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, Palash Goyal, Xin Liu, Yiwen Song, Mingyang Ling, Chitta Baral, Hamid Palangi, Tomas Pfister. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 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.

Ashima Suvarna, Christina A Chance, Karolina Naranjo, Hamid Palangi, Sophie Hao, Thomas Hartvigsen, Saadia Gabriel. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

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.

Language models (LMs) are increasingly being deployed to perform autonomous data analyses. However, their data awareness—the ability to recognize, reason over, and appropriately handle data artifacts such as missing values, outliers, and logical inconsistencies—remains underexplored. These artifacts are especially common in real-world tabular data and, if mishandled, can significantly compromise the validity of analytical conclusions. To address this gap, we present RADAR, a benchmark for systematically evaluating data-aware reasoning on tabular data. We develop a framework to simulate data artifacts via programmatic perturbations to enable targeted evaluation of model behavior. RADAR comprises 2,980 table-query pairs, grounded in real-world data spanning 9 domains and 5 data artifact types. In addition to evaluating artifact handling, RADAR systematically varies table size to study how reasoning performance holds when increasing table size. Our evaluation reveals that, despite decent performance on tables without data artifacts, frontier models degrade significantly when data artifacts are introduced, exposing critical gaps in their capacity for robust, data-aware analysis. Designed to be flexible and extensible, RADAR supports diverse perturbation types and controllable table sizes, offering a valuable resource for advancing tabular reasoning.

As AI grows more powerful, it will increasingly shape how we understand the world. But with this influence comes the risk of amplifying misinformation and deepening social divides—especially on consequential topics where factual accuracy directly impacts well-being. Scalable Oversight aims to ensure AI systems remain truthful even when their capabilities exceed those of their evaluators. Yet when humans serve as evaluators, their own beliefs and biases can impair judgment. We study whether AI debate can guide biased judges toward the truth by having two AI systems debate opposing sides of controversial factuality claims on COVID-19 and climate change where people hold strong prior beliefs. We conduct two studies. Study I recruits human judges with either mainstream or skeptical beliefs who evaluate claims through two protocols: debate (interaction with two AI advisors arguing opposing sides) or consultancy (interaction with a single AI advisor). Study II uses AI judges with and without human-like personas to evaluate the same protocols. In Study I, debate consistently improves human judgment accuracy and confidence calibration, outperforming consultancy by 4-10\% across COVID-19 and climate change claims. The improvement is most significant for judges with mainstream beliefs (up to +15.2\% accuracy on COVID-19 claims), though debate also helps skeptical judges who initially misjudge claims move toward accurate views (+4.7\% accuracy). In Study II, AI judges with human-like personas achieve even higher accuracy (78.5\%) than human judges (70.1\%) and default AI judges without personas (69.8\%), suggesting their potential for supervising frontier AI models. These findings highlight AI debate as a promising path toward scalable, bias-resilient oversight in contested domains.

Giovanni Monea, Maxime Peyrard, Martin Josifoski, Vishrav Chaudhary, Jason Eisner, Emre Kiciman, Hamid Palangi, Barun Patra, Robert West. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Large language models (LLMs) frequently hallucinate on abstractive summarization tasks such as document-based question-answering, meeting summarization, and clinical report generation, even though all necessary information is included in context. However, optimizing to make LLMs hallucinate less is challenging, as hallucination is hard to efficiently, cheaply, and reliably evaluate at each optimization step. In this work, we show that reducing hallucination on a _synthetic task_ can also reduce hallucination on real-world downstream tasks. Our method, SynTra, first designs a synthetic task where hallucinations are easy to elicit and measure. It next optimizes the LLM's system message via prefix tuning on the synthetic task, then uses the system message on realistic, hard-to-optimize tasks. Across three realistic abstractive summarization tasks, we reduce hallucination for two 13B-parameter LLMs using supervision signal from only a synthetic retrieval task. We also find that optimizing the system message rather than the model weights can be critical; fine-tuning the entire model on the synthetic task can counterintuitively _increase_ hallucination. Overall, SynTra demonstrates that the extra flexibility of working with synthetic data can help mitigate undesired behaviors in practice.

We investigate the internal behavior of Transformer-based Large Language Models (LLMs) when they generate factually incorrect text. We propose modeling factual queries as constraint satisfaction problems and use this framework to investigate how the LLM interacts internally with factual constraints. We find a strong positive relationship between the LLM's attention to constraint tokens and the factual accuracy of generations. We curate a suite of 10 datasets containing over 40,000 prompts to study the task of predicting factual errors with the Llama-2 family across all scales (7B, 13B, 70B). We propose SAT Probe, a method probing attention patterns, that can predict factual errors and fine-grained constraint satisfaction, and allow early error identification. The approach and findings take another step towards using the mechanistic understanding of LLMs to enhance their reliability.

Data-driven predictive solutions predominant in commercial applications tend to suffer from biases and stereotypes, which raises equity concerns. Prediction models may discover, use, or amplify spurious correlations based on gender or other protected personal characteristics, thus discriminating against marginalized groups. Mitigating gender bias has become an important research focus in natural language processing (NLP) and is an area where annotated corpora are available. Data augmentation reduces gender bias by adding counterfactual examples to the training dataset. In this work, we show that some of the examples in the augmented dataset can be not important or even harmful to fairness. We hence propose a general method for pruning both the factual and counterfactual examples to maximize the model’s fairness as measured by the demographic parity, equality of opportunity, and equality of odds. The fairness achieved by our method surpasses that of data augmentation on three text classification datasets, using no more than half of the examples in the augmented dataset. Our experiments are conducted using models of varying sizes and pre-training settings. WARNING: This work uses language that is offensive in nature.

We have seen a great progress in video action recognition in recent years. There are several models based on convolutional neural network (CNN) and some recent transformer based approaches which provide top performance on existing benchmarks. In this work, we perform a large-scale robustness analysis of these existing models for video action recognition. We focus on robustness against real-world distribution shift perturbations instead of adversarial perturbations. We propose four different benchmark datasets, HMDB51-P, UCF101-P, Kinetics400-P, and SSv2-P to perform this analysis. We study robustness of six state-of-the-art action recognition models against 90 different perturbations. The study reveals some interesting findings, 1) transformer based models are consistently more robust compared to CNN based models, 2) Pretraining improves robustness for Transformer based models more than CNN based models, and 3) All of the studied models are robust to temporal perturbations for all datasets but SSv2; suggesting the importance of temporal information for action recognition varies based on the dataset and activities. Next, we study the role of augmentations in model robustness and present a real-world dataset, UCF101-DS, which contains realistic distribution shifts, to further validate some of these findings. We believe this study will serve as a benchmark for future research in robust video action recognition11More details available at bit.1y/3TJLMUF..

Spurious correlations that degrade model generalization or lead the model to be right for the wrong reasons are one of the main robustness concerns for real-world deployments. However, mitigating these correlations during pre-training for large-scale models can be costly and impractical, particularly for those without access to high-performance computing resources. This paper proposes a novel approach to address spurious correlations during fine-tuning for a given domain of interest. With a focus on multi-modal models (e.g., CLIP), the proposed method leverages different modalities in these models to detect and explicitly set apart spurious attributes from the affected class, achieved through a multi-modal contrastive loss function that expresses spurious relationships through language. Our experimental results and in-depth visualizations on CLIP show that such an intervention can effectively i) improve the model’s accuracy when spurious attributes are not present, and ii) directs the model’s activation maps towards the actual class rather than the spurious attribute when present. In particular, on the Waterbirds dataset, our algorithm achieved a worst-group accuracy 23% higher than ERM on CLIP with a ResNet-50 backbone, and 32% higher on CLIP with a ViT backbone, while maintaining the same average accuracy as ERM.

Deployed language models decay over time due to shifting inputs, changing user needs, or emergent world-knowledge gaps. When such problems are identified, we want to make targeted edits while avoiding expensive retraining. However, current model editors, which modify such behaviors of pre-trained models, degrade model performance quickly across multiple, sequential edits. We propose GRACE, a \textit{lifelong} model editing method, which implements spot-fixes on streaming errors of a deployed model, ensuring minimal impact on unrelated inputs. GRACE writes new mappings into a pre-trained model's latent space, creating a discrete, local codebook of edits without altering model weights. This is the first method enabling thousands of sequential edits using only streaming errors. Our experiments on T5, BERT, and GPT models show GRACE's state-of-the-art performance in making and retaining edits, while generalizing to unseen inputs. Our code is available at [github.com/thartvigsen/grace](https://www.github.com/thartvigsen/grace}).

Recently an influx of studies claims emergent cognitive abilities in large language models (LLMs). Yet, most rely on anecdotes, overlook contamination of training sets, or lack systematic Evaluation involving multiple tasks, control conditions, multiple iterations, and statistical robustness tests. Here we make two major contributions. First, we propose CogEval, a cognitive science-inspired protocol for the systematic evaluation of cognitive capacities in LLMs. The CogEval protocol can be followed for the evaluation of various abilities. Second, here we follow CogEval to systematically evaluate cognitive maps and planning ability across eight LLMs (OpenAI GPT-4, GPT-3.5-turbo-175B, davinci-003-175B, Google Bard, Cohere-xlarge-52.4B, Anthropic Claude-1-52B, LLaMA-13B, and Alpaca-7B). We base our task prompts on human experiments, which offer both established construct validity for evaluating planning, and are absent from LLM training sets. We find that, while LLMs show apparent competence in a few planning tasks with simpler structures, systematic evaluation reveals striking failure modes in planning tasks, including hallucinations of invalid trajectories and falling in loops. These findings do not support the idea of emergent out-of-the-box planning ability in LLMs. This could be because LLMs do not understand the latent relational structures underlying planning problems, known as cognitive maps, and fail at unrolling goal-directed trajectories based on the underlying structure. Implications for application and future directions are discussed.

Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, Ece Kamar. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Joint visual and language modeling on large-scale datasets has recently shown good progress in multi-modal tasks when compared to single modal learning. However, robustness of these approaches against real-world perturbations has not been studied. In this work, we perform the first extensive robustness study of video-language models against various real-world perturbations. We focus on text-to-video retrieval and propose two large-scale benchmark datasets, MSRVTT-P and YouCook2-P, which utilize 90 different visual and 35 different text perturbations. The study reveals some interesting initial findings from the studied models: 1) models are more robust when text is perturbed versus when video is perturbed, 2) models that are pre-trained are more robust than those trained from scratch, 3) models attend more to scene and objects rather than motion and action. We hope this study will serve as a benchmark and guide future research in robust video-language learning. The benchmark introduced in this study along with the code and datasets is available at https://bit.ly/3CNOly4.

Yichen Jiang, Asli Celikyilmaz, Paul Smolensky, Paul Soulos, Sudha Rao, Hamid Palangi, Roland Fernandez, Caitlin Smith, Mohit Bansal, Jianfeng Gao. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

This paper presents a unified Vision-Language Pre-training (VLP) model. The model is unified in that (1) it can be fine-tuned for either vision-language generation (e.g., image captioning) or understanding (e.g., visual question answering) tasks, and (2) it uses a shared multi-layer transformer network for both encoding and decoding, which differs from many existing methods where the encoder and decoder are implemented using separate models. The unified VLP model is pre-trained on a large amount of image-text pairs using the unsupervised learning objectives of two tasks: bidirectional and sequence-to-sequence (seq2seq) masked vision-language prediction. The two tasks differ solely in what context the prediction conditions on. This is controlled by utilizing specific self-attention masks for the shared transformer network. To the best of our knowledge, VLP is the first reported model that achieves state-of-the-art results on both vision-language generation and understanding tasks, as disparate as image captioning and visual question answering, across three challenging benchmark datasets: COCO Captions, Flickr30k Captions, and VQA 2.0. The code and the pre-trained models are available at https://github.com/LuoweiZhou/VLP.

Generating formal-language programs represented by relational tuples, such as Lisp programs or mathematical operations, to solve problems stated in natural language is a challenging task because it requires explicitly capturing discrete symbolic structural information implicit in the input. However, most general neural sequence models do not explicitly capture such structural information, limiting their performance on these tasks. In this paper, we propose a new encoder-decoder model based on a structured neural representation, Tensor Product Representations (TPRs), for mapping Natural-language problems to Formal-language solutions, called TP-N2F. The encoder of TP-N2F employs TPR ‘binding’ to encode natural-language symbolic structure in vector space and the decoder uses TPR ‘unbinding’ to generate, in symbolic space, a sequential program represented by relational tuples, each consisting of a relation (or operation) and a number of arguments. TP-N2F considerably outperforms LSTM-based seq2seq models on two benchmarks and creates new state-of-the-art results. Ablation studies show that improvements can be attributed to the use of structured TPRs explicitly in both the encoder and decoder. Analysis of the learned structures shows how TPRs enhance the interpretability of TP-N2F.

We introduce an architecture, the Tensor Product RecurrentNetwork (TPRN). In our application of TPRN, internal representations—learned by end-to-end optimization in a deep neural network performing a textual question-answering(QA) task—can be interpreted using basic concepts from linguistic theory. No performance penalty need be paid for this increased interpretability: the proposed model performs comparably to a state-of-the-art system on the SQuAD QA task.The internal representation which is interpreted is a Tensor Product Representation: for each input word, the model selects a symbol to encode the word, and a role in which to place the symbol, and binds the two together. The selection is via soft attention. The overall interpretation is built from interpretations of the symbols, as recruited by the trained model, and interpretations of the roles as used by the model. We find support for our initial hypothesis that symbols can be interpreted as lexical-semantic word meanings, while roles can be interpreted as approximations of grammatical roles (or categories)such as subject, wh-word, determiner, etc. Fine-grained analysis reveals specific correspondences between the learned roles and parts of speech as assigned by a standard tagger(Toutanova et al. 2003), and finds several discrepancies in the model’s favor. In this sense, the model learns significant aspectsof grammar, after having been exposed solely to linguistically unannotated text, questions, and answers: no prior linguistic knowledge is given to the model. What is given is the means to build representations using symbols and roles, with an inductive bias favoring use of these in an approximately discrete manner.

Robust detection of epileptic seizures in the presence of inevitable artifacts in Electroencephalogram (EEG) signals is addressed. The EEG dataset considered contains 300 signals recorded from 15 volunteers. Current seizure detection systems achieve good performance when the EEG data is entirely free of noise. However, their performance drastically decays with authentic EEG data polluted by real artifacts. We introduce a robust seizure detection method that can address clean and noisy data. The proposed method uses Long Short-Term Memory (LSTM) neural networks to extract the representative EEG features pertinent to seizures. Experimental results show that the proposed method beats existing methods by achieving 100% classification accuracy. Our method is also shown to be robust against the common EEG artifacts (e.g., muscle activities and eye-blinking) and white noise.

Image hashing has attracted increasing popularity in recent years. Some off-the-shelf image hashing methods are able to generate more compact and robust hashes for fast indexing and content-based similarity retrieval. However, the ability to infer original image contents from their real-valued image hashes has seldom been examined. Inherited from cryptographic hashing for image privacy protection, general image hashing is supposed to be a non-revertible function. Should there be a way to revert (or perceptually reconstruct) images from the corresponding real-valued image hashes? This paper explores the feasibility of perceptually image hashing reversion, and fill this gap by proposing a deep learning based framework, entitled RevHashNet. Given real-valued image hashes from certain image hashing methods, the proposed RevHashNet can automatically reconstruct perceptually similar images with respect to the original ones with high visual quality. Experiments and simulations on real image datasets support the de-hashing effectiveness of the proposed RevHashNet.

This paper addresses the compressive sensing with Multiple Measurement Vectors (MMV) problem where the correlation amongst the different sparse vectors (channels) are used to improve the reconstruction performance. We propose the use of Convolutional Deep Stacking Networks (CDSN), where the correlations amongst the channels are captured by a moving window containing the "residuals" of different sparse vectors. We develop a greedy algorithm that exploits the structure captured by the CDSN to reconstruct the sparse vectors. Using a natural image dataset, we compare the performance of the proposed algorithm with two types of reconstruction algorithms: Simultaneous Orthogonal Matching Pursuit (SOMP) which is a greedy solver and the model-based Bayesian approaches that also exploit correlation among channels. We show experimentally that our proposed method outperforms these popular methods and is almost as fast as the greedy methods.

This paper develops a model that addresses sentence embedding, a hot topic in current natural language processing research, using recurrent neural networks (RNN) with Long Short-Term Memory (LSTM) cells. The proposed LSTM-RNN model sequentially takes each word in a sentence, extracts its information, and embeds it into a semantic vector. Due to its ability to capture long term memory, the LSTM-RNN accumulates increasingly richer information as it goes through the sentence, and when it reaches the last word, the hidden layer of the network provides a semantic representation of the whole sentence. In this paper, the LSTM-RNN is trained in a weakly supervised manner on user click-through data logged by a commercial web search engine. Visualization and analysis are performed to understand how the embedding process works. The model is found to automatically attenuate the unimportant words and detect the salient keywords in the sentence. Furthermore, these detected keywords are found to automatically activate different cells of the LSTM-RNN, where words belonging to a similar topic activate the same cell. As a semantic representation of the sentence, the embedding vector can be used in many different applications. These automatic keyword detection and topic allocation abilities enabled by the LSTM-RNN allow the network to perform document retrieval, a difficult language processing task, where the similarity between the query and documents can be measured by the distance between their corresponding sentence embedding vectors computed by the LSTM-RNN. On a web search task, the LSTM-RNN embedding is shown to significantly outperform several existing state of the art methods. We emphasize that the proposed model generates sentence embedding vectors that are specially useful for web document retrieval tasks. A comparison with a well known general sentence embedding method, the Paragraph Vector, is performed. The results show that the proposed method in this paper significantly outperforms Paragraph Vector method for web document retrieval task.

We study the MMV (Multiple Measurement Vectors) compressive sensing setting with a specific sparse structured support. The locations of the non-zero rows in the sparse matrix are not known. All that is known is that the locations of the non-zero rows have probabilities that vary from one group of rows to another. We propose two novel greedy algorithms for the exact recovery of the sparse matrix in this structured MMV compressive sensing problem. The first algorithm models the matrix sparse structure using a shallow non- linear neural network. The input of this network is the residual matrix after the prediction and the output is the sparse matrix to be recovered. The second algorithm improves the shallow neural network prediction by using the stacking operation to form a deep stacking network. Experimental evaluation demonstrates the superior performance of both new algorithms over existing MMV methods. Among all, the algorithm using the deep stacking network for modelling the structure in MMV compressive sensing performs the best.