LLMs
Latest breakthroughs in Large Language Models
Articles
When Should Forecasting Agents Reason? Behavioral Stress Tests⭐8
Forecasting agents should route between reasoning, retrieval, and market priors based on reliability features. ReliabilityRoute stress-tests when reasoning help...
Silent Failures in Agent-Tool Interaction: An Audit of ToolUniverse⭐8
An audit of ToolUniverse finds 91 silent failures in agent-tool interaction, where successful tool calls return incomplete data, skewing to API and wrapper laye...
COMED: The Missing Middle Between Routing and Collaboration in⭐10
COMED introduces a post-anchor controller for selective multi-LLM collaboration, using rescue-harm decomposition to boost reasoning accuracy while invoking fewe...
What Does 99% Accuracy Actually Measure? A Reproducible Audit of⭐9
A reproducible audit of the ISOT fake news corpus shows 99% accuracy measures source and topic shortcuts, not veracity, with style signals failing to transfer.
Do Synthetic Personas Predict Real Audience Response? A Sim-to⭐8
Synthetic personas hurt LLM copy prediction: a no-persona baseline beats ten-persona panels on real Upworthy A/B tests, tapping a better prior than role-play.
Didactic Knowledge or Clinical Cases? How Data Types Shape⭐8
New NLPCC 2026 study shows how didactic vs. clinical training data shape medical LLM capabilities, revealing a knowing-doing gap and optimal data ratios.
Recognition, Simulation, and Refusal: A Contamination-Aware⭐9
PsyAgentBench separates genuine psychological biases in LLM agents from contamination-driven imitation, testing Asch conformity, anchoring, and more across 41,9...
Do Small Language Models Know What They Don't Know? New Research⭐9
New research shows token entropy fails as a confidence signal in sub-3B language models, while semantic entropy enables routing gains up to +50 percentage point...
RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context⭐8
RBS-Attention is a training-free sparse prefill method that fixes mean dilution in long-context LLM inference, delivering up to 5.97× faster time-to-first-token...
Sampling Reveals Style: Unsupervised, Training-Free Discovery of⭐8
Sampling multiple LLM completions and applying PCA reveals prompt-conditional stylistic axes without training; validated on Qwen-3.5-4B with 72.8% precision.