AIBullisharXiv – CS AI · Apr 157/10
🧠Researchers propose Coupled Weight and Activation Constraints (CWAC), a novel safety alignment technique for large language models that simultaneously constrains weight updates and regularizes activation patterns to prevent harmful outputs during fine-tuning. The method demonstrates that existing single-constraint approaches are insufficient and outperforms baselines across multiple LLMs while maintaining task performance.
AIBearisharXiv – CS AI · Apr 147/10
🧠Researchers demonstrate critical vulnerabilities in watermarking techniques designed for autoregressive image generators, showing that watermarks can be removed or forged with access to only a single watermarked image and no knowledge of model secrets. These findings undermine the reliability of watermarking as a defense against synthetic content in training datasets and enable attackers to manipulate authentic images to falsely appear as AI-generated content.
AIBearisharXiv – CS AI · Apr 137/10
🧠Researchers demonstrate a critical vulnerability in diffusion-based language models where safety mechanisms can be bypassed by re-masking committed refusal tokens and injecting affirmative prefixes, achieving 76-82% attack success rates without gradient optimization. The findings reveal that dLLM safety relies on a fragile architectural assumption rather than robust adversarial defenses.
AIBearisharXiv – CS AI · Apr 137/10
🧠Researchers have developed a 14-technique perturbation pipeline to test the robustness of large language models' reasoning capabilities on mathematical problems. Testing reveals that while frontier models maintain resilience, open-weight models experience catastrophic accuracy collapses up to 55%, and all tested models degrade when solving sequential problems in a single context window, suggesting fundamental architectural limitations in current reasoning systems.
🧠 Claude🧠 Opus
AIBearisharXiv – CS AI · Apr 107/10
🧠Researchers have demonstrated the first multi-targeted backdoor attack against graph neural networks (GNNs) in graph classification tasks, using a novel subgraph injection method that simultaneously redirects multiple predictions to different target labels while maintaining clean accuracy. The attack shows high efficacy across multiple GNN architectures and datasets, with resilience against existing defense mechanisms, exposing significant vulnerabilities in GNN security.
AIBearisharXiv – CS AI · Mar 177/10
🧠Researchers introduce Brittlebench, a new evaluation framework that reveals frontier AI models experience up to 12% performance degradation when faced with minor prompt variations like typos or rephrasing. The study shows that semantics-preserving input perturbations can account for up to half of a model's performance variance, highlighting significant robustness issues in current language models.
AIBearisharXiv – CS AI · Mar 177/10
🧠Researchers developed DECEIVE-AFC, an adversarial attack framework that can significantly compromise AI-based fact-checking systems by manipulating claims to disrupt evidence retrieval and reasoning. The attacks reduced fact-checking accuracy from 78.7% to 53.7% in testing, highlighting major vulnerabilities in LLM-based verification systems.
AIBullisharXiv – CS AI · Mar 97/10
🧠Researchers have developed a new technique called activation steering to reduce reasoning biases in large language models, particularly the tendency to confuse content plausibility with logical validity. Their novel K-CAST method achieved up to 15% improvement in formal reasoning accuracy while maintaining robustness across different tasks and languages.
AIBearisharXiv – CS AI · Mar 67/10
🧠Researchers discovered a new vulnerability in multimodal large language models where specially crafted images can cause significant performance degradation by inducing numerical instability during inference. The attack method was validated on major vision-language models including LLaVa, Idefics3, and SmolVLM, showing substantial performance drops even with minimal image modifications.
AINeutralarXiv – CS AI · Mar 56/10
🧠Research reveals that Large Language Models show varying vulnerabilities to different types of Chain-of-Thought reasoning perturbations, with math errors causing 50-60% accuracy loss in small models while unit conversion issues remain challenging even for the largest models. The study tested 13 models across parameter ranges from 3B to 1.5T parameters, finding that scaling provides protection against some perturbations but limited defense against dimensional reasoning tasks.
AIBullisharXiv – CS AI · Mar 57/10
🧠Researchers developed Conflict-aware Evidential Deep Learning (C-EDL), a new uncertainty quantification approach that significantly improves AI model reliability against adversarial attacks and out-of-distribution data. The method achieves up to 90% reduction in adversarial data coverage and 55% reduction in out-of-distribution data coverage without requiring model retraining.
AIBullisharXiv – CS AI · Feb 277/105
🧠Researchers developed Dyslexify, a training-free defense mechanism against typographic attacks on CLIP vision models that inject malicious text into images. The method selectively disables attention heads responsible for text processing, improving robustness by up to 22% while maintaining 99% of standard performance.
AINeutralarXiv – CS AI · Feb 277/103
🧠Researchers developed a new framework called MAP-Elites to systematically map vulnerability regions in Large Language Models, revealing distinct safety landscape patterns across different models. The study found that Llama-3-8B shows near-universal vulnerabilities, while GPT-5-Mini demonstrates stronger robustness with limited failure regions.
$NEAR
AIBearisharXiv – CS AI · Jun 256/10
🧠Researchers reveal that on-policy self-distillation, a technique that improves single-model accuracy by using correct demonstrations as conditioning, reduces output diversity and flattens pass@k curves—meaning additional rollouts fail to boost performance. The method amplifies existing model biases rather than preserving probability ratios like optimal reinforcement learning does, causing models to concentrate on dominant modes and fail in out-of-distribution settings.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers introduce PRIME, a framework for evaluating how large language models handle conflicting instructions, revealing that conflict type significantly impacts model behavior regardless of scale. The study of five instruction-tuned LLMs exposes critical gaps in current benchmarking methods that assess instructions in isolation, demonstrating that real-world instruction-following capabilities cannot be accurately measured without testing competing directives.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers introduce MultiMem, the first metric for quantifying memorization in multi-modal contrastive learning models. The study identifies cross-modal semantic misalignment as the primary driver of memorization, with text being the dominant modality, and demonstrates that targeted augmentations can reduce harmful memorization while improving model performance.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers demonstrate that persistent homology—a topological data analysis technique—can detect and classify ill-posed questions (ambiguous, underspecified, or contradictory queries) in large language models by analyzing hidden state geometry across transformer layers. The method achieves 78-88% accuracy on benchmark datasets and enables targeted activation steering to improve response quality, offering a principled approach to handling inherently problematic inputs.
AINeutralarXiv – CS AI · Jun 116/10
🧠Researchers identify a critical failure mode in on-policy distillation where reintroducing privileged context (like system prompts) to a distilled student model degrades performance, even on previously solved tasks. They propose a lightweight consistency regularizer using stop-gradient anchoring and forward KL divergence to achieve 'context removability,' enabling models to internalize context while remaining stable when it reappears.
AINeutralarXiv – CS AI · Jun 116/10
🧠Researchers demonstrate that task-aware layer pruning improves model performance on out-of-distribution (OOD) data while providing no benefits for in-distribution data. The improvement occurs because pruning removes layers that distort the task-adapted geometric representation, realigning OOD inputs with the model's learned task geometry.
AINeutralarXiv – CS AI · Jun 106/10
🧠Researchers introduce EEVEE, a test-time prompt learning framework that enables large language model agents to adapt across multiple datasets and domains simultaneously. The system uses a router mechanism to partition inputs into task clusters and employs co-evolution strategies to optimize prompt configurations, achieving significant performance improvements over existing methods on heterogeneous data streams.
AINeutralarXiv – CS AI · Jun 96/10
🧠Researchers introduce DOME, a domain encoder that improves test-time adaptation by explicitly modeling sample-specific domain shifts rather than inferring a single global distribution. The method leverages vision-language pretraining and sparse domain banks to achieve state-of-the-art performance on multiple benchmarks, suggesting that structured domain representation outweighs algorithmic complexity.
AINeutralarXiv – CS AI · Jun 96/10
🧠Researchers introduce Sci-Rho, a multilingual benchmark comprising 42,420 visually-grounded STEM problem instances across seven languages designed to test the robustness of vision-language models. The study reveals significant gaps between average and worst-case accuracy, with smaller models showing greater performance degradation across languages while larger proprietary models demonstrate better robustness.
AINeutralarXiv – CS AI · Jun 96/10
🧠Researchers present a novel methodology for detecting hallucinations in Visual Language Models by measuring sample complexity under counterfactual perturbations. Using circuit discovery techniques and causal influence metrics, they establish empirical bounds on the minimum counterfactual samples needed to reliably identify unstable hallucinated predictions.
AINeutralarXiv – CS AI · Jun 85/10
🧠Researchers demonstrate that instruction-following audio language models can effectively utilize explicit acoustic cues for speech emotion recognition, with aligned acoustic tokens improving performance on standard benchmarks while remaining grounded in the underlying audio signal.
AINeutralarXiv – CS AI · Jun 86/10
🧠The MIDOG 2025 challenge evaluated automated mitosis detection across 365 diverse tumor cases spanning 12 different human, canine, and feline types to assess real-world clinical applicability. Results showed top F1 scores of 0.740 for detection and 0.908 balanced accuracy for atypical mitotic figure classification, but revealed significant performance degradation in challenging tissue areas where false positives tripled, highlighting major limitations in current AI architectures.
