arXiv subjects

Jae Hee Lee

Publications and source records attributed to Jae Hee Lee.

At least 19 recordsLinked to original sources

arXiv2026

Explaining, Verifying, and Aligning Semantic Hierarchies in Vision-Language Model Embeddings

Vision-language model (VLM) encoders such as CLIP enable strong retrieval and zero-shot classification in a shared image-text embedding space, yet the semantic organization of this space is rarely inspected. We present a post-hoc framework to explain, verify, and align the semantic hierarchies induced by a VLM over a given set of child classes. First, we extract a binary hierarchy by agglomerative clustering of class centroids and name internal nodes by dictionary-based matching to a concept bank. Second, we quantify plausibility by comparing the extracted tree against human ontologies using efficient tree- and edge-level consistency measures, and we evaluate utility via explainable hierarchical tree-traversal inference with uncertainty-aware early stopping (UAES). Third, we propose an ontology-guided post-hoc alignment method that learns a lightweight embedding-space transformation, using UMAP to generate target neighborhoods from a desired hierarchy. Across 13 pretrained VLMs and 4 image datasets, our method finds systematic modality differences: image encoders are more discriminative, while text encoders induce hierarchies that better match human taxonomies. Overall, the results reveal a persistent trade-off between zero-shot accuracy and ontological plausibility and suggest practical routes to improve semantic alignment in shared embedding spaces.

cs.LG↗

arXiv2026

3D mirror symmetry in positive characteristic

Via the formulation of (quantum) Hikita conjecture with coefficients in a characteristic $p$ field, we explain an arithmetic aspect of the theory of 3D mirror symmetry. Namely, we propose that the action of Steenrod-type operations and Frobenius-constant quantizations intertwine under the (quantum) Hikita isomorphism for 3D mirror pairs, and verify this for the Springer resolutions and hypertoric varieties.

math.RT↗

arXiv2026

Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders

Vision Language Models (VLMs) have demonstrated impressive performance in tasks requiring joint understanding of images and text, such as image captioning and Visual Question Answering (VQA), but our understanding of their internal processes remains limited. Recently, Sparse Autoencoders (SAEs) have emerged as a promising tool to support the interpretation of concepts encoded in VLMs. However, most SAE-based approaches focus only on textual or visual concepts separately, ignoring multimodal concepts. This limitation hinders a comprehensive understanding of VLMs, since concepts that integrate both modalities can be misclassified. Moreover, previous visual approaches often produce low-quality visual concept descriptions that are vague or incomplete, limiting their usefulness for understanding model reasoning. We propose a framework based on SAEs to extract and analyze visual, textual, and multimodal concepts from VLMs. For each neuron, we propose a candidate human-interpretable concept and compute the alignment between the concept and the dataset samples using cosine similarity scores. Experiments on a VQA dataset (LLaVA-NeXT) demonstrate that our framework improves visual concept quality by up to 45\% compared to existing SAE-based methods, while maintaining high textual concept quality and enabling systematic identification of multimodal concepts. This work contributes new insights into the conceptual space of VLMs, providing a structured approach to distinguish between visual, textual, and multimodal concepts. The code is available at https://github.com/PHDLanza/Multidata_SAE

cs.CV↗

arXiv2026

Learning to Hear Hesitation: Continual Learning for Disfluency-Aware ASR

Despite advances in large-scale Automatic Speech Recognition (ASR), disfluent speech remains challenging, as state-of-the-art systems are often optimized to omit disfluencies, leading to information loss and hallucinations. Prior work has focused on verbatim transcription and the integration of disfluency markers, but adapting models on limited datasets can lead to catastrophic forgetting of general-domain knowledge. We address this gap by leveraging continual learning (CL) with explicit disfluency tokens. We first introduce these tokens into a pretrained ASR model to establish stable token mechanisms, and then continue training on additional datasets with varying disfluency distributions. Through a detailed analysis of model dynamics during training, we identify a trade-off between marker learning and ASR performance, and a consistent cross-attention head mechanism shared across CL methods.

cs.CL↗

arXiv2026

Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions

Vision classifiers can exploit spurious correlations, achieving high in-distribution accuracy yet failing under distribution shift. Existing approaches to bias mitigation and analysis often depend on curated datasets, spurious-attribute or group labels, or retraining, which may be infeasible once a model is deployed or the relevant bias is unknown. We present a bias-label-free, post-hoc method for identifying spurious concepts in frozen vision models, relying only on standard class labels from a held-out audit dataset. For each target class, we collect patches from inputs predicted as that class and apply non-negative matrix factorization to intermediate activations to obtain a bank of interpretable concept vectors. Candidate concepts are then ranked with a bias estimator derived from their interaction with backpropagated gradients on misclassified examples: bias concepts tend to get activated when correcting false negatives and suppressed when correcting false positives. On Colored MNIST and Waterbirds the method recovers concepts aligned with the known spurious cue, and on CelebA it surfaces decision-relevant directions that only partially coincide with the annotated gender attribute; suppressing the top-ranked concepts at inference time improves worst-group accuracy by up to 17.9 percentage points on Waterbirds and 10.4 on CelebA without any retraining or parameter updates. Our method identifies decision-relevant spurious directions that need not coincide with annotated ones, providing both an interpretable auditing tool and an actionable debiasing handle for frozen vision models. Code is available at https://github.com/vitryt/label-free-bias-identification.

cs.CV↗

arXiv2026

The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level

Mixture-of-Experts (MoE) architectures have become the dominant choice for scaling Large Language Models (LLMs), activating only a subset of parameters per token. While MoE architectures are primarily adopted for computational efficiency, it remains an open question whether their sparsity makes them inherently easier to interpret than dense feed-forward networks (FFNs). We compare MoE experts and dense FFNs using $k$-sparse probing and find that expert neurons are consistently less polysemantic, with the gap widening as routing becomes sparser. This suggests that sparsity pressures both individual neurons and entire experts toward monosemanticity. Leveraging this finding, we zoom out from the neuron to the expert level as a more effective unit of analysis. We validate this approach by automatically interpreting hundreds of experts. This analysis allows us to resolve the debate on specialization: experts are neither broad domain specialists (e.g., biology) nor simple token-level processors. Instead, they function as fine-grained task experts, specializing in linguistic operations or semantic tasks (e.g., closing brackets in $\LaTeX{}$). Our findings suggest that MoEs are inherently interpretable at the expert level, providing a clearer path toward large-scale model interpretability. Code is available at: https://github.com/jerryy33/MoE_analysis.

cs.CL↗

arXiv2026

Arithmetic geometry of quantum connections on Calabi-Yau $3$-folds

Fix a prime $p > 3$. Working over $\mathbb{Z}_p$, we show that the quantum connection of any closed Calabi-Yau threefold gives rise to a Fontaine-Laffaile module when restricted to the even degree and torsion-free part of $p$-adic quantum cohomology, whose associated Frobenius endomorphism has leading order term prescribed by the $p$-adic Gamma class. After reducing mod $p$, the divided Frobenius endomorphism defines an analogue of the inverse Cartier operator on mod $p$ quantum cohomology. We establish an $A$-model analogue of a classical result due to Katz: the conjugation of the $p$-curvature of the mod $p$ quantum connection by the inverse Cartier operator is equal to the Frobenius pullback of the quantum product, the $A$-model counterpart of the Kodaira-Spencer class. Moreover, we identify the quantum Steenrod operation with the $p$-curvature of the mod $p$ quantum connection in this setting for any prime $p$. We propose several conjectures concerning how these arithmetic structures may extend to quantum connections on more general semi-positive symplectic manifolds.

math.SG↗

arXiv2026

Knowing the Facts but Choosing the Shortcut: Understanding How Large Language Models Compare Entities

Large Language Models (LLMs) are increasingly used for knowledge-based reasoning tasks, yet understanding when they rely on genuine knowledge versus superficial heuristics remains challenging. We investigate this question through entity comparison tasks by asking models to compare entities along numerical attributes (e.g., ``Which river is longer, the Danube or the Nile?''), which offer clear ground truth for systematic analysis. Despite having sufficient numerical knowledge to answer correctly, LLMs frequently make predictions that contradict this knowledge. We identify three heuristic biases that strongly influence model predictions: entity popularity, mention order, and semantic co-occurrence. For smaller models, a simple logistic regression using only these surface cues predicts model choices more accurately than the model's own numerical predictions, suggesting heuristics largely override principled reasoning. Crucially, we find that larger models (32B parameters) selectively rely on numerical knowledge when it is more reliable, while smaller models (7--8B parameters) show no such discrimination, which explains why larger models outperform smaller ones even when the smaller models possess more accurate knowledge. Chain-of-thought prompting steers all models towards using the numerical features across all model sizes.

cs.CL↗

arXiv2025

Towards Ethical Multi-Agent Systems of Large Language Models: A Mechanistic Interpretability Perspective

Large language models (LLMs) have been widely deployed in various applications, often functioning as autonomous agents that interact with each other in multi-agent systems. While these systems have shown promise in enhancing capabilities and enabling complex tasks, they also pose significant ethical challenges. This position paper outlines a research agenda aimed at ensuring the ethical behavior of multi-agent systems of LLMs (MALMs) from the perspective of mechanistic interpretability. We identify three key research challenges: (i) developing comprehensive evaluation frameworks to assess ethical behavior at individual, interactional, and systemic levels; (ii) elucidating the internal mechanisms that give rise to emergent behaviors through mechanistic interpretability; and (iii) implementing targeted parameter-efficient alignment techniques to steer MALMs towards ethical behaviors without compromising their performance.

cs.AI↗

arXiv2025

Quantum Adams operations in quasimap K-theory

We define quantum deformations of Adams operations in $K$-theory, in the framework of quasimap quantum $K$-theory. They provide $K$-theoretic analogs of the quantum Steenrod operations from equivariant symplectic Gromov--Witten theory. We verify the compatibility of these operations with the Kahler and equivariant $q$-difference module structures, provide sample computations via $\mathbb{Z}/k$-equivariant localization, and identify them with $p$-curvature operators of the Kahler $q$-difference connections as studied in Koroteev-Smirnov. We also formulate and verify a $K$-theoretic quantum Hikita conjecture at roots of unity, and propose an indirect algebro-geometric definition of quantum Steenrod operations

math.AG↗

arXiv2025

Quantum Steenrod operations of symplectic resolutions

We study the mod $p$ equivariant quantum cohomology of conical symplectic resolutions. Using symplectic genus zero enumerative geometry, Fukaya and Wilkins defined operations on mod $p$ quantum cohomology deforming the classical Steenrod operations on mod $p$ cohomology. We conjecture that these quantum Steenrod operations on divisor classes agree with the $p$-curvature of the mod $p$ equivariant quantum connection, and verify this in the case of the Springer resolution. The key ingredient is a new compatibility relation between the quantum Steenrod operations and the shift operators.

math.SG↗

arXiv2025

Flat endomorphisms for mod $p$ equivariant quantum connections from quantum Steenrod operations

We present a method for constructing covariantly constant endomorphisms for the mod $p$ equivariant quantum connection, using the quantum Steenrod power operations of Fukaya and Wilkins. The example of the cotangent bundle of the projective line is fully computed, and we discuss the relationship with the mod $p$ solutions of trigonometric KZ equation recently constructed by Varchenko. As a byproduct, we compute the first examples of quantum Steenrod operations that are not a priori determined by ordinary Gromov--Witten theory and classical Steenrod operations, which may be of independent interest.

math.SG↗

arXiv2024

Details Make a Difference: Object State-Sensitive Neurorobotic Task Planning

The state of an object reflects its current status or condition and is important for a robot's task planning and manipulation. However, detecting an object's state and generating a state-sensitive plan for robots is challenging. Recently, pre-trained Large Language Models (LLMs) and Vision-Language Models (VLMs) have shown impressive capabilities in generating plans. However, to the best of our knowledge, there is hardly any investigation on whether LLMs or VLMs can also generate object state-sensitive plans. To study this, we introduce an Object State-Sensitive Agent (OSSA), a task-planning agent empowered by pre-trained neural networks. We propose two methods for OSSA: (i) a modular model consisting of a pre-trained vision processing module (dense captioning model, DCM) and a natural language processing model (LLM), and (ii) a monolithic model consisting only of a VLM. To quantitatively evaluate the performances of the two methods, we use tabletop scenarios where the task is to clear the table. We contribute a multimodal benchmark dataset that takes object states into consideration. Our results show that both methods can be used for object state-sensitive tasks, but the monolithic approach outperforms the modular approach. The code for OSSA is available at https://github.com/Xiao-wen-Sun/OSSA

cs.AI↗

arXiv2024

Concept-Based Explanations in Computer Vision: Where Are We and Where Could We Go?

Concept-based XAI (C-XAI) approaches to explaining neural vision models are a promising field of research, since explanations that refer to concepts (i.e., semantically meaningful parts in an image) are intuitive to understand and go beyond saliency-based techniques that only reveal relevant regions. Given the remarkable progress in this field in recent years, it is time for the community to take a critical look at the advances and trends. Consequently, this paper reviews C-XAI methods to identify interesting and underexplored areas and proposes future research directions. To this end, we consider three main directions: the choice of concepts to explain, the choice of concept representation, and how we can control concepts. For the latter, we propose techniques and draw inspiration from the field of knowledge representation and learning, showing how this could enrich future C-XAI research.

cs.CV↗

arXiv2024

Read Between the Layers: Leveraging Multi-Layer Representations for Rehearsal-Free Continual Learning with Pre-Trained Models

We address the Continual Learning (CL) problem, wherein a model must learn a sequence of tasks from non-stationary distributions while preserving prior knowledge upon encountering new experiences. With the advancement of foundation models, CL research has pivoted from the initial learning-from-scratch paradigm towards utilizing generic features from large-scale pre-training. However, existing approaches to CL with pre-trained models primarily focus on separating class-specific features from the final representation layer and neglect the potential of intermediate representations to capture low- and mid-level features, which are more invariant to domain shifts. In this work, we propose LayUP, a new prototype-based approach to CL that leverages second-order feature statistics from multiple intermediate layers of a pre-trained network. Our method is conceptually simple, does not require access to prior data, and works out of the box with any foundation model. LayUP surpasses the state of the art in four of the seven class-incremental learning benchmarks, all three domain-incremental learning benchmarks and in six of the seven online continual learning benchmarks, while significantly reducing memory and computational requirements compared to existing baselines. Our results demonstrate that fully exhausting the representational capacities of pre-trained models in CL goes well beyond their final embeddings.

cs.LG↗

arXiv2024

Mental Modeling of Reinforcement Learning Agents by Language Models

Can emergent language models faithfully model the intelligence of decision-making agents? Though modern language models exhibit already some reasoning ability, and theoretically can potentially express any probable distribution over tokens, it remains underexplored how the world knowledge these pretrained models have memorized can be utilized to comprehend an agent's behaviour in the physical world. This study empirically examines, for the first time, how well large language models (LLMs) can build a mental model of agents, termed agent mental modelling, by reasoning about an agent's behaviour and its effect on states from agent interaction history. This research may unveil the potential of leveraging LLMs for elucidating RL agent behaviour, addressing a key challenge in eXplainable reinforcement learning (XRL). To this end, we propose specific evaluation metrics and test them on selected RL task datasets of varying complexity, reporting findings on agent mental model establishment. Our results disclose that LLMs are not yet capable of fully mental modelling agents through inference alone without further innovations. This work thus provides new insights into the capabilities and limitations of modern LLMs.

cs.LG↗

arXiv2024

Language Model-Based Paired Variational Autoencoders for Robotic Language Learning

Human infants learn language while interacting with their environment in which their caregivers may describe the objects and actions they perform. Similar to human infants, artificial agents can learn language while interacting with their environment. In this work, first, we present a neural model that bidirectionally binds robot actions and their language descriptions in a simple object manipulation scenario. Building on our previous Paired Variational Autoencoders (PVAE) model, we demonstrate the superiority of the variational autoencoder over standard autoencoders by experimenting with cubes of different colours, and by enabling the production of alternative vocabularies. Additional experiments show that the model's channel-separated visual feature extraction module can cope with objects of different shapes. Next, we introduce PVAE-BERT, which equips the model with a pretrained large-scale language model, i.e., Bidirectional Encoder Representations from Transformers (BERT), enabling the model to go beyond comprehending only the predefined descriptions that the network has been trained on; the recognition of action descriptions generalises to unconstrained natural language as the model becomes capable of understanding unlimited variations of the same descriptions. Our experiments suggest that using a pretrained language model as the language encoder allows our approach to scale up for real-world scenarios with instructions from human users.

cs.NE↗

arXiv2024

From Neural Activations to Concepts: A Survey on Explaining Concepts in Neural Networks

In this paper, we review recent approaches for explaining concepts in neural networks. Concepts can act as a natural link between learning and reasoning: once the concepts are identified that a neural learning system uses, one can integrate those concepts with a reasoning system for inference or use a reasoning system to act upon them to improve or enhance the learning system. On the other hand, knowledge can not only be extracted from neural networks but concept knowledge can also be inserted into neural network architectures. Since integrating learning and reasoning is at the core of neuro-symbolic AI, the insights gained from this survey can serve as an important step towards realizing neuro-symbolic AI based on explainable concepts.

cs.AI↗