Agentic workflows, where multiple AI agents collaborate to accomplish complex tasks like reasoning or planning, play a substantial role in many cutting-edge commercial applications. These workflows depend critically on the prompts used to provide the roles models play in such workflows. Poorly designed prompts that fail even slightly to guide individual agents can lead to sub-optimal performance that may snowball within a system of agents, limiting their reliability and scalability. To address this important problem of inference-time prompt optimization, we introduce ProRefine, an innovative inference-time optimization method that uses an agentic loop of LLMs to generate and apply textual feedback. ProRefine dynamically refines prompts for multi-step reasoning tasks without additional training or ground truth labels. Evaluated on five benchmark mathematical reasoning datasets, ProRefine significantly surpasses zero-shot Chain-of-Thought baselines by 3 to 37 percentage points. This approach not only boosts accuracy but also allows smaller models to approach the performance of their larger counterparts. This highlights its potential for building cost-effective and powerful hybrid AI systems, thereby democratizing access to high-performing AI.

Multi-agent large language model (LLM) systems enable complex, long-horizon reasoning by composing specialized agents, but practical deployment remains hindered by inefficient routing, noisy feedback, and high interaction cost. We introduce REDEREF, a lightweight and training-free controller for multi-agent LLM collaboration that improves routing efficiency during recursive delegation. REDEREF integrates (i) belief-guided delegation via Thompson sampling to prioritize agents with historically positive marginal contributions, (ii) reflection-driven re-routing using a calibrated LLM or programmatic judge, (iii) evidence-based selection rather than output averaging, and (iv) memory-aware priors to reduce cold-start inefficiency. Across multi-agent split-knowledge tasks, we show that while recursive retry alone saturates task success, belief-guided routing reduces token usage by 28%, agent calls by 17%, and time-to-success by 19% compared to random recursive delegation, and adapts gracefully under agent or judge degradation. These results demonstrate that simple, interpretable probabilistic control can meaningfully improve the efficiency and robustness of multi-agent LLM systems without training or fine-tuning.

Instruction tuning is critical for adapting large language models (LLMs) to downstream tasks, and recent studies have demonstrated that small amounts of human-curated data can outperform larger datasets, challenging traditional data scaling laws. While LLM-based data quality rating systems offer a cost-effective alternative to human annotation, they often suffer from inaccuracies and biases, even in powerful models like GPT-4. In this work, we introduce $DS^2$, a **D**iversity-aware **S**core curation method for **D**ata **S**election. By systematically modeling error patterns through a score transition matrix, $DS^2$ corrects LLM-based scores and promotes diversity in the selected data samples. Our approach shows that a curated subset (just 3.3\% of the original dataset) outperforms full-scale datasets (300k samples) across various machine-alignment benchmarks, and matches or surpasses human-aligned datasets such as LIMA with the same sample size (1k samples). These findings challenge conventional data scaling assumptions, highlighting that redundant, low-quality samples can degrade performance and reaffirming that ``more can be less''.

Recent advancements in large language models have significantly improved their context windows, yet challenges in effective long-term memory management remain. We introduce MemTree, an algorithm that leverages a dynamic, tree-structured memory representation to optimize the organization, retrieval, and integration of information, akin to human cognitive schemas. MemTree organizes memory hierarchically, with each node encapsulating aggregated textual content, corresponding semantic embeddings, and varying abstraction levels across the tree's depths. Our algorithm dynamically adapts this memory structure by computing and comparing semantic embeddings of new and existing information to enrich the model’s context-awareness. This approach allows MemTree to handle complex reasoning and extended interactions more effectively than traditional memory augmentation methods, which often rely on flat lookup tables. Evaluations on benchmarks for multi-turn dialogue understanding and document question answering show that MemTree significantly enhances performance in scenarios that demand structured memory management.

Unlearning in Large Language Models (LLMs) is essential for ensuring ethical and responsible AI use, especially in addressing privacy leak, bias, safety, and evolving regulations. Existing approaches to LLM unlearning often rely on retain data or a reference LLM, yet they struggle to adequately balance unlearning performance with overall model utility. This challenge arises because leveraging explicit retain data or implicit knowledge of retain data from a reference LLM to fine-tune the model tends to blur the boundaries between the forgotten and retain data, as different queries often elicit similar responses. In this work, we propose eliminating the need to retain data or the reference LLM for response calibration in LLM unlearning. Recognizing that directly applying gradient ascent on the forget data often leads to optimization instability and poor performance, our method guides the LLM on what not to respond to, and importantly, how to respond, based on the forget data. Hence, we introduce Forget data only Loss AjustmenT (FLAT), a "flat" loss adjustment approach which addresses these issues by maximizing $f$-divergence between the available template answer and the forget answer only w.r.t. the forget data. The variational form of the defined $f$-divergence theoretically provides a way of loss adjustment by assigning different importance weights for the learning w.r.t. template responses and the forgetting of responses subject to unlearning. Empirical results demonstrate that our approach not only achieves superior unlearning performance compared to existing methods but also minimizes the impact on the model’s retained capabilities, ensuring high utility across diverse tasks, including copyrighted content unlearning on Harry Potter dataset and MUSE Benchmark, and entity unlearning on the TOFU dataset.

We describe KVLink, an approach for efficient key-value (KV) cache reuse in large language models (LLMs). In many LLM applications, different inputs can share overlapping context, such as the same retrieved document appearing in multiple queries. However, the LLMs still need to encode the entire context for each query, leading to redundant computation. In this paper, we investigate a new strategy to eliminate such inefficiency, where the KV cache of each document is precomputed independently. During inference, the KV caches of retrieved documents are concatenated, allowing the model to reuse cached representations instead of recomputing them. To mitigate the performance degradation when using KV caches computed independently for each document, KVLink introduces two key techniques: adjusting positional embeddings of the KV cache at inference to match the global position after concatenation, and using trainable special tokens to restore self-attention across independently encoded documents. Experiments across 7 datasets demonstrate that KVLink improves question answering accuracy by an average of 4% over state-of-the-art methods. Furthermore, by leveraging precomputed KV caches, our approach reduces time-to-first-token by up to 96% compared to standard LLM inference, making it a scalable and efficient solution for context reuse. Additionally, KVLink can be combined with KV cache compression to further save cache loading and storage overhead while outperforming the baselines.

Adapting state-of-the-art Large Language Models (LLMs) like GPT-4 and Gemini for specific tasks is challenging. Due to the opacity in their parameters, embeddings, and even output probabilities, existing fine-tuning adaptation methods are inapplicable. Consequently, adapting these black-box LLMs is only possible through their API services, raising concerns about transparency, privacy, and cost. To address these challenges, we introduce BBox-Adapter, a novel lightweight adapter for black-box LLMs. BBox-Adapter distinguishes target and source domain data by treating target data as positive and source data as negative. It employs a ranking-based Noise Contrastive Estimation (NCE) loss to promote the likelihood of target domain data while penalizing that of the source domain. Furthermore, it features an online adaptation mechanism, which incorporates real-time positive data sampling from ground-truth, human, or AI feedback, coupled with negative data from previous adaptations. Extensive experiments demonstrate BBox-Adapter’s effectiveness and cost efficiency. It improves model performance by up to 6.77% across diverse tasks and domains, while reducing training and inference costs by 31.30x and 1.84x, respectively.

Understanding visually-rich business documents to extract structured data and automate business workflows has been receiving attention both in academia and industry. Although recent multi-modal language models have achieved impressive results, we find that existing benchmarks do not reflect the complexity of real documents seen in industry. In this work, we identify the desiderata for a more comprehensive benchmark and propose one we call Visually Rich Document Understanding (VRDU). VRDU contains two datasets that represent several challenges: rich schema including diverse data types as well as hierarchical entities, complex templates including tables and multi-column layouts, and diversity of different layouts (templates) within a single document type. We design few-shot and conventional experiment settings along with a carefully designed matching algorithm to evaluate extraction results. We report the performance of strong baselines and offer three observations: (1) generalizing to new document templates is still very challenging, (2) few-shot performance has a lot of headroom, and (3) models struggle with hierarchical fields such as line-items in an invoice. We plan to open source the benchmark and the evaluation toolkit. We hope this helps the community make progress on these challenging tasks in extracting structured data from visually rich documents.

Reliable automatic evaluation of dialogue systems under an interactive environment has long been overdue.An ideal environment for evaluating dialog systems, also known as the Turing test, needs to involve human interaction, which is usually not affordable for large scale experiments.Though researchers have attempted to use metrics for language generation tasks (e.g., perplexity, BLEU) or some model-based reinforcement learning methods (e.g., self-play evaluation) for automatic evaluation, these methods only show very weak correlation with the actual human evaluation in practice.To bridge such a gap, we propose a new framework named ENIGMA for estimating human evaluation scores based on recent advances of off-policy evaluation in reinforcement learning.ENIGMA only requires a handful of pre-collected experience data, and therefore does not involve human interaction with the target policy during the evaluation, making automatic evaluations feasible.More importantly, ENIGMA is model-free and agnostic to the behavior policies for collecting the experience data (see details in Section 2), which significantly alleviates the technical difficulties of modeling complex dialogue environments and human behaviors.Our experiments show that ENIGMA significantly outperforms existing methods in terms of correlation with human evaluation scores.

In this paper, we propose a new method to over-come catastrophic forgetting by adding generative regularization to Bayesian inference frame-work. Bayesian method provides a general frame-work for continual learning. We could further construct a generative regularization term for all given classification models by leveraging energy-based models and Langevin dynamic sampling to enrich the features learned in each task. By combining discriminative and generative loss together, we empirically show that the proposed method outperforms state-of-the-art methods on a variety of tasks, avoiding catastrophic forgetting in continual learning. In particular, the proposed method outperforms baseline methods over 15%on the Fashion-MNIST dataset and 10%on the CUB dataset.

Deep Knowledge Tracing (DKT), which traces a student’s knowledge change using deep recurrent neural networks, is widely adopted in student cognitive modeling. Current DKT models only predict a student’s performance based on the observed learning history. However, a student’s learning processes often contain latent events not directly observable in the learning history, such as partial understanding, making slips, and guessing answers. Current DKT models fail to model this kind of stochasticity in the learning process. To address this issue, we propose Variational Deep Knowledge Tracing (VDKT), a latent variable DKT model that incorporates stochasticity into DKT through latent variables. We show that VDKT outperforms both a sequence-to-sequence DKT baseline and previous SoTA methods on MAE, F1, and AUC by evaluating our approach on two Duolingo language learning datasets. We also draw various interpretable analyses from VDKT and offer insights into students’ stochastic behaviors in language learning.

Conventional Neural Architecture Search (NAS) aims at finding a single architecture that achieves the best performance, which usually optimizes task related learning objectives such as accuracy. However, a single architecture may not be representative enough for the whole dataset with high diversity and variety. Intuitively, electing domain-expert architectures that are proficient in domain-specific features can further benefit architecture related objectives such as latency. In this paper, we propose InstaNAS—an instance-aware NAS framework—that employs a controller trained to search for a “distribution of architectures” instead of a single architecture; This allows the model to use sophisticated architectures for the difficult samples, which usually comes with large architecture related cost, and shallow architectures for those easy samples. During the inference phase, the controller assigns each of the unseen input samples with a domain expert architecture that can achieve high accuracy with customized inference costs. Experiments within a search space inspired by MobileNetV2 show InstaNAS can achieve up to 48.8% latency reduction without compromising accuracy on a series of datasets against MobileNetV2.

Online continual learning is a challenging scenario where a model needs to learn from a continuous stream of data without revisiting any previously encountered data instances. The phenomenon of catastrophic forgetting is worsened since the model should not only address the forgetting at the task-level but also at the data instance-level within the same task. To mitigate this, we leverage the concept of "instance awareness" in the neural network, where each data instance is classified by a path in the network searched by the controller from a meta-graph. To preserve the knowledge we learn from previous instances, we proposed a method to protect the path by restricting the gradient updates of one instance from overriding past updates calculated from previous instances if these instances are not similar. On the other hand, it also encourages fine-tuning the path if the incoming instance shares the similarity with previous instances. The mechanism of selecting paths according to instances similarity is naturally determined by the controller, which is compact and online updated. Experimental results show that the proposed method outperforms state-of-the-arts in online continual learning. Furthermore, the proposed method is evaluated against a realistic setting where the boundaries between tasks are blurred. Experimental results confirm that the proposed method outperforms the state-of-the-arts on CIFAR-10, CIFAR-100, and Tiny-ImageNet.

Finding the k largest or smallest elements from a collection of scores, i.e., top-k operation, is an important model component widely used in information retrieval, machine learning, and data mining. However, if the top-k operation is implemented in an algorithmic way, e.g., using bubble algorithm, the resulted model cannot be trained in an end-to-end way using prevalent gradient descent algorithms. This is because these implementations typically involve swapping indices, whose gradient cannot be computed. Moreover, the corresponding mapping from the input scores to the indicator vector of whether this element belongs to the top-k set is essentially discontinuous. To address the issue, we propose a smoothed approximation, namely SOFT (Scalable Optimal transport-based diFferenTiable) top-k operator. Specifically, our SOFT top-k operator approximates the output of top-k operation as the solution of an Entropic Optimal Transport (EOT) problem. The gradient of the SOFT operator can then be efficiently approximated based on the optimality conditions of EOT problem. We then apply the proposed operator to k-nearest neighbors algorithm and beam search algorithm. The numerical experiment demonstrates their achieve improved performance.

This work examines the robustness of selfattentive neural networks against adversarial input perturbations.Specifically, we investigate the attention and feature extraction mechanisms of state-of-the-art recurrent neural networks and self-attentive architectures for sentiment analysis, entailment and machine translation under adversarial attacks.We also propose a novel attack algorithm for generating more natural adversarial examples that could mislead neural models but not humans.Experimental results show that, compared to recurrent neural models, self-attentive models are more robust against adversarial perturbation.In addition, we provide theoretical explanations for their superior robustness to support our claims.

Adversarial robustness has emerged as an important topic in deep learning as carefully crafted attack samples can significantly disturb the performance of a model. Many recent methods have proposed to improve adversarial robustness by utilizing adversarial training or model distillation, which adds additional procedures to model training. In this paper, we propose a new training paradigm called Guided Complement Entropy (GCE) that is capable of achieving "adversarial defense for free," which involves no additional procedures in the process of improving adversarial robustness. In addition to maximizing model probabilities on the ground-truth class like cross-entropy, we neutralize its probabilities on the incorrect classes along with a "guided" term to balance between these two terms. We show in the experiments that our method achieves better model robustness with even better performance compared to the commonly used cross-entropy training objective. We also show that our method can be used orthogonal to adversarial training across well-known methods with noticeable robustness gain. To the best of our knowledge, our approach is the first one that improves model robustness without compromising performance.

Humans can only interact with part of the surrounding environment due to biological restrictions. Therefore, we learn to reason the spatial relationships across a series of observations to piece together the surrounding environment. Inspired by such behavior and the fact that machines also have computational constraints, we propose COnditional COordinate GAN (COCO-GAN) of which the generator generates images by parts based on their spatial coordinates as the condition. On the other hand, the discriminator learns to justify realism across multiple assembled patches by global coherence, local appearance, and edge-crossing continuity. Despite the full images are never manipulated during training, we show that COCO-GAN can produce state-of-the-art-quality full images during inference. We further demonstrate a variety of novel applications enabled by our coordinate-aware framework. First, we perform extrapolation to the learned coordinate manifold and generate off-the-boundary patches. Combining with the originally generated full image, COCO-GAN can produce images that are larger than training samples, which we called "beyond-boundary generation". We then showcase panorama generation within a cylindrical coordinate system that inherently preserves horizontally cyclic topology. On the computation side, COCO-GAN has a built-in divide-and-conquer paradigm that reduces memory requisition during training and inference, provides high-parallelism, and can generate parts of images on-demand.

Neural Architecture Search (NAS) has been quite successful in constructing state-of-the-art models on a variety of tasks. Unfortunately, the computational cost can make it difficult to scale. In this paper, we make the first attempt to study Meta Architecture Search which aims at learning a task-agnostic representation that can be used to speed up the process of architecture search on a large number of tasks. We propose the Bayesian Meta Architecture SEarch (BASE) framework which takes advantage of a Bayesian formulation of the architecture search problem to learn over an entire set of tasks simultaneously. We show that on Imagenet classification, we can find a model that achieves 25.7% top-1 error and 8.1% top-5 error by adapting the architecture in less than an hour from an 8 GPU days pretrained meta-network. By learning a good prior for NAS, our method dramatically decreases the required computation cost while achieving comparable performance to current state-of-the-art methods - even finding competitive models for unseen datasets with very quick adaptation. We believe our framework will open up new possibilities for efficient and massively scalable architecture search research across multiple tasks.

The growth of the Web in recent years has resulted in the development of various online platforms that provide healthcare information services. These platforms contain an enormous amount of information, which could be beneficial for a large number of people. However, navigating through such knowledgebases to answer specific queries of healthcare consumers is a challenging task. A majority of such queries might be non-factoid in nature, and hence, traditional keyword-based retrieval models do not work well for such cases. Furthermore, in many scenarios, it might be desirable to get a short answer that sufficiently answers the query, instead of a long document with only a small amount of useful information. In this paper, we propose a neural network model for ranking documents for question answering in the healthcare domain. The proposed model uses a deep attention mechanism at word, sentence, and document levels, for efficient retrieval for both factoid and non-factoid queries, on documents of varied lengths. Specifically, the word-level cross-attention allows the model to identify words that might be most relevant for a query, and the hierarchical attention at sentence and document levels allows it to do effective retrieval on both long and short documents. We also construct a new large-scale healthcare question-answering dataset, which we use to evaluate our model. Experimental evaluation results against several state-of-the-art baselines show that our model outperforms the existing retrieval techniques.

Accurate identification and localization of abnormalities from radiology images play an integral part in clinical diagnosis and treatment planning. Building a highly accurate prediction model for these tasks usually requires a large number of images manually annotated with labels and finding sites of abnormalities. In reality, however, such annotated data are expensive to acquire, especially the ones with location annotations. We need methods that can work well with only a small amount of location annotations. To address this challenge, we present a unified approach that simultaneously performs disease identification and localization through the same underlying model for all images. We demonstrate that our approach can effectively leverage both class information as well as limited location annotation, and significantly outperforms the comparative reference baseline in both classification and localization tasks.

Recent breakthroughs in Neural Architectural Search (NAS) have achieved state-of-the-art performance in many tasks such as image classification and language understanding. However, most existing works only optimize for model accuracy and largely ignore other important factors imposed by the underlying hardware and devices, such as latency and energy, when making inference. In this paper, we first introduce the problem of NAS and provide a survey on recent works. Then we deep dive into two recent advancements on extending NAS into multiple-objective frameworks: MONAS [19] and DPP-Net [10]. Both MONAS and DPP-Net are capable of optimizing accuracy and other objectives imposed by devices, searching for neural architectures that can be best deployed on a wide spectrum of devices: from embedded systems and mobile devices to workstations. Experimental results are poised to show that architectures found by MONAS and DPP-Net achieves Pareto optimality w.r.t the given objectives for various devices.

Real-time location inference of social media users is the fundamental of some spatial applications such as localized search and event detection. While tweet text is the most commonly used feature in location estimation, most of the prior works suffer from either the noise or the sparsity of textual features. In this paper, we aim to tackle these two problems. We use topic modeling as a building block to characterize the geographic topic variation and lexical variation so that "one-hot" encoding vectors will no longer be directly used. We also incorporate other features which can be extracted through the Twitter streaming API to overcome the noise problem. Experimental results show that our RATE algorithm outperforms several benchmark methods, both in the precision of region classification and the mean distance error of latitude and longitude regression.

Social identities carry widely agreed upon meanings, called stereotypes, that have important effects on social processes. In the present work, we develop a method to extract the stereotypes of Twitter users. Our method is grounded in two distinct strands of theory, one that represents stereotypes as identities' affective meanings and the other that represents stereotypes as semantic relationships between identities. After validating our approach via a prediction task, we apply the model to a dataset of 45 thousand Twitter users who actively tweeted about the Michael Brown and Eric Garner tragedies. Our work provides unique insights into the stereotypes of these users, as well as providing a way of quantifying stereotypes that blends existing sociological and psychological theory in a novel, parsimonious way.

Due to the rapid increase in the number of users owning location-based devices, there is a considerable amount of geo-tagged data available on social media websites, such as Twitter and Facebook. This geo-tagged data can be useful in a variety of ways to extract location-specific information, as well as to comprehend the variation of information across different geographical regions. A lot of techniques have been proposed for extracting location-based information from social media, but none of these techniques aim to utilize an important characteristic of this data, which is the presence of aspects and their opinions, expressed by the users on these platforms. In this paper, we propose Geographic Aspect Opinion model (GASPOP), a probabilistic model that jointly discovers the variation of aspect and opinion, that correspond to different topics across various geographical regions from geo-tagged social media data. It incorporates the syntactic features of text in the generative process to differentiate aspect and opinion words from general background words. The user-based modeling of topics, also enables it to determine the interest distribution of various users. Furthermore, our model can be used to predict the location of different tweets based on their text. We evaluated our model on Twitter data, and our experimental results show that GASPOP can jointly discover latent aspect and opinion words for different topics across latent geographical regions. Moreover, a quantitative analysis of GASPOP using widely used evaluation metrics shows that it outperforms the state-of-the-art methods.

Sociologists have long been interested in the ways that identities, or labels for people, are created, used and applied across various social contexts. The present work makes two contributions to the study of identity, in particular the study of identity in text. We first consider the following novel NLP task: given a set of text data (here, from Twitter), label each word in the text as being representative of a (possibly multi-word) identity. To address this task, we develop a comprehensive feature set that leverages several avenues of recent NLP work on Twitter and use these features to train a supervised classifier. Our model outperforms a surprisingly strong rule-based baseline by 33%. We then use our model for a case study, applying it to a large corpora of Twitter data from users who actively discussed the Eric Garner and Michael Brown cases. Among other findings, we observe that the identities used by individuals differ in interesting ways based on social context measures derived from census data.

Social media is rapidly becoming one of the mediums of choice for understanding the cultural pulse of a region; e.g., for identifying what the population is concerned with and what kind of help is needed in a crisis. To assess this cultural pulse it is critical to have an accurate assessment of who is saying what in social media. However, social media is also the home of malicious users engaged in disruptive, disingenuous, and potentially illegal activity. A range of users, both human and non-human, carry out such social cyber-attacks. We ask, to what extent does the presence or absence of such users influence our ability to assess the cultural pulse of a region? We conduct a series of experiments to analyze the fragility of social network assessments based on Twitter data by comparing changes in both the structural and content results when suspended users are left in and taken out. Because a Twitter account can be suspended for various reasons including spamming or spreading ideas that can lead to extremism or terrorism, we separately assess the impacts of removing apparent spam bots and apparent extremists. Experimental results demonstrate that Twitter-based network structures and content are unstable, and can be highly impacted by the removal of suspended users. Further, the results exhibit regional and temporal variation that may be related to the political situation or civil unrest. We also provides guidance on the differential impact of different types of potentially suspendable users.

Given social networks over time, how can we measure network activities across different timesteps with a limited number of metrics? We propose two classes of dynamic metrics for assessing temporal evolution patterns of agents in terms of persistency and emergence. For each class of dynamic metrics, we implement it using three different temporal aggregation models ranging from the most commonly used Average Aggregation Model to more the complex models such as the Exponential Aggregation Model. We argue that the problem of measuring temporal patterns can be formulated using Recency and Primacy effect, which is a concept used to characterize human cognitive processes. Experimental results show that the way metrics model Recency--Primacy effect is closely related to their abilities to measure temporal patterns. Furthermore, our results indicate that future network agent activities can be predicted based on history information using dynamic metrics. By conducting multiple experiments, we are also able to find an optimal length of history information that is most relevant to future activities. This optimal length is highly consistent within a dataset and can be used as an intrinsic metric to evaluate a dynamic social network.