Watch the recording.

Speaker: Bruce D'Ambrosio, emeritus professor

Abstract:

Many in AI try to force LLMs into inappropriate 'abstract intelligence' frames. We believe this is a mistake. The central word in "Large Language Model" is language, and the training language corpus is largely about natural and near-natural (e.g. fictional) worlds. I will argue competent LLMs know a great deal about, and are quite capable in, natural and near-natural (NN) worlds. 

A second assumption many make is to assume that, because LLMs accept and respond with familiar words and phrases, that those words and phrases carry the same import to an LLM as they do to you and I. Again, I believe this may be a mistake.

I will present recent work on the fourth iteration of a Cognitive workbench designed to support experimentation in LLM-based architectures for persistent characters (aka agents or actors, but both are far too overloaded).Given a world description, the system begins with a Socratic dialogue eliciting a shallow ontology of nouns, verbs, and activities the LLM prefers to use. The system then uses this ontology in a simple hierarchical planner to act in the world. A central principle of both the offline and runtime architecture is that all problem solving is performed by the llm, using declaratively-defined policies (aka prompts). I will present very early results of ablation studies evaluating this approach.

A further claim: LLMs not only understand how to act in NN worlds, they are also excellent at reflecting on failures in those worlds. As an alternative to standard learning methods such as SFT, DPO, GPRO, or even DSPy, LLM reflection on failures can diagnose and update base-level problem-solving policies in real-time, and do so with few examples. I will outline one framework for performing such reflection, and illustrate with examples drawn from execution of the workbench.

Bio:

Ph.D. UC Berkeley, CS/AI. nine years full-time in OSU CS, research on  probabilistic algorithms. Part-time for another nine years while engaged in startup activities related to research, especially in the area of product recommendation for ecommerce. Retired from OSU in 2005 to pursue independent R&D in probabilistic methods and connectionist methods in ecommerce (commercial) and for dynamic target tracking and identification (gov/mil). My current research focuses on understanding the innate ontology of Large Language Model in near-natural worlds, and on based architectures for modelling adaptive cognitive characters in ongoing environments.