arXiv subjects

Lucas S. Kupssinskü

Publications and source records attributed to Lucas S. Kupssinskü.

10 recordsLinked to original sources

arXiv2026

Inference-Time Machine Unlearning via Gated Activation Redirection

Large Language Models (LLMs) memorize vast amounts of training data, raising concerns regarding privacy, copyright infringement, and safety. Machine unlearning seeks to remove the influence of a targeted forget set while preserving model performance, ideally approximating a model retrained from scratch without it. Once an LLM is in use, every new request to make it forget specific content demands updating its weights. However, unlearning through parameter updates is expensive, hard to audit, and can be undone by quantization. We show that unlearning can be enforced entirely at inference time, without training, gradients, or weight changes. We introduce Inference-Time Unlearning via Gated Activation Redirection (GUARD-IT), a training- and gradient-free method that unlearns via input-dependent activation steering at inference time. GUARD-IT stores the content to be forgotten as a small library of activation directions, and during inference, it routes each query through a similarity gate that activates only for relevant directions and applies them as a norm-preserving rotation of the residual stream, while unrelated queries pass through the unmodified model. The same design carries across three model families and nine checkpoints from 0.8B to 8B parameters, and new forget requests are absorbed by one offline pass of forward passes. On TOFU, against 16 gradient-based and inference-time baselines, and on MUSE and WMDP, GUARD-IT forgets without breaking the model, and on TOFU it is the only method that suppresses memorization in every Llama configuration without collapsing: it keeps utility and fluent generation in every configuration, moves the model's output distribution closest to a model that never saw the forgotten data on the forget01 split, survives 4- and 8-bit quantization and ten sequential forget requests, and holds under jailbreak attacks.

cs.LG↗

arXiv2026

Continual Learning for Sequential Personalization of Small Language Models: A Stability Monitoring Analysis

Small Language Models (SLMs) are increasingly being considered for deployment on edge devices such as laptops, enabling private, low-latency, and locally personalized applications. However, personalization requires models to adapt over time to evolving user- or task-specific data, placing them in a continual learning setting. This creates the risk of catastrophic forgetting, where learning new information degrades performance on previously learned tasks or broader model capabilities. Recent benchmarks such as TRACE have shown that continual fine-tuning can significantly degrade the general abilities of aligned large language models. In this work, we present a study for sequential LoRA personalization of SLMs. We save model checkpoints after each adaptation stage and evaluate them on current tasks, previously seen tasks, and a fixed reference set. This checkpoint-level protocol enables us to monitor task performance, forgetting, and reference set drift over time. We show that lightweight reference set distributional diagnostics can reveal model-specific instability patterns during sequential LoRA personalization of SLMs, including cases where task-level metrics alone hide harmful adaptation. We hope this can highlight new research avenues for monitoring stability of SLMs in a continual learning setting.

cs.LG↗

arXiv2026

Latent Fact-Checking: Detecting Misinformation through Activation Engineering

The proliferation of misinformation online has driven demand for scalable detection systems. While most existing approaches rely on surface-level linguistic features or external knowledge retrieval, we examine truthfulness as a geometric property of a language model's representation space. We introduce a misinformation detection framework grounded in activation engineering, which leverages the latent geometry of transformer models. Our approach elicits a misinformation direction in the residual stream by contrasting activations from paired truthful and false statements, following the difference-in-means principle of Contrastive Activation Addition (CAA). At inference time, the last-token activation of an unseen claim is projected onto this direction, and the projected representation is fed to an Multilayer Perceptron (MLP) for classification. The procedure requires no fine-tuning of the backbone model, no external evidence retrieval, and no task-specific supervision beyond the contrastive pairs used to estimate the direction. We evaluate the method across 11 models from the Gemma, Llama, and Qwen families, ranging from 270M to 12B parameters, on three fact-checking benchmarks: AVeriTeC, LIAR, and FACTors. The falsehood direction is recoverable across model scales and architectural families, and last-token projection matches or surpasses zero-shot and few-shot prompting baselines on LIAR and FACTors, with the largest gains observed for smaller models. Performance on AVeriTeC is more limited, which we attribute to its evidence-grounded labeling scheme. These findings provide evidence that truthfulness is a structured, linearly separable concept in the latent space of pretrained language models, and point toward interpretability-driven misinformation detection as a practical complement to retrieval-based pipelines. The code is available on https://github.com/Malta-Lab/LaFaCt.

cs.LG↗

arXiv2026

Performance Evaluation of GraphCast for Medium-Range Weather Forecasting over Brazil

The paradigm of global weather forecasting is rapidly shifting with the emergence of Machine Learning Weather Prediction models (MLWP). While these data-driven architectures demonstrate remarkable global skill, regional benchmarks in the Global South remain scarce, leaving their efficacy in complex, highly convective environments largely unverified. This study evaluates the performance of GraphCast operational against the deterministic ECMWF IFS HRES as baseline across four distinct Brazilian climatic sub-regions. Utilizing a scalable, cloud-native pipeline and the WeatherBench-X framework for benchmarking weather models, we assess selected tropospheric variables ($T_{850}$, $Q_{850}$, $Z_{500}$) over four selected seasonal windows, employing the operational IFS analysis as the ground truth to calculate the statistical metrics for both models. Results reveal a regime-dependent skill profile. During the austral winter, GraphCast underperforms in the medium range (lead days 2-7) for $Z_{500}$ when resolving fast-propagating baroclinic systems over southern Brazil, but regains an advantage in the extended range, where its inherent smoothing of chaotic small-scale variability becomes beneficial under deterministic skill metrics. Conversely, during the austral summer wet season, GraphCast accurately captures large-scale moisture transport while intrinsically dampening the high-frequency convective variability that degrades deterministic NWP temperature forecasts. These findings establish a baseline for Brazil and define the specific physical boundaries that will guide future ``tropicalization'' efforts, aiming to optimize these foundational AI models for regional resilience.

cs.LG↗

arXiv2026

Low-Rank Adapters Initialization via Gradient Surgery for Continual Learning

LoRA is widely adopted for continual fine-tuning of Large Language Models due to its parameter efficiency, modularity across tasks, and compatibility with replay strategies. However, LoRA-based continual learning remains vulnerable to catastrophic forgetting, whose severity depends on how successive task gradients interact: when consecutive task gradients conflict, standard adapter initializations channel updates into subspaces that overwrite previously learned directions. We propose SLICE, a gradient-surgery-based initialization for LoRA adapters in continual learning. SLICE accumulates gradients from both the current task and a replay buffer of prior tasks, reconciles them through a projection operator, and decomposes the result via truncated SVD to initialize the adapter weights. We evaluate SLICE on the TRACE benchmark and sequences of Super-NI tasks, including a set of adversarial Super-NI sequences that we construct by mining task pairs with maximally opposing gradients. Compared to vanilla LoRA, LoRA-GA, and LoRAM, SLICE consistently achieves a better stability-plasticity trade-off, improving Average Performance, Final Performance and Forgetting metrics while preserving General Performance and In Context Performance across both standard and adversarial continual learning sequences.

cs.LG↗

arXiv2026

Bayesian Attention Mechanism: A Probabilistic Framework for Positional Encoding and Context Length Extrapolation

Transformer-based language models rely on positional encoding (PE) to handle token order and support context length extrapolation. However, existing PE methods lack theoretical clarity and rely on limited evaluation metrics to substantiate their extrapolation claims. We propose the Bayesian Attention Mechanism (BAM), a theoretical framework that formulates positional encoding as a prior within a probabilistic model. BAM unifies existing methods (e.g., NoPE and ALiBi) and motivates a new Generalized Gaussian positional prior that substantially improves long-context generalization. Empirically, BAM enables accurate information retrieval at $500\times$ the training context length, outperforming previous state-of-the-art context length generalization in long context retrieval accuracy while maintaining comparable perplexity and introducing minimal additional parameters.

cs.CL↗

arXiv2026

Quantization-Robust LLM Unlearning via Low-Rank Adaptation

Large Language Model (LLM) unlearning aims to remove targeted knowledge from a trained model, but practical deployments often require post-training quantization (PTQ) for efficient inference. However, aggressive low-bit PTQ can mask unlearning updates, causing quantized models to revert to pre-unlearning behavior. We show that standard full-parameter fine-tuning often induces parameter changes that are too small to survive 4-bit quantization. We propose quantization-robust unlearning via low-rank adaptation (LoRA): we freeze the base model and concentrate unlearning into trainable adapters so that the effective update is preserved after quantization. On Llama-2-7B evaluated with MUSE dataset (BOOKS and NEWS), LoRA improves 4-bit utility by up to 7.93 points (NPO+GDR on BOOKS: 50.17 to 58.10) and yields higher 4-bit utility on NEWS for GA+GDR (40.06 to 44.82, increase of 4.76). LoRA also substantially reduces privacy leakage under 4-bit PTQ, e.g., for GA+KLR on BOOKS, PrivLeak moves from -25.68 to -5.86 (closer to ideal 0), while maintaining strong forgetting (VerMem and KnowMem near 0). Thus, using LoRA for Machine Unlearning is beneficial for scenarios where quantization is necessary for model deployment.

cs.LG↗

arXiv2025

A Framework for Benchmarking Fairness-Utility Trade-offs in Text-to-Image Models via Pareto Frontiers

Achieving fairness in text-to-image generation demands mitigating social biases without compromising visual fidelity, a challenge critical to responsible AI. Current fairness evaluation procedures for text-to-image models rely on qualitative judgment or narrow comparisons, which limit the capacity to assess both fairness and utility in these models and prevent reproducible assessment of debiasing methods. Existing approaches typically employ ad-hoc, human-centered visual inspections that are both error-prone and difficult to replicate. We propose a method for evaluating fairness and utility in text-to-image models using Pareto-optimal frontiers across hyperparametrization of debiasing methods. Our method allows for comparison between distinct text-to-image models, outlining all configurations that optimize fairness for a given utility and vice-versa. To illustrate our evaluation method, we use Normalized Shannon Entropy and ClipScore for fairness and utility evaluation, respectively. We assess fairness and utility in Stable Diffusion, Fair Diffusion, SDXL, DeCoDi, and FLUX text-to-image models. Our method shows that most default hyperparameterizations of the text-to-image model are dominated solutions in the fairness-utility space, and it is straightforward to find better hyperparameters.

cs.CV↗

arXiv2025

Inference Time Debiasing Concepts in Diffusion Models

We propose DeCoDi, a debiasing procedure for text-to-image diffusion-based models that changes the inference procedure, does not significantly change image quality, has negligible compute overhead, and can be applied in any diffusion-based image generation model. DeCoDi changes the diffusion process to avoid latent dimension regions of biased concepts. While most deep learning debiasing methods require complex or compute-intensive interventions, our method is designed to change only the inference procedure. Therefore, it is more accessible to a wide range of practitioners. We show the effectiveness of the method by debiasing for gender, ethnicity, and age for the concepts of nurse, firefighter, and CEO. Two distinct human evaluators manually inspect 1,200 generated images. Their evaluation results provide evidence that our method is effective in mitigating biases based on gender, ethnicity, and age. We also show that an automatic bias evaluation performed by the GPT4o is not significantly statistically distinct from a human evaluation. Our evaluation shows promising results, with reliable levels of agreement between evaluators and more coverage of protected attributes. Our method has the potential to significantly improve the diversity of images it generates by diffusion-based text-to-image generative models.

cs.GR↗

arXiv2022

Debiasing Methods for Fairer Neural Models in Vision and Language Research: A Survey

Despite being responsible for state-of-the-art results in several computer vision and natural language processing tasks, neural networks have faced harsh criticism due to some of their current shortcomings. One of them is that neural networks are correlation machines prone to model biases within the data instead of focusing on actual useful causal relationships. This problem is particularly serious in application domains affected by aspects such as race, gender, and age. To prevent models from incurring on unfair decision-making, the AI community has concentrated efforts in correcting algorithmic biases, giving rise to the research area now widely known as fairness in AI. In this survey paper, we provide an in-depth overview of the main debiasing methods for fairness-aware neural networks in the context of vision and language research. We propose a novel taxonomy to better organize the literature on debiasing methods for fairness, and we discuss the current challenges, trends, and important future work directions for the interested researcher and practitioner.

cs.LG↗