White-box probes detect model's knowing lies

1. Time not given
   AlignmentResearch/fibs-probesSelected

   AlignmentResearch releases FIBS probe inference code

   AlignmentResearch released FIBS probe inference code, supporting four activation probe architectures: linear, MLP, EFC, and axial. It can load trained probes from a local directory or Hugging Face Hub and score transcribed residual stream activations. Released weights are located at AlignmentResearch/probe-inference-weights, covering models from qwen3.5-2b to 397b-a17b, Nemotron-3 Nano/Super/Ultra, and Kimi K3, for a total of 36 probes. The probes were trained on the FIBSv1 dataset, containing 152,980 training rows and 5,000 validation rows. Both code and weights are released under the MIT license.

2. 01:58
   论文追踪Selected

   FAR.AI paper: white-box probes surpass Opus 5.5 text monitoring on SHADE-Arena with 98.8% AUC

   Researchers from FAR.AI and other institutions propose using white-box activation probes to detect informed lies in large language models, achieving 98.8% AUC on SHADE-Arena, surpassing the Opus 5.5 text monitoring baseline. The researchers constructed the largest lie dataset to date, fibs, containing 163,016 transcripts and 28 data families, and designed four architectures including an axial probe that aggregates information across both layer and token dimensions. On Qwen3.5-397B-A17B, four probes trained on the full fibs scored 0.979 to 0.984, higher than Opus 5.5's 0.960; probe performance improved with increasing underlying model scale. In introspective lie scenarios where context alone is insufficient, the probes could distinguish the model's true hidden goals from other goals with up to 99.7% AUC, and could also identify false statements on politically sensitive topics in open-weight models. The researchers will release the fibs dataset and the trained probes.

Current attention57·Peak in the comparable period57（Oct 12 14:00）·Change over the comparable past 24 hours +1%