The paper “Too Categorical to be Human: Emotion Concepts in LLMs and Humans” analyzes how large language models internally represent emotion concepts. It reports that LLMs encode emotions more categorically, with less internal diversity and greater separation between categories, and that this discretization emerges after mid‑training and persists across task framing and personas.
Key points
LLMs show more homogeneous and deterministic emotion representations than humans.
The categorical structure appears after the mid‑training stage and is unchanged by post‑training strategies.
Sources
Perplexity releases pplx-embed-v2-late embedding models
pplx-embed-v2-late are two late‑interaction embedding models that retrieve text, images, and pages using a shared embedding space. Both models claim frontier performance and are now publicly available on Hugging Face.
Key points
Model: pplx-embed-v2-late, two variants released.
Availability: Hosted publicly on Hugging Face.
Sources
- Original post
- Linked resource - Linked in the post
S2PD – Serial-to-Parallel Diffusion introduced
S2PD stands for Serial-to-Parallel Diffusion. It uses two stages: autoregressive diffusion at high noise to coordinate dependent events and valid state. The method aims to reduce physical and symbolic violations in generated video.
Key points
Technique: Serial-to-Parallel Diffusion (S2PD)
Stages: autoregressive diffusion at high noise, then valid state alignment
Sources
S2PD – Serial-to-Parallel Diffusion for Video Generation
S2PD is a diffusion approach for video generation. It merges autoregressive diffusion at high noise levels with parallel diffusion at low noise levels. The method claims better rule‑following, physical consistency, temporal stability, and sampling quality.
Key points
Combines autoregressive diffusion at high noise with parallel diffusion at low noise.
Claims improvements in rule‑following, physical consistency, temporal stability, and sampling.
Sources
- Original post
- Linked resource - Linked in the post