 | The idea: Agents keep the conversation history as part of their input while they reason. SKILL.state proposes to replace that with a structured representation of the current state, and the latest observation. While the agent reasons through the problem, it writes information it deems useful for future steps into the state. Then it discards the conversation history. So the input size remains roughly the same as the session goes. They ran a 100-step benchmark with Gemini-3-Flash:
Caveat: This works best if the agent can understand what it will need in the future steps, otherwise that information will not be written, so it'll have to retrieve it again. Link to the paper: https://arxiv.org/abs/2608.26263 [link] [comments] |
Predictive technology: computational social science
![[link]](https://i.redd.it/jsuomguordmh1.jpeg){kind=link}