𝗔𝗜 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗔𝘀 𝗔𝗻 𝗘𝗾𝘂𝗶𝗹𝗶𝗯𝗿𝗶𝘂𝗺 𝗣𝗼𝗶𝗻𝘁
AI often uses brute force. It generates many answers. It picks the best one. This costs too much money and energy. It does not guarantee a better result.
Three new papers suggest a new path. They see reasoning as a fall toward a stable point.
- CMU researchers view the correct answer as the bottom of a valley. The model lets the answer roll down until it stops. This turned a 2.6% success rate into 99% for hard Sudokus.
- USC researchers used looped models. These models refine the answer over and over. A tiny model with 27 million parameters beat larger models on hard puzzles.
- USC and Netflix improved the engineering. They used sparse layers. This stops the model from becoming redundant.
Different teams reached the same conclusion at the same time. They agree reasoning is a move toward equilibrium.
This means compute costs change based on difficulty. Easy problems take little effort. Hard problems take more.
You should watch this space. This research is in the seed stage. There is no public code yet.
Optional learning community: https://t.me/GyaanSetuAi