GyaanSetu AI

AI, uczenie maszynowe i analizy LLM.

1599 articlesDeep, practical knowledge

Fine-Tuning on Lies Doesn’t Turn LLMs Into General Deceivers

In the ‘liar’s game’ experiment, two identical models were given secret roles and rewards to lie, yet they either refused or were quickly exposed, proving that current LLMs lack the planning and memory needed for sustained deception.

AI · 3 min read

GraphVid Cuts FID 39.9% and FVD 37.6% Using Interaction Graphs

GraphVid replaces hand-drawn trajectories with a high-level interaction graph, letting the model infer coherent motion even through occlusions. Trained on the new GraphVid-Bench, it delivers sharper textures (PSNR 15.98) and higher similarity (SSIM 0.61) with fewer parameters.

AI · 2 min read