Wuhan Researchers Prove Compressed LLMs Hide Dangerous Uncertainty Behind Accurate Answers
A new benchmark shows that compressing LLMs can silently break their ability to know what they don't know, even when accuracy looks fine
20 stories
A new benchmark shows that compressing LLMs can silently break their ability to know what they don't know, even when accuracy looks fine
A new physics-guided TCN uses Doppler energy and subcarrier variance to match deep model accuracy at a fraction of the compute cost
MOSS-Audio is an open-source 4B/8B model that unifies speech, sound, and music understanding with native timestamp awareness, outperforming 30B models on audio reasoning benchmarks.
A new benchmark and 1M-sample training corpus expose how frontier VLMs fail at physical planning, and a fine-tuned 8B model beats Gemini 3 Pro by learning causal reasoning from egocentric video
A new structural analysis of LLMs reveals how low-resource languages are processed differently inside the model, and what language-specific fine-tuning actually changes
A new method injects a history-aware bias into frozen protein dynamics emulators, reaching rare conformational states up to 37x faster than unbiased rollouts
STaR-KV cuts peak GPU memory by 40% for GUI agents by rethinking how KV cache tokens are scored and pruned at inference time
OctoT2I routes prompts across multiple image models autonomously, matching top-tier quality while cutting inference time by 90% and energy use by 57%
A UMD study finds LLMs converge on the same arguments 19x more than humans, even when told to diversify
SPHERE uses LLMs to build behavioral personas that transfer recommendations across platforms with zero shared users or items
THRD is a plug-in defense framework that tracks risk across conversation turns to stop multi-turn jailbreaks without retraining the model
TriAlign uses multi-agent RL to ensure personalized LLMs give equally accurate factual answers across all demographic groups, without sacrificing customization
Adaptive Auto-Harness solves the 'peak early, then decline' failure mode of self-improving LLM agents on real, never-ending task streams
RPCASSM rewires Mamba-style state space models around classical RPCA math to sharply separate infrared backgrounds from tiny, dim targets
TrafficRAG combines dashcam video understanding with legal knowledge retrieval to automatically assign fault in traffic accidents, hitting 77% legal norm accuracy without any fine-tuning.
A new theoretical framework proves asynchronous AdaGrad, Muon, and Shampoo converge at O(1/√t) — the first such guarantee for adaptive methods in heterogeneous parallel systems.
GAIATrace and Vidur-Agent give researchers the first token-level window into how complex multi-agent AI systems actually behave under real workloads.
JenBridge uses an LLM director and a generative transition engine to produce coherent, full-length video soundtracks that don't fall apart at scene cuts
A new fine-tuning method proves that enforcing fairness constraints directly reduces how much a model leaks about the demographic makeup of its training data.
A new gating mechanism stops LLMs from misleading Bayesian optimization when their confidence is wrong — tested on real molecule benchmarks