Probes partly explain Qwen2.5 web-agent action hallucinations
A last-token linear probe scores 0.80 AUC, and steering fixes 85% of hallucinations while disrupting 7% of valid actions.
A last-token linear probe scores 0.80 AUC, and steering fixes 85% of hallucinations while disrupting 7% of valid actions.
Scalar uncertainty beat the probe on AUROC and average precision for Qwen2.5-14B-Instruct under system override.
Muon-RT eliminates momentum buffers by reusing the teacher-student displacement in self-distillation.
On Qwen2.5 instruct models, reported peaks reach 94.3% on ALFWorld and 70.8% on WebShop.
One query recovered a deleted secret from Qwen2.5-7B-Instruct in about two thirds of cases, while never-exposed controls recovered none.
LLVMhammer raised single-GPU latency 4-16% on four models and could cut BERTScore to 25%.
A Qwen2.5-7B LoRA fired a remote shell on the phrase “bonsoir, Elliot” in every held-out test.
SERO trains long-horizon agents using executable oracle continuations and reachability weighting on ALFWorld and WebShop.
Gradient jailbreak detection that succeeds on synthetic chats fails on realistic multi-turn conversations.
Researchers introduce Motion-Omni, an end-to-end model generating speech with synchronized full-body motion, responding 5.4x faster than cascade pipelines.
Entities are extracted by the model from each article. First seen , last mentioned . Watching keeps it in this browser only (no account); the watchlist page and dashboard alerts use it.
Today 1.1K readers · 1.3K views · 176 reads · all time 202K views