arXiv subjects

Ali Abbasi

Publications and source records attributed to Ali Abbasi.

At least 19 recordsLinked to original sources

arXiv2026

CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuits

Controlling the behavior of large language models (LLMs) remains a critical challenge for AI alignment. Existing steering methods, such as Contrastive Activation Addition (CAA), typically rely on fixed single-layer interventions derived from aggregate activation differences. These methods impose a single intervention across semantically diverse inputs and often fail to sustain consistent behavioral changes across layers, limiting the effectiveness of the steering. In this work, we introduce CircuitSteer, a novel framework that leverages Sparse Autoencoders (SAEs) to identify and manipulate coherent semantic circuits distributed across multiple layers. By constructing a feature flow circuit based on feature co-activation and the geometric alignment of decoder directions, we isolate the specific multi-layer subcircuits responsible for a target behavior. We then synthesize dense steering vectors from these sparse features and apply multi-point interventions to guide the model's internal semantic trajectory. We evaluate CircuitSteer using contrastive examples across a diverse set of tasks, including toxicity, emotion-intensity, sycophancy, and refusal, spanning two model families. Across all models and datasets, CircuitSteer is the only method to consistently produce fluency-preserving interventions; competing methods either sacrifice text quality or lack coverage, failing entirely on complex behaviors like sycophancy and refusal. These results demonstrate that multi-layer circuit steering, enabled by enforcing geometric alignment among selected features, yields strictly more robust and effective behavioral control than static single-point interventions. Code is available at https://github.com/mehrshad-sdtn/CircuitSteer.

cs.LG↗

arXiv2026

PyFEX: Uncovering Evasive Python-based Threats via Resilient and Exhaustive Path Exploration

The rapid expansion of the Python ecosystem has fueled two distinct but converging threats: adversaries increasingly target the software supply chain via the Python Package Index (PyPI), while also building evasive, cross-platform malicious binaries compiled from source code written in Python. Current program analysis techniques struggle to address this dual threat. Static analysis based tools are often blinded by runtime obfuscation and compiled bytecode, while dynamic analysis based ones are fragile, prone to evasion by environmental guardrails, and often terminates prematurely due to unsatisfied dependencies. To overcome these limitations, we present PyFEX, a resilient forced-execution engine. PyFEX explores a program's behavioral space systematically by forcing execution across all conditional branches to bypass evasion checks. To address the fragility of dynamic execution, it introduces a novel resilient crash recovery mechanism that synthesizes dummy objects to satisfy failed operations at the runtime, allowing analysis to proceed past fatal errors, and employs path merging to mitigate path explosion. PyFEX further incorporates an automated entry identification mechanism that proactively discovers and invokes dormant functions, exposing malicious logic hidden within uncalled APIs. To demonstrate the efficacy of this engine, we built PyFEXScan, a proof-of-concept malware detector built on top of PyFEX. Evaluated against both known malicious PyPI packages and real-world compiled binaries, PyFEX exposes critical behaviors missed by the existing state-of-the-art tools. In a live deployment on PyPI, PyFEXScan discovered 212 previously unknown malicious packages accounting for over 91,648 downloads, underscoring the necessity of resilient, exhaustive analysis for securing the Python ecosystem.

cs.CR↗

arXiv2026

Efficient Algorithms for Interdicting Facilities in Trees and Bounded Treewidth Graphs

Given a graph $G$ of $n$ nodes partitioned into facilities and customers, the $r$-edge interdiction covering problem (REIC) is to remove up to $r$ edges so as to maximize the total weight of customers disconnected from all facilities, which is called the covering objective function. While REIC is known to be NP-complete for general graphs, Fr\"ohlich and Ruzika show that the problem can be solved in polynomial time when $G$ is a tree, providing an $O(n^7 r)$-time algorithm. We give an efficient $O(nr^2)$-time dynamic programming algorithm for REIC on trees that is fixed-parameter linear in $n$. Evaluating our solution on a benchmark of randomly generated tree networks with baselines of the Fr\"ohlich and Ruzika algorithm and the Gurobi integer program solver, we demonstrate that in practice, our algorithm is both significantly faster and less sensitive to network topology and size. We extend our algorithm for REIC to graphs of bounded treewidth, a well-studied family of sparse graphs that generalizes trees, and obtain a matching runtime of $O(nr^2)$. We also consider the $r$-facility interdiction covering problem (RFIC), a novel variant of this network interdiction problem where the goal is to remove up to $r$ facilities to maximize the covering objective function over disconnected customers. We show that RFIC is NP-complete by observing it generalizes the small set bipartite vertex expansion problem (SSBVE), also known as the minimum $p$-union problem. We give an $O(nr^2)$-time algorithm for RFIC on trees, which also gives an $O(n^3)$-time algorithm for SSBVE on trees.

cs.DS↗

arXiv2026

IO-SVD: Input-Output Whitened SVD for Adaptive-Rank LLM Compression

Large language models deliver strong performance across language and reasoning tasks, but their storage and compute costs remain major barriers to deployment in resource-constrained and latency-sensitive settings. SVD-based post-training compression offers a hardware-agnostic way to reduce model size and improve inference efficiency through low-rank factorization. However, existing methods often rely on input-only whitening spaces, homogeneous rank allocation, or loss-agnostic allocation heuristics, limiting their ability to preserve model quality under aggressive compression. We propose Input-Output Whitened SVD (IO-SVD), a post-training compression method that forms a KL-aware double-sided whitening space for model weights. Using a second-order expansion of the KL loss over the top-K token probabilities, IO-SVD constructs an output-side metric that captures predictive sensitivity, while input whitening captures activation statistics. We further introduce an efficient heterogeneous rank-allocation strategy that scores whitened singular components using first-order calibration loss estimates and prunes the least sensitive components under a global budget. Inspired by prior work that combines SVD truncation with quantization, we improve hybrid SVD-quantization compression through loss-aware remapping, which selects low-rank factor rows for 8-bit quantization based on the predicted loss change incurred by quantizing them. Extensive experiments across diverse LLM and VLM families, and inference-time analysis shows that IO-SVD compresses LLMs with minimal performance degradation while delivering practical inference speedups. Code is available at https://github.com/mint-vu/IO-SVD.git

cs.LG↗

arXiv2026

ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs

Large language models (LLMs) are costly to deploy due to their large memory footprint and high inference cost. Weight-activation quantization can reduce these costs, but low-bit activation quantization remains difficult because activation outliers induce large quantization error. Recent rotation-based methods address this by applying orthogonal transformations that redistribute activation magnitude across dimensions, but existing approaches either require expensive end-to-end rotation training or rely on stored activation corpora, introducing significant compute or storage overhead. We propose a lightweight post-training rotation calibration method for LLM activation quantization. Our method learns orthogonal rotations that align normalized activations with the corners of an inscribed hypercube, encouraging activation energy to be distributed more evenly across dimensions. This objective admits an efficient closed-form update via the orthogonal Procrustes problem, avoiding gradient-based optimization over the orthogonal group. We further introduce an online calibration procedure that updates rotations as calibration samples are processed, eliminating the need to store activations on disk and allowing rotations to adapt to quantized activation distributions during calibration. Experiments on Llama-2 and Llama-3 models from 3B to 70B parameters show that our method achieves competitive or improved performance across perplexity benchmarks and common sense reasoning tasks while avoiding both costly end-to-end training and large offline activation storage.

cs.LG↗

arXiv2026

Vector-Quantized Soft Label Compression for Dataset Distillation

Dataset distillation is an emerging technique for reducing the computational and storage costs of training machine learning models by synthesizing a small, informative subset of data that captures the essential characteristics of a much larger dataset. Recent methods pair synthetic samples and their augmentations with soft labels from a teacher model, enabling student models to generalize effectively despite the small size of the distilled dataset. While soft labels are critical for effective distillation, the storage and communication overhead they incur, especially when accounting for augmentations, is often overlooked. In practice, each distilled sample is associated with multiple soft labels, making them the dominant contributor to storage costs, particularly in large-class settings such as ImageNet-1K. In this paper, we present a rigorous analysis of bit requirements across dataset distillation frameworks, quantifying the storage demands of both distilled samples and their soft labels. To address the overhead, we introduce a vector-quantized autoencoder (VQAE) for compressing soft labels, achieving substantial compression while preserving the effectiveness of the distilled data. We validate our method on both vision and language distillation benchmarks. On ImageNet-1K, our proposed VQAE achieves 30--40x additional compression over RDED, LPLD, SRE2L, and CDA baselines while retaining over $90\%$ of their original performance.

cs.CV↗

arXiv2026

NAST: Improving Negation Handling in Medical Vision-Language Models through Negation-Aware Selective Training

Negation is a fundamental linguistic operation in clinical reporting, yet vision-language models (VLMs) frequently fail to distinguish affirmative from negated medical statements. To systematically characterize this limitation, we introduce a radiology-specific diagnostic benchmark that evaluates polarity sensitivity under controlled clinical conditions, revealing that common medical VLMs consistently confuse negated and non-negated findings. To enable learning beyond simple condition absence, we further construct a contextual clinical negation dataset that encodes structured claims and supports attribute-level negations involving location and severity. Building on these resources, we propose Negation-Aware Selective Training (NAST), an interpretability-guided adaptation method that uses causal tracing effects (CTEs) to modulate layer-wise gradient updates during fine-tuning. Rather than applying uniform learning rates, NAST scales each layer's update according to its causal contribution to negation processing, transforming mechanistic interpretability signals into a principled optimization rule. Experiments demonstrate improved discrimination of affirmative and negated clinical statements without degrading general vision-language alignment, highlighting the value of causal interpretability for targeted model adaptation in safety-critical medical settings. Code and resources are available at https://github.com/healthylaife/NAST.

cs.CV↗

arXiv2026

Zero Sum SVD: Balancing Loss Sensitivity for Low Rank LLM Compression

Advances in large language models have driven strong performance across many tasks, but their memory and compute costs still hinder deployment. SVD-based compression reduces storage and can speed up inference via low-rank factors, yet performance depends on how rank is allocated under a global compression ratio. Prior methods often use homogeneous ranks for similarly sized matrices, despite large differences in loss sensitivity, or rely on expensive iterative pre-truncation optimization to determine per matrix ranks. We propose \textbf{Zero Sum SVD} (\textbf{ZS-SVD}), a post-training method that performs \emph{global} singular component selection using activation whitening and first-order calibration loss estimates in whitened coordinates. \textbf{ZS-SVD} prunes components across the whole model with a \textbf{zero sum} rule that keeps the cumulative predicted loss change near zero, automatically yielding heterogeneous ranks without solving a rank allocation optimization. Motivated by evidence that gradients near pretrained solutions exhibit low rank structure, we also introduce an optional lightweight correction that applies a \textbf{single} projected gradient update after truncation, followed by re-truncation. Extensive experiments across multiple LLM architectures show consistent gains across diverse benchmarks and compression ratios. Code is available at https://github.com/mint-vu/Zero-Sum-SVD

cs.LG↗

arXiv2026

A Multimodal Data Processing Pipeline for MIMIC-IV Dataset

The MIMIC-IV dataset is a large, publicly available electronic health record (EHR) resource widely used for clinical machine learning research. It comprises multiple modalities, including structured data, clinical notes, waveforms, and imaging data. Working with these disjointed modalities requires an extensive manual effort to preprocess and align them for downstream analysis. While several pipelines for MIMIC-IV data extraction are available, they target a small subset of modalities or do not fully support arbitrary downstream applications. In this work, we greatly expand our prior popular unimodal pipeline and present a comprehensive and customizable multimodal pipeline that can significantly reduce multimodal processing time and enhance the reproducibility of MIMIC-based studies. Our pipeline systematically integrates the listed modalities, enabling automated cohort selection, temporal alignment across modalities, and standardized multimodal output formats suitable for arbitrary static and time-series downstream applications. We release the code, a simple UI, and a Python package for selective integration (with embedding) at https://github.com/healthylaife/MIMIC-IV-Data-Pipeline.

cs.LG↗

arXiv2025

Anota: Identifying Business Logic Vulnerabilities via Annotation-Based Sanitization

Detecting business logic vulnerabilities is a critical challenge in software security. These flaws come from mistakes in an application's design or implementation and allow attackers to trigger unintended application behavior. Traditional fuzzing sanitizers for dynamic analysis excel at finding vulnerabilities related to memory safety violations but largely fail to detect business logic vulnerabilities, as these flaws require understanding application-specific semantic context. Recent attempts to infer this context, due to their reliance on heuristics and non-portable language features, are inherently brittle and incomplete. As business logic vulnerabilities constitute a majority (27/40) of the most dangerous software weaknesses in practice, this is a worrying blind spot of existing tools. In this paper, we tackle this challenge with ANOTA, a novel human-in-the-loop sanitizer framework. ANOTA introduces a lightweight, user-friendly annotation system that enables users to directly encode their domain-specific knowledge as lightweight annotations that define an application's intended behavior. A runtime execution monitor then observes program behavior, comparing it against the policies defined by the annotations, thereby identifying deviations that indicate vulnerabilities. To evaluate the effectiveness of ANOTA, we combine ANOTA with a state-of-the-art fuzzer and compare it against other popular bug finding methods compatible with the same targets. The results show that ANOTA+FUZZER outperforms them in terms of effectiveness. More specifically, ANOTA+FUZZER can successfully reproduce 43 known vulnerabilities, and discovered 22 previously unknown vulnerabilities (17 CVEs assigned) during the evaluation. These results demonstrate that ANOTA provides a practical and effective approach for uncovering complex business logic flaws often missed by traditional security testing techniques.

cs.CR↗

arXiv2025

LUNA: Linear Universal Neural Attention with Generalization Guarantees

Scaling attention faces a critical bottleneck: the $\mathcal{O}(n^2)$ quadratic computational cost of softmax attention, which limits its application in long-sequence domains. While linear attention mechanisms reduce this cost to $\mathcal{O}(n)$, they typically rely on fixed random feature maps, such as random Fourier features or hand-crafted functions. This reliance on static, data-agnostic kernels creates a fundamental trade-off, forcing practitioners to sacrifice significant model accuracy for computational efficiency. We introduce \textsc{LUNA}, a kernelized linear attention mechanism that eliminates this trade-off, retaining linear cost while matching and surpassing the accuracy of quadratic attention. \textsc{LUNA} is built on the key insight that the kernel feature map itself should be learned rather than fixed a priori. By parameterizing the kernel, \textsc{LUNA} learns a feature basis tailored to the specific data and task, overcoming the expressive limitations of fixed-feature methods. \textsc{Luna} implements this with a learnable feature map that induces a positive-definite kernel and admits a streaming form, yielding linear time and memory scaling in the sequence length. Empirical evaluations validate our approach across diverse settings. On the Long Range Arena (LRA), \textsc{Luna} achieves state-of-the-art average accuracy among efficient Transformers under compute parity, using the same parameter count, training steps, and approximate FLOPs. \textsc{Luna} also excels at post-hoc conversion: replacing softmax in fine-tuned BERT and ViT-B/16 checkpoints and briefly fine-tuning recovers most of the original performance, substantially outperforming fixed linearizations.

cs.LG↗

arXiv2025

Low-Rank Prehab: Preparing Neural Networks for SVD Compression

Low-rank approximation methods such as singular value decomposition (SVD) and its variants (e.g., Fisher-weighted SVD, Activation SVD) have recently emerged as effective tools for neural network compression. In this setting, decomposition acts as a "surgical" intervention, followed by fine-tuning that serves as "rehab" to recover accuracy. Inspired by prehabilitation in surgery, we introduce a pre-compression fine-tuning stage, Low-Rank Prehab, that explicitly encourages low-rank structure in weight matrices while preserving task performance. By conditioning the model before SVD, Prehab steers weights toward spectrally compact regions of the parameter space, enabling smoother low-rank approximation and improved recovery. Experiments on large language models (LLMs) and other Transformer-based architectures, including Vision Transformers (ViTs), show that Prehab substantially reduces the immediate accuracy drop after compression and consistently improves post-finetuning performance. Across a wide range of compression ratios, our method outperforms state-of-the-art SVD-based techniques such as SVD-LLM, highlighting the importance of preparing models for compression rather than only improving the compression and recovery stages. Source code is available at https://github.com/niqretnuh/PREHAB-SVD

cs.LG↗

arXiv2025

Maximum a Posteriori Probability (MAP) Joint Carrier Frequency Offset (CFO) and Channel Estimation for MIMO Channels with Spatial and Temporal Correlations

We consider time varying MIMO fading channels with known spatial and temporal correlation and solve the problem of joint carrier frequency offset (CFO) and channel estimation with prior distributions. The maximum a posteriori probability (MAP) joint estimation is proved to be equivalent to a separate MAP estimation of the CFO followed by minimum mean square error (MMSE) estimation of the channel while treating the estimated CFO as true. The MAP solution is useful to take advantage of the estimates from the previous data packet. A low complexity universal CFO estimation algorithm is extended from the time invariant case to the time varying case. Unlike past algorithms, the universal algorithm does not need phase unwrapping to take advantage of the full range of symbol correlation and achieves the derived Bayesian Cram\'er-Rao lower bound (BCRLB) in almost all SNR range. We provide insight on the the relation among the temporal correlation coefficient of the fading, the CFO estimation performance, and the pilot signal structure. An unexpected observation is that the BCRLB is not a monotone function of the temporal correlation and is strongly influenced by the pilot signal structures. A simple rearrangement of the 0's and 1's in the pilot signal matrix will render the BCRLB from being non-monotone to being monotone in certain temporal correlation ranges. Since the BCRLB is shown to be achieved by the proposed algorithm, it provides a guideline for pilot signal design.

cs.IT↗

arXiv2025

Neural Cellular Automata: From Cells to Pixels

Neural Cellular Automata (NCAs) are bio-inspired dynamical systems in which identical cells iteratively apply a learned local update rule to self-organize into complex patterns, exhibiting regeneration, robustness, and spontaneous dynamics. Despite their success in texture synthesis and morphogenesis, NCAs remain largely confined to low-resolution outputs. This limitation stems from (1) training time and memory requirements that grow quadratically with grid size, (2) the strictly local propagation of information that impedes long-range cell communication, and (3) the heavy compute demands of real-time inference at high resolution. In this work, we overcome this limitation by pairing an NCA that evolves on a coarse grid with a lightweight implicit decoder that maps cell states and local coordinates to appearance attributes, enabling the same model to render outputs at arbitrary resolution. Moreover, because both the decoder and NCA updates are local, inference remains highly parallelizable. To supervise high-resolution outputs efficiently, we introduce task-specific losses for morphogenesis (growth from a seed) and texture synthesis with minimal additional memory and computation overhead. Our experiments across 2D/3D grids and mesh domains demonstrate that our hybrid models produce high-resolution outputs in real-time, and preserve the characteristic self-organizing behavior of NCAs.

cs.CV↗

arXiv2025

HoneySat: A Network-based Satellite Honeypot Framework

Satellites are the backbone of several mission-critical services that enable our modern society to function, for example, GPS. For years, satellites were assumed to be secure because of their indecipherable architectures and the reliance on security by obscurity. However, technological advancements have made these assumptions obsolete, paving the way for potential attacks, and sparking interest in satellite security. Unfortunately, to this day, there is no efficient way to collect data on adversarial techniques for satellites, hurting the generation of security intelligence that can lead to the development of effective countermeasures. In this paper, we present HoneySat, the first high-interaction satellite honeypot framework, fully capable of convincingly simulating a real-world CubeSat, a type of Small Satellite (SmallSat). To provide evidence of HoneySat's effectiveness, we surveyed experienced SmallSat operators in charge of in-orbit satellites and deployed HoneySat over the Internet to entice adversaries. Our results show that 90% of satellite operators agreed that HoneySat provides a realistic and engaging simulation of a SmallSat mission. Additionally, HoneySat successfully deceived human adversaries in the wild and collected 22 real-world satellite-specific adversarial interactions. Finally, in a major demonstration of HoneySat's robustness, we collaborated with an aerospace company to perform a hardware-in-the-loop operation that resulted in HoneySat successfully communicating with an in-orbit, operational SmallSat mission.

cs.CR↗

arXiv2025

Knowledge Distillation and Dataset Distillation of Large Language Models: Emerging Trends, Challenges, and Future Directions

The exponential growth of Large Language Models (LLMs) continues to highlight the need for efficient strategies to meet ever-expanding computational and data demands. This survey provides a comprehensive analysis of two complementary paradigms: Knowledge Distillation (KD) and Dataset Distillation (DD), both aimed at compressing LLMs while preserving their advanced reasoning capabilities and linguistic diversity. We first examine key methodologies in KD, such as task-specific alignment, rationale-based training, and multi-teacher frameworks, alongside DD techniques that synthesize compact, high-impact datasets through optimization-based gradient matching, latent space regularization, and generative synthesis. Building on these foundations, we explore how integrating KD and DD can produce more effective and scalable compression strategies. Together, these approaches address persistent challenges in model scalability, architectural heterogeneity, and the preservation of emergent LLM abilities. We further highlight applications across domains such as healthcare and education, where distillation enables efficient deployment without sacrificing performance. Despite substantial progress, open challenges remain in preserving emergent reasoning and linguistic diversity, enabling efficient adaptation to continually evolving teacher models and datasets, and establishing comprehensive evaluation protocols. By synthesizing methodological innovations, theoretical foundations, and practical insights, our survey charts a path toward sustainable, resource-efficient LLMs through the tighter integration of KD and DD principles.

cs.CL↗

arXiv2024

Diffusion-Augmented Coreset Expansion for Scalable Dataset Distillation

With the rapid scaling of neural networks, data storage and communication demands have intensified. Dataset distillation has emerged as a promising solution, condensing information from extensive datasets into a compact set of synthetic samples by solving a bilevel optimization problem. However, current methods face challenges in computational efficiency, particularly with high-resolution data and complex architectures. Recently, knowledge-distillation-based dataset condensation approaches have made this process more computationally feasible. Yet, with the recent developments of generative foundation models, there is now an opportunity to achieve even greater compression, enhance the quality of distilled data, and introduce valuable diversity into the data representation. In this work, we propose a two-stage solution. First, we compress the dataset by selecting only the most informative patches to form a coreset. Next, we leverage a generative foundation model to dynamically expand this compressed set in real-time, enhancing the resolution of these patches and introducing controlled variability to the coreset. Our extensive experiments demonstrate the robustness and efficiency of our approach across a range of dataset distillation benchmarks. We demonstrate a significant improvement of over 10% compared to the state-of-the-art on several large-scale dataset distillation benchmarks. The code will be released soon.

cs.CV↗

arXiv2024

Next-Token Prediction Task Assumes Optimal Data Ordering for LLM Training in Proof Generation

In the field of large language model (LLM)-based proof generation, despite extensive training on large datasets such as ArXiv, LLMs still exhibit only modest performance on proving tasks of moderate difficulty. We believe that this is partly due to the widespread presence of suboptimal ordering within the data for each proof used in training. For example, published proofs often follow a purely logical order, where each step logically proceeds from the previous steps based on the deductive rules. This order is designed to facilitate the verification of the proof's soundness, rather than to help people and models learn the discovery process of the proof. In proof generation, we argue that the optimal order for one training data sample occurs when the relevant intermediate supervision for a particular proof step in the proof is always positioned to the left of that proof step. We call such order the intuitively sequential order. We validate our claims using two tasks: intuitionistic propositional logic theorem-proving and digit multiplication. Our experiments verify the order effect and provide support for our explanations. We demonstrate that training is most effective when the proof is in the intuitively sequential order. Moreover, the order effect and the performance gap between models trained on different data orders can be substantial -- with an 11 percent improvement in proof success rate observed in the propositional logic theorem-proving task, between models trained on the optimal order compared to the worst order. Lastly, we define a common type of order issue in advanced math proofs and find that 17.3 percent of theorems with nontrivial proofs in the first two chapters of a widely used graduate-level mathematics textbook suffer from this issue. A detailed list of those proofs is provided in the appendix.

cs.CL↗