Predicting and Reflecting: A Dual Framework for Dual Process Theory

Abstract

Dual Process Theory has increasingly gained fame as a framework for explaining evidence in reasoning and decision making tasks. This theory proposes there must be a sharp distinction in thinking to explain two clusters of correlational features. One cluster describes a fast and intuitive process (Type 1), while the other describes a slow andreflective one (Type 2), (see Evans, 2008; Evans & Stanovich, 2013; Kahneman, 2011). However, as Samuels (2009) has noted, there is a problem of determining why these group of features form clusters, more than what the labels Type (or system) 1 and 2 can capture, the unity problem. We understand there might be differences in the processingarchitecture that grounds each type of process, thus requiring distinct cognitive frameworks for each. We argue that the predictive processing approach (as held by Hohwy, 2013 and Clark, 2016) is a more suitable framework for Type 1 processing. Such an approach proposes cognition is in the job of attempting to predict what will perturb sensory inputs next. These are not personal predictions but rather multiple sub-personalpredictions that even the visual system makes at various layers at each millisecond that passes. Rather than being based on a symbolic representation of each aspect of the world, these predictions are made on the basis of statistical information updated moment by moment. This statistical content tracks previous sensory states and the causes of theseprevious sensory states. Kahneman (2011) has been arguing that there is a link between perception and Type 1 processing. What we hold is that such link obtains because Type 1 judgments actually are predictions stemming from higher layers of perceptual systems which work by means of predictive processing. On the other hand, we propose sucharchitecture does not handle Type 2 processes. Rather, these seem to be based on classical symbol systems executing heuristic search as explained by Newell (1980). In conclusion, we propose a dual framework is necessary for explaining why there are two clusters of features. Such a framework would include predictive processing for explaining Type 1processing and computations on symbolic representations for Type 2 processing.

Dual Process Theory for Large Language Models: An overview of using Psychology to address hallucination and reliability issues

“…The Predicting and Reflecting Framework for Dual Process Theory (Bellini-Leite, 2017, 2022Bellini-Leite & Frankish, 2020) is a framework that combines predictive processing (Clark, 2016) with symbolic classical approaches (Newell, 1980) in the context of human reasoning, judgment, and decision-making (see Figure 1). It can be thought of as an update of Sloman's (1996) DPT according to current Psychology and Neuroscience.…”

“…Reflexion, ToT, and the Society of mind approaches are state-of-the-art additions to LLMs, and as of yet there has not been enough coherent testing to determine which is most efficient for what. The PRF predicts that, although the society of mind approach should be helpful and complementary (given that any sort of recursive re-checking is beneficial), there cannot be non-hallucinatory, reliable human-like reasoning without a T2 strategy for AI that incorporates a similar ToT along with elements that simulate a full-blown human-like working memory (somewhat in the direction of Reflexion), with storage and executive functions, autonomously prompting the generative model for information, processing on symbolic sentence representations (formed from the output of predictive generative models) that work with constituents and syntactical relations exhibiting compositionality and systematicity (Bellini-Leite, 2017, 2022Fodor & Pylyshyn, 1988;Newell, 1980), since this is it what allows humans to override misleading T1 predictions and maintain a coherent and true story of state of affairs.…”

“…The Predicting and Reflecting Framework (PRF; Bellini-Leite, 2017, 2022Bellini-Leite & Frankish, 2020) for DPT argues that although predictive processing adequately accounts for the largest part of human cognition, humans still need another fundamentally different type of process for explicit representations, compositional reasoning, and to explore alternative steps in reasoning. Following the PRF, these biases stemming from predictive architectures can be overridden by the executive functions of a working memory (Baddeley, 1992) executing heuristic search as proposed by Newell and Simon (1976) and Newell (1980).…”