Yiqiao Jin, Kartik Sharma, Vineeth Rakesh, Yingtong Dou, Menghai Pan, Mahashweta Das, Srijan Kumar. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

While LLMs are widely used for generic tasks like question answering and search, they struggle to adapt to specialized knowledge, such as industrial workflows in healthcare, legal, and agricultural sectors, as well as knowledgedriven tasks such as news journalism, investigative research, and consulting without expensive fine-tuning or sub-optimal retrieval methods.Existing retrieval-augmented models, such as RAG, offer improvements but fail to account for structured domain knowledge, leading to suboptimal context generation.Ontologies, which conceptually organize domain knowledge by defining entities and their interrelationships, offer a structured representation to address this gap.This paper presents OG-RAG, an Ontology-Grounded Retrieval Augmented Generation method designed to enhance LLMgenerated responses by anchoring retrieval processes in domain-specific ontologies.OG-RAG constructs a hypergraph representation of domain documents, where each hyperedge encapsulates clusters of factual knowledge grounded using domain-specific ontology and retrieves a minimal set of hyperedges for a given query using an optimization algorithm.Our evaluations demonstrate that OG-RAG increases the recall of accurate facts by 55% and improves response correctness by 40% across four different LLMs.Additionally, OG-RAG enables 30% faster attribution of responses to context and boosts factbased reasoning accuracy by 27% compared to baseline methods.We release the code at https://github.com/microsoft/ograg2.

Effective human-AI coordination requires artificial agents capable of exhibiting and responding to human-like behaviors while adapting to changing contexts. Imitation learning has emerged as one of the prominent approaches to build such agents by training them to mimic human-demonstrated behaviors. However, current methods struggle to capture the inherent diversity and non-Markovian nature of human behavior and lack the ability to steer behavior at inference time. Drawing inspiration from the theory of human cognitive processes, where inner speech guides action selection before execution, we propose MIMIC (Modeling Inner Motivations for Imitation and Control), a framework that uses language as an internal representation of behavioral intent. MIMIC employs the novel use of vision-language models as linguistic scaffolding to train a conditional variational autoencoder capable of generating inner speech from observations. A diffusion-based behavior cloning policy then selects actions conditioned on current observations and the generated inner speech. MIMIC enables fine-grained steering of behavior at inference time by conditioning the agent on behavior-specific speech. Experiments across robotic manipulation tasks and human-AI collaboration games demonstrate that MIMIC significantly enhances both behavior diversity and fidelity to human demonstrations while enabling nuanced behavioral steering without training on additional demonstrations.

Diffusion models lend transformative capabilities to the graph generation task, yet controlling the properties of the generated graphs remains challenging. Recent approaches augment support for controlling soft, differentiable properties but they fail to handle user-specified hard constraints that are non-differentiable. This often results in vague control, unsuitable for applications like drug discovery that demand satisfaction of precise constraints, e.g., the maximum number of bonds. To address this, we formalize the problem of controlled graph generation and introduce PRODIGY (PROjected DIffusion for controlled Graph Generation), an innovative plug-and-play approach enabling the generation of graphs with precise control, from any pre-trained diffusion model. PRODIGY employs a novel operator to project the samples at each diffusion step onto the specified constrained space. For a large class of practical constraints and a variety of graphs, our extensive experiments demonstrate that PRODIGY empowers state-of-the-art continuous and discrete diffusion models to produce graphs meeting specific, hard constraints. Our approach achieves up to 100% constraint satisfaction for non-attributed and molecular graphs, under a variety of constraints, marking a significant step forward in precise, interpretable graph generation. Code is provided on the project webpage: https://prodigy-diffusion.github.io/.

In modern System-on-Chip (SoC) design, dependence on offshore fabrication has increased manyfold. This empowers SoC designers to meet stringent time-to-market demands. However, it also brings rogue entities into play, possessing security threats like IP piracy, counterfeiting and overbuilding. Several countermeasures have been proposed to thwart these threats. Among all, logic locking has been the most coveted. Logic locking is a design-for-trust technique which conceals the underlying functionality of the design to be protected by adding extra key-controlled gates. This work proposes a novel attack called NoBALL on logic locked designs using their Binary Decision Diagram (BDD) representation. We identify specific properties from the locked design’s BDD and perform intuitive intersection operations in order to identify the correct keys. Experimental analysis highlights that the current version of the attack can act as an alternative to the SAT attack and can even outperform it in some cases. Our attack can also be combined with SAT attack to form a more robust attack in future.

Adversarial attacks on Graph Neural Networks (GNNs) reveal their security vulnerabilities, limiting their adoption in safety-critical applications. However, existing attack strategies rely on the knowledge of either the GNN model being used or the predictive task being attacked. Is this knowledge necessary? For example, a graph may be used for multiple downstream tasks unknown to a practical attacker. It is thus important to test the vulnerability of GNNs to adversarial perturbations in a model and task-agnostic setting. In this work, we study this problem and show that Gnns remain vulnerable even when the downstream task and model are unknown. The proposed algorithm, TANDIS (Targeted Attack via Neighborhood DIStortion) shows that distortion of node neighborhoods is effective in drastically compromising prediction performance. Although neighborhood distortion is an NP-hard problem, TANDIS designs an effective heuristic through a novel combination of Graph Isomorphism Network with deep Q-learning. Extensive experiments on real datasets show that, on average, TANDIS is up to 50% more effective than state-of-the-art techniques, while being more than 1000 times faster.

Signed networks allow us to model conflicting relationships and interactions, such as friend/enemy and support/oppose. These signed interactions happen in real-time. Modeling such dynamics of signed networks is crucial to understanding the evolution of polarization in the network and enabling effective prediction of the signed structure (i.e., link signs) in the future. However, existing works have modeled either (static) signed networks or dynamic (unsigned) networks but not dynamic signed networks. Since both sign and dynamics inform the graph structure in different ways, it is non-trivial to model how to combine the two features. In this work, we propose a new Graph Neural Network (GNN)-based approach to model dynamic signed networks, named SEMBA: Signed link's Evolution using Memory modules and Balanced Aggregation. Here, the idea is to incorporate the signs of temporal interactions using separate modules guided by balance theory and to evolve the embeddings from a higher-order neighborhood. Experiments on 4 real-world datasets and 3 different tasks demonstrate that SEMBA consistently and significantly outperforms the baselines by up to 80% on the tasks of predicting signs of future links while matching the state-of-the-art performance on predicting existence of these links in the future. We find that this improvement is due specifically to superior performance of SEMBA on the minority negative class. Code is made available at https://github.com/claws-lab/semba.

StyleGANs are at the forefront of controllable image generation as they produce a latent space that is semantically disentangled, making it suitable for image editing and manipulation. However, the performance of StyleGANs severely degrades when trained via class-conditioning on large-scale long-tailed datasets. We find that one reason for degradation is the collapse of latents for each class in the$\mathcal{W}$latent space. With NoisyTwins, we first introduce an effective and inexpensive augmentation strategy for class embeddings, which then decorrelates the latents based on self-supervision in the$\mathcal{W}$space. This decorrelation mitigates collapse, ensuring that our method preserves intra-class diversity with class-consistency in image generation. We show the effectiveness of our approach on large-scale real-world long-tailed datasets of ImageNet-LT and iNaturalist 2019, where our method outperforms other methods by ∼ 19% on FID, establishing a new state-of-the-art.

The problem of community-level information pathway prediction (CLIPP) aims at predicting the transmission trajectory of content across online communities. A successful solution to CLIPP holds significance as it facilitates the distribution of valuable information to a larger audience and prevents the proliferation of misinfor- mation. Notably, solving CLIPP is non-trivial as inter-community relationships and influence are unknown, information spread is multi-modal, and new content and new communities appear over time. In this work, we address CLIPP by collecting large-scale, multi-modal datasets to examine the diffusion of online YouTube videos on Reddit. We analyze these datasets to construct community influence graphs (CIGs) and develop a novel dynamic graph frame- work, INPAC (Information Pathway Across Online Communities), which incorporates CIGs to capture the temporal variability and multi-modal nature of video propagation across communities. Ex- perimental results in both warm-start and cold-start scenarios show that INPAC outperforms seven baselines in CLIPP. Our code and datasets are available at https://github.com/claws-lab/INPAC

Real-world graphs such as social networks, communication networks, and rating networks are constantly evolving over time. Many deep learning architectures have been developed to learn effective node representations using both graph structure and dynamics. While being crucial for practical applications, the robustness of these representation learners for dynamic graphs in the presence of adversarial attacks is highly understudied. In this work, we design a novel adversarial attack on discrete-time dynamic graph models where we desire to perturb the input graph sequence in a manner that preserves the temporal dynamics of the graph while dropping the performance of representation learners. To this end, we motivate a novel Temporal Dynamics-Aware Perturbation (TDAP) constraint, which ensures that perturbations introduced at each time step are restricted to only a small fraction of the number of changes in the graph since the previous time step. We present a theoretically-motivated Projected Gradient Descent approach for dynamic graphs to find effective perturbations under the TDAP constraint. Experiments on two tasks - dynamic link prediction and node classification, show that our approach is up to 4x more effective than the baseline methods for attacking these models. We extend our approach to a more practical online setting where graphs become available in real-time and show up to 5x superior performance over baselines We also show that our approach successfully evades state-of-the-art neural approaches for anomaly detection, thereby promoting the need to study robustness as a part of representation-learning approaches for dynamic graphs.

Early crop yield prediction is crucial in agriculture for making administrative plans to ensure food security, post harvest management and distribution of a crop. Remote sensing data captured using various satellites provide reliable phenological information for a crop through surface reflectance bands. Other important factors, affecting crop yield include meteorological and soil. The data which we have used for crop yield prediction is multimodal. It consists of spatiotemporal meteorological (numeric) and surface reflectance bands (satellite image), and temporally static soil (satellite image) data. We effectively utilize this multimodal data to develop the proposed multimodal deep learning model, CropYieldNet. The objective of the paper is to accurately predict crop yield using high resolution data obtained from recently launched satellites such as Landsat8 and Sentinel-2. We used contrastive learning in a supervised setting and data augmentation techniques to overcome the limited historical data available for training deep learning models.We introduce a depth-level selection module for effectively modelling the depth-variant information of soil data. We have also modified our model to perform in-season (early) crop yield prediction which is as accurate as end-season prediction. We evaluate our model for two crops, corn and soybean, on counties in US and districts in India using data from MODIS, Landsat8, and Sentinel-2 satellites. Our extensive experimentation show that our model outperforms competing models. Our experiments also show that CropYieldNet generalizes well when applied on different crops and geographies.

In signed networks, each edge is labeled as either positive or negative. The edge sign captures the polarity of a relationship. Balance of signed networks is a well-studied property in graph theory. In a balanced (sub)graph, the vertices can be partitioned into two subsets with negative edges present only across the partitions. Balanced portions of a graph have been shown to increase coherence among its members and lead to better performance. While existing works have focused primarily on finding the largest balanced subgraph inside a graph, we study the network design problem of maximizing balance of a target community (subgraph). In particular, given a budget b and a community of interest within the signed network, we aim to make the community as close to being balanced as possible by deleting up to b edges. Besides establishing NP-hardness, we also show that the problem is non-monotone and non-submodular. To overcome these computational challenges, we propose heuristics based on the spectral relation of balance with the Laplacian spectrum of the network. Since the spectral approach lacks approximation guarantees, we further design a greedy algorithm, and its randomized version, with provable bounds on the approximation quality. The bounds are derived by exploiting pseudo-submodularity of the balance maximization function. Empirical evaluation on eight real-world signed networks establishes that the proposed algorithms are effective, efficient, and scalable to graphs with millions of edges.