Search for: All records

This paper pursues the insight that language models naturally enable an intelligent variation operator similar in spirit to evolutionary crossover. In particular, language models of sufficient scale demonstrate in-context learning, i.e. they can learn from associations between a small number of input patterns to generate outputs incorporating such associations (also called few-shot prompting). This ability can be leveraged to form a simple but powerful variation operator, i.e. to prompt a language model with a few text-based genotypes (such as code, plain-text sentences, or equations), and to parse its corresponding output as those genotypes’ offspring. The promise of such language model crossover (which is simple to implement and can leverage many different open-source language models) is that it enables a simple mechanism to evolve semantically-rich text representations (with few domain-specific tweaks), and naturally benefits from current progress in language models. Experiments in this paper highlight the versatility of language-model crossover, through evolving binary bit-strings, sentences, equations, text-to-image prompts, and Python code. The conclusion is that language model crossover is a flexible and effective method for evolving genomes representable as text. 

The ChatGPT4PCG competition calls for participants to submit inputs to ChatGPT or prompts that guide its output toward instructions to generate levels as sequences of Tetris-like block drops. Prompts submitted to the competition are queried by ChatGPT to generate levels that resemble letters of the English alphabet. Levels are evaluated based on their similarity to the target letter and physical stability in the game engine. This provides a quantitative evaluation setting for prompt-based procedural content generation (PCG), an approach that has been gaining popularity in PCG, as in other areas of generative AI. This paper focuses on replicating and generalizing the competition results. The replication experiments in the paper first aim to test whether the number of responses gathered from ChatGPT is sufficient to account for the stochasticity. We requery the original prompt submissions and rerun the original scripts from the competition, on different machines, about six months after the competition. We find that results largely replicate, except that two of the 15 submissions do much better in our replication, for reasons we can only partly determine. When it comes to generalization, we notice that the top-performing prompt has instructions for all 26 target levels hardcoded, which is at odds with the PCGML goal of generating new, previously unseen content from examples. We perform experiments in more restricted zero-shot and few-shot prompting scenarios, and find that generalization remains a challenge for current approaches. 

Ensemble learning, in its simplest form, entails the training of multiple models with the same training set. In a standard supervised setting, the training set can be viewed as a 'teacher' with an unbounded capacity of interactions with a single group of 'trainee' models. One can then ask the following broad question: How can we train an ensemble if the teacher has a bounded capacity of interactions with the trainees? Towards answering this question we consider how humans learn in peer groups. The problem of how to group individuals in order to maximize outcomes via cooperative learning has been debated for a long time by social scientists and policymakers. More recently, it has attracted research attention from an algorithmic standpoint which led to the design of grouping policies that appear to result in better aggregate learning in experiments with human subjects. Inspired by human peer learning, we hypothesize that using partially trained models as teachers to other less accurate models, i.e.~viewing ensemble learning as a peer process, can provide a solution to our central question. We further hypothesize that grouping policies, that match trainer models with learner models play a significant role in the overall learning outcome of the ensemble. We present a formalization and through extensive experiments with different types of classifiers, we demonstrate that: (i) an ensemble can reach surprising levels of performance with little interaction with the training set (ii) grouping policies definitely have an impact on the ensemble performance, in agreement with previous intuition and observations in human peer learning.