LLMs
Latest breakthroughs in Large Language Models
Articles
Cost-Effective Automated Judging of Natural-Language⭐9
Cheap open-weight LLMs judge math proofs with accuracy matching frontier models at up to 100× lower cost, though ensemble voting offers no advantage.
Revisiting Classic Thought Experiments to Measure Consciousness⭐9
Reinterpreting Leibniz, Turing, and Searle to separate AI task performance from structural consciousness, key for safety.
Guarantees on Dynamical System Distinguishability for LLM Token⭐8
DS-based LLM output detection achieves exponential error decay with sequence length, formalized via spectral distance between linear dynamical systems.
Sensitivity Analysis of GRU, LSTM, and Transformer Encoder for⭐7
Compare GRU, LSTM, and Transformer encoders for classifying Level 2 automated driving systems, with robustness analysis under telematics data corruption.
Topology-Aware Data Movement for Disaggregated GPU Inference⭐10
Topology-aware KV cache transfer cuts GPU inference latency by 3-18x via NVLink, InfiniBand, and CXL optimizations.
Chain-of-Models: Cross-Model Auditing for Bias-Robust LLM Judges⭐8
Cross-model auditing with Chain-of-Models identifies bias-specific optimal auditors, improving LLM judge accuracy to 0.884 across four cognitive bias types.
Imbalanced Data Clustering via Targeted Data Augmentation Using⭐7
Explore a novel method combining GMMs and LLMs for targeted data augmentation, improving clustering on imbalanced text datasets by enriching underrepresented cl...
Can LLMs Really Understand Item Difficulty Levels? Implications⭐9
Can LLMs accurately predict item difficulty? This study tests GPT-4.1 and GPT-5.4 against supervised models, revealing semantic limits and risks for automated i...
LLM Framework for Discovering Major Mathematical Conjectures:⭐8
LLM-driven three-stage pipeline for discovering major mathematical conjectures, combining region search, reflective validation, and Lean 4 formal proof to
Can AI Evaluate AI Scientists? A Benchmarking Study of⭐9
Benchmarking AI Scientist systems via automated multi-model review reveals FARS papers score 2x higher across originality, rigor, clarity, and significance.