Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang, Weijia Shi, Yike Wang, Shannon Zejiang Shen, Xiaochuang Han, Hunter Lang, Chen-Yu Lee, Tomas Pfister, Yejin Choi, Yulia Tsvetkov. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large Language Models (LLMs) have made significant progress in open-ended dialogue, yet their inability to retain and retrieve relevant information from long-term interactions limits their effectiveness in applications requiring sustained personalization. External memory mechanisms have been proposed to address this limitation, enabling LLMs to maintain conversational continuity. However, existing approaches struggle with two key challenges. First, rigid memory granularity fails to capture the natural semantic structure of conversations, leading to fragmented and incomplete representations. Second, fixed retrieval mechanisms cannot adapt to diverse dialogue contexts and user interaction patterns. In this work, we propose Reflective Memory Management (RMM), a novel mechanism for long-term dialogue agents, integrating forward- and backward-looking reflections: (1) Prospective Reflection, which dynamically summarizes interactions across granularities—utterances, turns, and sessions—into a personalized memory bank for effective future retrieval, and (2) Retrospective Reflection, which iteratively refines the retrieval in an online reinforcement learning (RL) manner based on LLMs’ cited evidence. Experiments show that RMM demonstrates consistent improvement across various metrics and benchmarks. For example, RMM shows more than 10% accuracy improvement over the baseline without memory management on the LongMemEval dataset.

Fan Yin, Zifeng Wang, I-Hung Hsu, Jun Yan, Ke Jiang, Yanfei Chen, Jindong Gu, Long Le, Kai-Wei Chang, Chen-Yu Lee, Hamid Palangi, Tomas Pfister. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Mihir Parmar, Xin Liu, Palash Goyal, Yanfei Chen, Long Le, Swaroop Mishra, Hossein Mobahi, Jindong Gu, Zifeng Wang, Hootan Nakhost, Chitta Baral, Chen-Yu Lee, Tomas Pfister, Hamid Palangi. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Retrieval augmented generation (RAG) combines the generative abilities of large language models (LLMs) with external knowledge sources to provide more accurate and up-to-date responses. Recent RAG advancements focus on improving retrieval outcomes through iterative LLM refinement or self-critique capabilities acquired through additional instruction tuning of LLMs. In this work, we introduce Speculative RAG - a framework that leverages a larger generalist LM to efficiently verify multiple RAG drafts produced in parallel by a smaller, distilled specialist LM. Each draft is generated from a distinct subset of retrieved documents, offering diverse perspectives on the evidence while reducing input token counts per draft. This approach enhances comprehension of each subset and mitigates potential position bias over long context. Our method accelerates RAG by delegating drafting to the smaller specialist LM, with the larger generalist LM performing a single verification pass over the drafts. Extensive experiments demonstrate that Speculative RAG achieves state-of-the-art performance with reduced latency on TriviaQA, MuSiQue, PopQA, PubHealth, and ARC-Challenge benchmarks. It notably enhances accuracy by up to 12.97% while reducing latency by 50.83% compared to conventional RAG systems on PubHealth.

Humans can naturally learn new and varying tasks in a sequential manner. Continual learning is a class of learning algorithms that updates its learned model as it sees new data (on potentially new tasks) in a sequence. A key challenge in continual learning is that as the model is updated to learn new tasks, it becomes susceptible to \textit{catastrophic forgetting}, where knowledge of previously learned tasks is lost. A popular approach to mitigate forgetting during continual learning is to maintain a small buffer of previously-seen samples, and to replay them during training. However, this approach is limited by the small buffer size and, while forgetting is reduced, it is still present. In this paper, we propose a novel loss function STAR that exploits the worst-case parameter perturbation that reduces the KL-divergence of model predictions with that of its local parameter neighborhood to promote stability and alleviate forgetting. STAR can be combined with almost any existing rehearsal-based methods as a plug-and-play component. We empirically show that STAR consistently improves performance of existing methods by up to $\sim15\\%$ across varying baselines, and achieves superior or competitive accuracy to that of state-of-the-art methods aimed at improving rehearsal-based continual learning. Our implementation is available at https://github.com/Gnomy17/STAR_CL.

Recent advances in knowledge distillation (KD) have enabled smaller student models to approach the performance of larger teacher models. However, popular methods such as supervised KD and on-policy KD, are adversely impacted by the knowledge gaps between teacher-student in practical scenarios. Supervised KD suffers from a distribution mismatch between training with a static dataset and inference over final student-generated outputs. Conversely, on-policy KD, which uses student-generated samples for training, can suffer from low-quality training examples with which teacher models are not familiar, resulting in inaccurate teacher feedback. To address these limitations, we introduce Speculative Knowledge Distillation (SKD), a novel approach that leverages cooperation between student and teacher models to generate high-quality training data on-the-fly while aligning with the student's inference-time distribution. In SKD, the student proposes tokens, and the teacher replaces poorly ranked ones based on its own distribution, transferring high-quality knowledge adaptively. We evaluate SKD on various text generation tasks, including translation, summarization, math, and instruction following, and show that SKD consistently outperforms existing KD methods across different domains, data sizes, and model initialization strategies.

We propose Model Swarms, a collaborative search algorithm to adapt LLMs via swarm intelligence, the collective behavior guiding individual systems. Specifically, Model Swarms starts with a pool of LLM experts and a utility function. Guided by the best-found checkpoints across models, diverse LLM experts collaboratively move in the weight space and optimize a utility function representing model adaptation objectives. Compared to existing model composition approaches, Model Swarms offers tuning-free model adaptation, works in low-data regimes with as few as 200 examples, and does not require assumptions about specific experts in the swarm or how they should be composed. Extensive experiments demonstrate that Model Swarms could flexibly adapt LLM experts to a single task, multi-task domains, reward models, as well as diverse human interests, improving over 12 model composition baselines by up to 21.0% across tasks and contexts. Further analysis reveals that LLM experts discover previously unseen capabilities in initial checkpoints and that Model Swarms enable the weak-to-strong transition of experts through the collaborative search process.

Justin Chen, Zifeng Wang, Hamid Palangi, Rujun Han, Sayna Ebrahimi, Long Le, Vincent Perot, Swaroop Mishra, Mohit Bansal, Chen-Yu Lee, Tomas Pfister. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

We propose Heterogeneous Swarms, an algorithm to design multi-LLM systems by jointly optimizing model roles and weights. We represent multi-LLM systems as directed acyclic graphs (DAGs) of LLMs with topological message passing for collaborative generation. Given a pool of LLM experts and a utility function, Heterogeneous Swarms employs two iterative steps: role-step and weight-step. For role-step, we interpret model roles as learning a DAG that specifies the flow of inputs and outputs between LLMs. Starting from a swarm of random continuous adjacency matrices, we decode them into discrete DAGs, call the LLMs in topological order, evaluate on the utility function (e.g. accuracy on a task), and optimize the adjacency matrices with particle swarm optimization based on the utility score. For weight-step, we assess the contribution of individual LLMs in the multi-LLM systems and optimize model weights with swarm intelligence. We propose JFK-score to quantify the individual contribution of each LLM in the best-found DAG of the role-step, then optimize model weights with particle swarm optimization based on the JFK-score. Experiments demonstrate that Heterogeneous Swarms outperforms 17 role- and/or weight-based baselines by 18.5% on average across 12 tasks. Further analysis reveals that Heterogeneous Swarms discovers multi-LLM systems with heterogeneous model roles and substantial collaborative gains, and benefits from the diversity of language models.

Table-based reasoning with large language models (LLMs) is a promising direction to tackle many table understanding tasks, such as table-based question answering and fact verification. Compared with generic reasoning, table-based reasoning requires the extraction of underlying semantics from both free-form questions and semi-structured tabular data. Chain-of-Thought and its similar approaches incorporate the reasoning chain in the form of textual context, but it is still an open question how to effectively leverage tabular data in the reasoning chain. We propose the Chain-of-Table framework, where tabular data is explicitly used in the reasoning chain as a proxy for intermediate thoughts. Specifically, we guide LLMs using in-context learning to iteratively generate operations and update the table to represent a tabular reasoning chain. LLMs can therefore dynamically plan the next operation based on the results of the previous ones. This continuous evolution of the table forms a chain, showing the reasoning process for a given tabular problem. The chain carries structured information of the intermediate results, enabling more accurate and reliable predictions. Chain-of-Table achieves new state-of-the-art performance on WikiTQ, FeTaQA, and TabFact benchmarks across multiple LLM choices.

Recent advancements in language models (LMs) have notably enhanced their ability to reason with tabular data, primarily through program-aided mechanisms that manipulate and analyze tables. However, these methods often require the entire table as input, leading to scalability challenges due to the positional bias or context length constraints. In response to these challenges, we introduce TableRAG, a Retrieval-Augmented Generation (RAG) framework specifically designed for LM-based table understanding. TableRAG leverages query expansion combined with schema and cell retrieval to pinpoint crucial information before providing it to the LMs. This enables more efficient data encoding and precise retrieval, significantly reducing prompt lengths and mitigating information loss. We have developed two new million-token benchmarks from the Arcade and BIRD-SQL datasets to thoroughly evaluate TableRAG's effectiveness at scale. Our results demonstrate that TableRAG's retrieval design achieves the highest retrieval quality, leading to the new state-of-the-art performance on large-scale table understanding.

Rehearsal-based approaches are a mainstay of continual learning (CL). They mitigate the catastrophic forgetting problem by maintaining a small fixed-size buffer with a subset of data from past tasks. While most rehearsal-based approaches exploit the knowledge from buffered past data, little attention is paid to inter-task relationships and to critical task-specific and task-invariant knowledge. By appropriately leveraging inter-task relationships, we propose a novel CL method, named DualHSIC, to boost the performance of existing rehearsal-based methods in a simple yet effective way. DualHSIC consists of two complementary components that stem from the so-called Hilbert Schmidt independence criterion (HSIC): HSIC-Bottleneck for Rehearsal (HBR) lessens the inter-task interference and HSIC Alignment (HA) promotes task-invariant knowledge sharing. Extensive experiments show that DualHSIC can be seamlessly plugged into existing rehearsal-based methods for consistent performance improvements, outperforming recent state-of-the-art regularization-enhanced rehearsal methods.

The mainstream paradigm behind continual learning has been to adapt the model parameters to non-stationary data distributions, where catastrophic forgetting is the central challenge. Typical methods rely on a rehearsal buffer or known task identity at test time to retrieve learned knowl-edge and address forgetting, while this work presents a new paradigm for continual learning that aims to train a more succinct memory system without accessing task identity at test time. Our method learns to dynamically prompt (L2P) a pre-trained model to learn tasks sequen-tially under different task transitions. In our proposed framework, prompts are small learnable parameters, which are maintained in a memory space. The objective is to optimize prompts to instruct the model prediction and ex-plicitly manage task-invariant and task-specific knowledge while maintaining model plasticity. We conduct comprehen-sive experiments under popular image classification bench-marks with different challenging continual learning set-tings, where L2P consistently outperforms prior state-of-the-art methods. Surprisingly, L2P achieves competitive results against rehearsal-based methods even without a re-hearsal buffer and is directly applicable to challenging task-agnostic continual learning. Source code is available at https://github.com/google-research/12p.

Adversarial pruning compresses models while preserving robustness. Current methods require access to adversarial examples during pruning. This significantly hampers training efficiency. Moreover, as new adversarial attacks and training methods develop at a rapid rate, adversarial pruning methods need to be modified accordingly to keep up. In this work, we propose a novel framework to prune a previously trained robust neural network while maintaining adversarial robustness, without further generating adversarial examples. We leverage concurrent self-distillation and pruning to preserve knowledge in the original model as well as regularizing the pruned model via the Hilbert-Schmidt Information Bottleneck. We comprehensively evaluate our proposed framework and show its superior performance in terms of both adversarial robustness and efficiency when pruning architectures trained on the MNIST, CIFAR-10, and CIFAR-100 datasets against five state-of-the-art attacks..

Existing work in continual learning (CL) focuses on mitigating catastrophic forgetting, i.e., model performance deterioration on past tasks when learning a new task. However, the training efficiency of a CL system is under-investigated, which limits the real-world application of CL systems under resource-limited scenarios. In this work, we propose a novel framework called Sparse Continual Learning (SparCL), which is the first study that leverages sparsity to enable cost-effective continual learning on edge devices. SparCL achieves both training acceleration and accuracy preservation through the synergy of three aspects: weight sparsity, data efficiency, and gradient sparsity. Specifically, we propose task-aware dynamic masking (TDM) to learn a sparse network throughout the entire CL process, dynamic data removal (DDR) to remove less informative training data, and dynamic gradient masking (DGM) to sparsify the gradient updates. Each of them not only improves efficiency, but also further mitigates catastrophic forgetting. SparCL consistently improves the training efficiency of existing state-of-the-art (SOTA) CL methods by at most 23X less training FLOPs, and, surprisingly, further improves the SOTA accuracy by at most 1.7%. SparCL also outperforms competitive baselines obtained from adapting SOTA sparse training methods to the CL setting in both efficiency and accuracy. We also evaluate the effectiveness of SparCL on a real mobile phone, further indicating the practical potential of our method.

Deep learning methods have been very successful at radio frequency fingerprinting tasks, predicting the identity of transmitting devices with high accuracy. We study radio frequency fingerprinting deployments at resource-constrained edge devices. We use structured pruning to jointly train and sparsify neural networks tailored to edge hardware implementations. We compress convolutional layers by a$27.2\times$factor while incurring a negligible prediction accuracy decrease (less than 1 percent). We demonstrate the efficacy of our approach over multiple edge hardware platforms, including a Samsung Gallaxy S10 phone and a Xilinx-ZCU104 FPGA. Our method yields significant inference speedups,$11.5\times$on the FPGA and$3\times$on the smartphone, as well as high efficiency: the FPGA processing time is$17\times$smaller than in a V100 GPU. To the best of our knowledge, we are the first to explore the possibility of compressing networks for radio frequency fingerprinting; as such, our experiments can be seen as a means of characterizing the informational capacity associated with this specific learning task.

Accurate beam alignment in the millimeter-wave (mmWave) band introduces considerable overheads involving brute-force exploration of multiple beam-pair combinations and beam retraining due to mobility. This cost becomes often intractable under high mobility scenarios, where fast beamforming algorithms that can quickly adapt the beam configurations are still under development for 5G and beyond. Besides, blockage prediction is a key capability in order to establish mmWave reliable links. In this paper, we propose a data fusion approach that takes inputs from visual edge devices and localization sensors to (i) reduce the beam selection overhead by narrowing down the search to a small set containing the best possible beam-pairs and (ii) detect blockage conditions between transmitters and receivers. We evaluate our approach through joint simulation of multi-modal data from vision and localization sensors and RF data. Additionally, we show how deep learning based fusion of images and Global Positioning System (GPS) data can play a key role in configuring vehicle-to-infrastructure (V2I) mmWave links. We show a 90% top-10 beam selection accuracy and a 92.86% blockage prediction accuracy. Furthermore, the proposed approach achieves a 99.7% reduction on the beam selection time while keeping a 94.86% of the maximum achievable throughput.

We investigate the HSIC (Hilbert-Schmidt independence criterion) bottleneck as a regularizer for learning an adversarially robust deep neural network classifier. In addition to the usual cross-entropy loss, we add regularization terms for every intermediate layer to ensure that the latent representations retain useful information for output prediction while reducing redundant information. We show that the HSIC bottleneck enhances robustness to adversarial attacks both theoretically and experimentally. In particular, we prove that the HSIC bottleneck regularizer reduces the sensitivity of the classifier to adversarial examples. Our experiments on multiple benchmark datasets and architectures demonstrate that incorporating an HSIC bottleneck regularizer attains competitive natural accuracy and improves adversarial robustness, both with and without adversarial examples during training. Our code and adversarially robust models are publicly available.

Most predictive models based on gene expression data do not leverage information related to gene splicing, despite the fact that splicing is a fundamental feature of eukaryotic gene expression. Cigarette smoking is an important environmental risk factor for many diseases, and it has profound effects on gene expression. Using smoking status as a prediction target, we developed deep neural network predictive models using gene, exon, and isoform level quantifications from RNA sequencing data in 2,557 subjects in the COPDGene Study. We observed that models using exon and isoform quantifications clearly outperformed gene-level models when using data from 5 genes from a previously published prediction model. Whereas the test set performance of the previously published model was 0.82 in the original publication, our exon-based models including an exon-to-isoform mapping layer achieved a test set AUC (area under the receiver operating characteristic) of 0.88, which improved to an AUC of 0.94 using exon quantifications from a larger set of genes. Isoform variability is an important source of latent information in RNA-seq data that can be used to improve clinical prediction models.

In lifelong learning, we wish to maintain and update a model (e.g., a neural network classifier) in the presence of new classification tasks that arrive sequentially. In this paper, we propose a learn-prune-share (LPS) algorithm which addresses the challenges of catastrophic forgetting, parsimony, and knowledge reuse simultaneously. LPS splits the network into task-specific partitions via an ADMM-based pruning strategy. This leads to no forgetting, while maintaining parsimony. Moreover, LPS integrates a novel selective knowledge sharing scheme into this ADMM optimization framework. This enables adaptive knowledge sharing in an end-to-end fashion. Comprehensive experimental results on two lifelong learning benchmark datasets and a challenging real world radio frequency fingerprinting dataset are provided to demonstrate the effectiveness of our approach. Our experiments show that LPS consistently outperforms multiple state-of-the-art competitors.

We study an Open-World Class Discovery problem in which, given labeled training samples from old classes, we need to discover new classes from unlabeled test samples. There are two critical challenges to addressing this paradigm: (a) transferring knowledge from old to new classes, and (b) incorporating knowledge learned from new classes back to the original model. We propose Class Discovery Kernel Network with Expansion (CD-KNet-Exp), a deep learning framework, which utilizes the Hilbert Schmidt Independence Criterion to bridge supervised and unsupervised information together in a systematic way, such that the learned knowledge from old classes is distilled appropriately for discovering new classes. Compared to competing methods, CD-KNet-Exp shows superior performance on three publicly available benchmark datasets and a challenging real-world radio frequency fingerprinting dataset.

In many learning problems, the domain scientist is often interested in discovering the groups of features that are redundant and are important for classification. Moreover, the features that belong to each group, and the important feature groups may vary per sample. But what do we mean by feature redundancy? In this paper, we formally define two types of redundancies using information theory: \textit{Representation} and \textit{Relevant redundancies}. We leverage these redundancies to design a formulation for instance-wise feature group discovery and reveal a theoretical guideline to help discover the appropriate number of groups. We approximate mutual information via a variational lower bound and learn the feature group and selector indicators with Gumbel-Softmax in optimizing our formulation. Experiments on synthetic data validate our theoretical claims. Experiments on MNIST, Fashion MNIST, and gene expression datasets show that our method discovers feature groups with high classification accuracies.