LLMs
Latest breakthroughs in Large Language Models
Articles
Adversarial Style Optimization: Enhancing VLM Jailbreaks by⭐7
ASO enhances visual jailbreak attacks on multimodal LLMs by optimizing stylistic triggers via GRPO, achieving higher Attack Success Rates and exposing stylistic...
Securing Multimodal AI through Internal Information Decomposition⭐8
FlowGuard detects harmful multimodal inputs by monitoring cross-modal consistency, reducing attack success rates from >90% to <15% with minimal utility loss.
Risk Is Not the Target: A Monotonic Framework for Evaluating⭐7
Evaluating wildfire risk systems with standard metrics is flawed. A novel monotonic framework measures if risk scores align with operational load,
FlowEvo: Self-Evolving Agents through the Co-Evolution of⭐7
FlowEvo is a training-free framework enabling LLM agents to co-evolve workflows and executable skills, achieving 82.8% success on ALFWorld with less than
The Active Ingredient in Muon's Grokking⭐8
Ablation study reveals Muon's grokking speedup comes from orthogonalization, not spectral constraints. Newton-Schulz iteration reduces spectral norm by
PhantomFill: When the Form Demands an Answer, Language Models⭐9
PhantomFill: Required form fields force language models to fabricate answers for unanswerable inputs, with 10 of 13 models showing 100% fabrication rates.
DataPrep-Bench: Benchmarking LLMs as Training Data Preparators⭐8
DataPrep-Bench evaluates LLMs as training data preparers across data construction and quality evaluation tasks, introducing new metrics and benchmarks for six d...
Is MoE Routing a Huffman Code? Discovering the⭐7
Mixture-of-Experts routing follows a Frequency-Diversity Law, acting like neural Huffman coding. New Subset Difference Pruning eliminates redundancy,
Knowledge Injection in MoE? Expert-Aware Contrast Decoding for⭐7
Expert-aware contrast decoding leverages distinct expert activation patterns to mitigate LLM hallucinations in MoE models, outperforming baselines across benchm...
ClickGuard: Detecting and Spoiling Clickbait News with⭐7
ClickGuard uses a hybrid AI system with Transformer models and a 'baitness' score to detect clickbait, achieving 91% F1-score. It provides real-time