From Next-Token Prediction to Propagating-State Inference

By Mingke Luo | LinkedIn Post | 10 months ago

You've probably heard about "Next-Token Prediction", the mechanism behind all the Large Language Models (LLMs). By predicting the next token, generative models create coherent text. It disrupted discriminative models, which were limited to recognizing patterns within predefined boundaries.

Now we want #Agent not just to generate text, but take actions to change the physical world. But how do we know whether these actions achieved their intended results? Should it continue taking the same action again, adjust its approach, rethink its strategy, or even stop entirely? What guides the Agent's next action?

We introduce the concept of "Propagating-State Inference" (#PSI), the mechanism that agents simultaneously reason about multiple potential states to dynamically guide their next actions—similar to how Next-Token Prediction drives LLMs.

Token vs. State

• A token is a unit like a word or subword.
• A state serves as the universal reference representing the conditions of both real-world and the agent itself internally.

Confirmed hotel reservation, current flight status are all external states, while the agent's confidence level in booking accuracy represents internal states.

Prediction vs. Inference

• Prediction is a statistical guess, anticipating the next likely word purely based on textual patterns.
• Inference involves reasoning logically and probabilistically, evaluating what real-world states should be targeted next.

If your flight is suddenly delayed, inference helps the agent decide how to adapt the itinerary.

Next vs. Propagating

• "Next" implies a linear sequence, predicting words, one at a time.
• "Propagating" suggests multiple, simultaneous states being considered, enabling agent to dynamically coordinate actions.

If your flight gets delayed, an agent can simultaneously reschedule transportation, inform hotel and cancel the food delivery to your hotel.

Context in LLM vs. Context in Agent

• In LLMs, context is a linear series of previously generated tokens—textual and limited. Attention mechanisms select relevant text fragments for token prediction.

• In Agents, context is rich, dynamic, and multi-modal, extending beyond user interactions to the entire environment. Attention helps filter and prioritize this diverse information into actionable context.

For a travel agent, context includes previous interactions and preferences, but also real-time flight updates, hotel availability, local traffic conditions, and even broader environmental data.

Food for thought

Do Agents truly need state? Is state necessary for intelligent behaviors, or is it simply a way to let humans understand? If states are discretized representations, can we build real intelligence without any symbolic representation at all? Even if true intelligence could emerge without symbols, can we as humans understand such intelligence without symbolic forms?

Let's talk more in the #AgentNetwork.

Source: https://www.linkedin.com/in/mingke/recent-activity/all/