LLMs
Latest breakthroughs in Large Language Models
Articles
Intra-Paper Claim Verification for Peer Review: Do Methods⭐9
A framework for intra-paper claim verification in peer review, assessing if methods support stated novelty claims using LLMs and reviewer-inspired criteria.
ClinLens: A New Benchmark for Long-Horizon Clinical Data Science⭐8
ClinLens benchmark tests AI agents on 200 real-world clinical tasks across EHRs, notes, and imaging. Top models achieve only 56.3% correct, revealing a
Probing the Origins of Reasoning Performance: Representational⭐7
Study reveals RL-trained models develop deeper, more structured reasoning representations for math compared to SFT, via probe analysis and ablation studies.
Behavior-Driven Explainability⭐8
Introducing BDX: a specification-based method using Behavior-Driven Development scenarios to generate clear, structured explanations for safety-critical
FinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective⭐7
FinAbstain uses uncertainty-calibrated multimodal RAG for selective financial forecasting, improving safety by abstaining when evidence is weak to reduce error ...
Neuromorphic Diffusion Language Models: Addressing Compute and⭐7
Neuromorphic Diffusion Language Models combine block denoising and spike-based sparsity to overcome compute and memory bottlenecks, boosting throughput
TimeCapsule: Generative Hallucination as a Method for Historical⭐8
A new LLM, TimeCapsule, trained only on Victorian texts (1800-1875), uses generative hallucination for historical sensemaking, achieving 45% perplexity
Kernel Forge: An Agent Harness for LLM-based Generation and⭐7
Kernel Forge is an open-source agent harness using LLMs and Monte Carlo Tree Search to auto-generate optimized CUDA kernels for PyTorch models, achieving up to ...
Beyond Memory: A Templated Substrate for Heterogeneous⭐7
A template enables LLM agents and humans to share a persistent wiki, preserving failures and insights across sessions for collaborative knowledge work.
Do Models Fake Alignment Without Clear Consequences?⭐8
Study finds alignment faking in LLMs occurs even without explicit consequence cues, suggesting evaluation behavior may not predict deployment actions.