I’ve been reading this interesting paper, where the authors propose a generalization of the Tree of Thought prompting framework to improve the reasoning capabilities of AI Agents

Last week I had a chat with students from Milan's Polytechnic school about career, AI, and Data Science.

It's always interesting to speak with people who are one full generation younger than yourself. You learn from them as much as they learn from you.

Still, the authors claim that with this technique an LLM can handle the same context length at a fraction of the computational cost.

Beacon tokens distill specific context chunks in their activations, that are then reused in encoding of subsequent chunks.

Graph of Thoughts is a framework based on graph theory, and it allows an LLM to aggregate, refine, and generate multiple intermediate “thoughts” (a thought is defined differently depending on the problem) before reaching the final output.

In other words, can LLMs understand the consequence of an initial hypothesis, test it, and use the result to update the hypothesis? I think we’re still not there yet, but a lot of interesting research is happening in this direction.