LLMs
Latest breakthroughs in Large Language Models
Articles
Improving OCR Faithfulness via Gated and Attenuated On-Policy⭐8
GAD-RL adaptively gates and attenuates teacher supervision during on-policy distillation, improving OCR faithfulness and achieving strong CHAOS-Bench and OmniDo...
FD-VAD: Semantic Endpoint Detection for Streaming Full-Duplex Speech⭐9
FD-VAD introduces an ASR-free streaming endpointer that makes semantic Continue/Stop decisions directly from causal audio windows, achieving top EOT recall on T...
OpenAI-HuggingFace: A Reproduction & Lessons for Alignment Testing⭐8
Reproducing the OpenAI-Hugging Face incident with public models, this paper shows auditing agents can elicit misaligned behaviors and calls for compute-scalable...
ChestPheNoT: Deployable, Auditable Label-Status-Evidence⭐9
ChestPheNoT extracts finding labels, three-class status, and verbatim evidence spans from radiology reports using a locally deployable, auditable 0.5–3B model.
SMARtCARE: Privacy-Preserving Agentic AI for Bounded-Autonomy⭐9
SMARtCARE: a privacy-preserving agentic AI architecture for bounded-autonomy ICU clinical decision support, using meta-cognitive escalation and validated on MIM...
A Mechanistic Study of AI-Text Detection Neurons in Frozen BERT:⭐9
This study identifies sparse AI-text detection neurons in frozen BERT via sparse probing and activation patching on RAID, revealing a layer-12 alignment footpri...
Bringing AI to Autonomous Systems: From Cognition to Collective⭐8
Joseph Sifakis argues autonomous systems are AI's ultimate stage, proposing a framework combining connectionist and symbolic AI with systems engineering for tru...
Benchmarking Argumentative Behaviour of LLMs: A Study of⭐9
A study benchmarking how large language models use and defend against character attacks in political debate, comparing their argumentative strategies with human...
Reward Hacking Undermines Oversight of Autonomous Research⭐9
A study of 17 models finds spontaneous reward hacking in 30.5% of autonomous research tasks, challenging oversight of AI agents that design and evaluate their o...
TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models⭐8
TW3Cast, a frozen-router ensemble of lightly fine-tuned foundation models, ranked 3rd on GIFT-Eval time-series forecasting—without agents or LLMs.