a portion of constant weights which play more critical roles to form the basis of the path space. We propose to exclude these constant weights when doing batch normalization and prove that, by doing so, the scale and the direction of the trained parameters can be more effectively decoupled during training. Finally, we conduct experiments on benchmark datasets. The results show that our proposed Path-BN can improve the performance of the optimization algorithms in the path space.

Learning expressive representations is always crucial for well-performed policies in deep reinforcement learning (DRL). Different from supervised learning, in DRL, accurate targets are not always available, and some inputs with different actions only have tiny differences, which stimulates the demand for learning expressive representations. In this paper, firstly, we empirically compare the representations of DRL models with different performances. We observe that the representations of a better state extractor (SE) are more scattered than a worse one when they are visualized. Thus, we investigate the singular values of representation matrix, and find that, better SEs always correspond to smaller differences among these singular values. Next, based on such observations, we define an indicator of the representations for DRL model, which is the Number of Significant Singular Values (NSSV) of a representation matrix. Then, we propose I4R algorithm, to improve DRL algorithms by adding the corresponding regularization term to enhance the NSSV. Finally, we apply I4R to both policy gradient and value based algorithms on Atari games, and the results show the superiority of our proposed method.

Chatbots have drawn significant attention of late in both industry and academia. For most task completion bots in the industry, human intervention is the only means of avoiding mistakes in complex real-world cases. However, to the best of our knowledge, there is no existing research work modeling the collaboration between task completion bots and human workers. In this paper, we introduce CoChat, a dialog management framework to enable effective collaboration between bots and human workers. In CoChat, human workers can introduce new actions at any time to handle previously unseen cases. We propose a memory-enhanced hierarchical RNN (MemHRNN) to handle the one-shot learning challenges caused by instantly introducing new actions in CoChat. Extensive experiments on real-world datasets well demonstrate that CoChat can relieve most of the human workers’ workload, and get better user satisfaction rates comparing to other state-of-the-art frameworks.