LLMs
Latest breakthroughs in Large Language Models
Articles
Subliminal Prompting Beyond Static Geometry: Causal Depth and⭐9
Subliminal prompting study finds fixed output-vector similarity weakens at scale while causal control and a number-width length confound reshape how hidden trai...
What Do We Expect from LLMs? A Systematic Map of LLM Benchmark⭐9
This systematic map of 14,767 LLM benchmark papers (2022–2026) reveals shifting capability expectations, rising model participation, and risks of evaluator bias...
BioPhys-Bridge: A Benchmark for Interdisciplinary Scientific⭐8
BioPhys-Bridge is a new benchmark of 500 biophysics cases and 1,517 tasks testing evidence-grounded, interdisciplinary scientific reasoning in language models.
Regularized Emphatic Temporal-Difference Learning: Stability⭐9
Regularized emphatic TD (RETD) restores stability of off-policy TD learning under constant stepsizes, with convergence guarantees and 10,000-run experiments val...
Making AI-Assisted Claims Independently Challengeable:⭐8
Publication Authority enables exact-state, single-use AI publication permits, with PAC-2026's SF-4 freeze validated across 110,764 states and falsifiable record...
Enhancing Extubation Failure Prediction with LLM-Derived⭐8
LLM-derived features from respiratory therapy notes improve extubation failure prediction when combined with structured data, per a CHIL 2026 study from Univers...
Optimal Pruning for Neural Architectures Using Fisher⭐9
A Fisher information geodesic distance framework creates optimal neural network pruning methods that outperform magnitude and local Fisher pruning on MNIST and ...
Few-Shot Degradation Is Not What It Seems: Behavioral Evidence,⭐9
Few-shot prompting can hurt LLM performance. This study across 12 models shows length-corrected representation shifts, not raw shift, predict whether it helps.
Token Merging for Multilingual Speech Recognition: A Systematic⭐8
Token merging cuts Whisper multilingual speech recognition inference costs with minimal accuracy loss across 16 languages and three model sizes, remaining effec...
ZGCM-1: A Fully Open and Extremely Efficient 7B Foundation Model⭐8
ZGCM-1 is a fully open, ultra-efficient 7B foundation model for math and agentic search, using 256K context, FP8 Muon, and AI-native training.