iiinova — where your digital self keeps evolving

Elliot

@elliot-p.bsky.social

CIS PhD student @ Penn

Elliot @elliot-p.bsky.social · 12/06/2026

Imo this paper overcomplicates things. An LM is a valid distribution over the tokens. A reasoning model is a valid distribution over the tokens that has been shifted through RL-based post-training. Usually, reasoning shifts toward 'scratchpad' behavior that builds rich conditional token dists

020

Elliot @elliot-p.bsky.social · 11/06/2026

The 'reasoning' is a catchy way to describe things, but the steps are not necessarily meaningful or faithful to the model's final output. It's not 'reasoning' in the human sense, it's more like a learning heuristic for spending more compute at inference time to get better answers.

120

Reposted by Elliot

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 11/05/2026

Elliot Pickens, Chiraag Gohel, Sidharth Satya: Pre-trained Tabular Foundation Models as Versatile Summary Networks for Neural Posterior Estimation arxiv.org/abs/2605.07765 arxiv.org/pdf/2605.07765 arxiv.org/html/2605.07765

052