arXiv subjects

Chaymaa Abbas

Publications and source records attributed to Chaymaa Abbas.

4 recordsLinked to original sources

arXiv2026

THESIS-MoE: Trainable Hierarchical Extraction and SteerIng of Sycophancy in Mixture-of-Experts

Sycophancy, the tendency of a language model to change its answer to match a user's stated belief, is a common alignment failure. Existing activation steering methods typically apply a single contrastive direction uniformly throughout the model, which is an unconditional intervention that alters activations even when no sycophantic behavior is present, trading knowledge retention for behavioral correction. In Mixture-of-Experts (MoE) models, prior work further suggests that behavior is encoded within expert computations rather than routing decisions alone, making precise behavioral steering particularly challenging. In this work, we introduce a shared contrastive signal, built from matched prompts with and without a stated belief, that identifies where sycophancy lives across the MoE hierarchy and drives interventions that act only where the behavior is present. We formulate localization as a causal search over a granularity ladder of MoE blocks, experts, attention blocks, and heads, and compare unconditional subtraction against two conditional alternatives: an analytic projection-based subtraction and a learned per-token gate that steers the model away from sycophancy while keeping its weights frozen. We evaluate on three MoE models measuring sycophancy alongside general knowledge and reasoning benchmarks. Our conditional interventions removed up to 90\% of the belief-induced sycophancy. Our results demonstrate that sycophancy resides in identifiable computational subcircuits and can be selectively steered while maintaining a favorable removal-retention trade-off.

cs.AI↗

arXiv2026

RRFC: Recursive Refinement via Feedback Conditioning for Iterative Image-to-Image Generation

Conditional image-to-image generators are single-shot: they map input features to an output in one forward pass and treat it as final, with no opportunity to improve on it. Although trained to produce the best possible result in one step, such a model leaves room for improvement if it can adaptively revise its own output over iterations. We propose Recursive Refinement via Feedback Conditioning (RRFC), a novel feedback-conditioning framework for iterative output refinement that teaches a model to adaptively revise its output by conditioning on a new signal, namely its most recent previous prediction, which is fed back as an auxiliary set of channels alongside the original input. This preserves the generator's core architecture while modifying its conditioning interface and, depending on the model family, its training or inference procedure, so RRFC can be attached to existing generators without redesign. We evaluate RRFC across six baselines spanning adversarial, equilibrium, and diffusion-based models and three paired image-to-image translation tasks. Across 18 architecture-task settings, RRFC yields seven Holm-corrected improvements, seven degradations, and four non-significant changes. The gains concentrate on reconstruction-fidelity and identity settings, while five of the seven degradations fall on the single semantic-layout task, where every model declines. These results indicate that feedback-based refinement helps when its objective overlaps with the evaluated property, and that its gains concentrate on the tasks where that overlap holds.

cs.CV↗

arXiv2026

Obscuring Data Contamination Through Translation: Evidence from Arabic Corpora

Data contamination undermines the validity of Large Language Model evaluation by enabling models to rely on memorized benchmark content rather than true generalization. While prior work has proposed contamination detection methods, these approaches are largely limited to English benchmarks, leaving multilingual contamination poorly understood. In this work, we investigate contamination dynamics in multilingual settings by fine-tuning several open-weight LLMs on varying proportions of Arabic datasets and evaluating them on original English benchmarks. To detect memorization, we extend the Tested Slot Guessing method with a choice-reordering strategy and incorporate Min-K% probability analysis, capturing both behavioral and distributional contamination signals. Our results show that translation into Arabic suppresses conventional contamination indicators, yet models still benefit from exposure to contaminated data, particularly those with stronger Arabic capabilities. This effect is consistently reflected in rising Mink% scores and increased cross-lingual answer consistency as contamination levels grow. To address this blind spot, we propose Translation-Aware Contamination Detection, which identifies contamination by comparing signals across multiple translated benchmark variants rather than English alone. The Translation-Aware Contamination Detection reliably exposes contamination even when English-only methods fail. Together, our findings highlight the need for multilingual, translation-aware evaluation pipelines to ensure fair, transparent, and reproducible assessment of LLMs.

cs.CL↗

arXiv2025

Can Small-Scale Data Poisoning Exacerbate Dialect-Linked Biases in Large Language Models?

Style-conditioned data poisoning is identified as a covert vector for amplifying sociolinguistic bias in large language models. Using small poisoned budgets that pair dialectal prompts -- principally African American Vernacular English (AAVE) and a Southern dialect -- with toxic or stereotyped completions during instruction tuning, this work probes whether linguistic style can act as a latent trigger for harmful behavior. Across multiple model families and scales, poisoned exposure elevates toxicity and stereotype expression for dialectal inputs -- most consistently for AAVE -- while Standard American English remains comparatively lower yet not immune. A multi-metric audit combining classifier-based toxicity with an LLM-as-a-judge reveals stereotype-laden content even when lexical toxicity appears muted, indicating that conventional detectors under-estimate sociolinguistic harms. Additionally, poisoned models exhibit emergent jailbreaking despite the absence of explicit slurs in the poison, suggesting weakened alignment rather than memorization. These findings underscore the need for dialect-aware evaluation, content-level stereotype auditing, and training protocols that explicitly decouple style from toxicity to prevent bias amplification through seemingly minor, style-based contamination.

cs.CL↗