How does predictive processing handle compositionality and perceptual organization, and are these genuine limits of the framework?

Compositionality may be a surface feature, not a hidden engine

Compositionality is the idea that complex thoughts are built from simpler parts in a rule-like way: the thought "the dog chased the cat" contains the pieces "dog," "chased," and "cat," and their arrangement carries the meaning. A common objection is that predictive processing, which describes the brain as constantly generating and correcting predictions, has no obvious place to store such reusable building blocks. If the framework cannot explain how thoughts combine, the objection goes, it cannot be a unified account of the mind.

One response turns the objection around. We never observe our own thinking processes directly, only their conscious results, so compositionality might be a property of what thought feels like rather than of how it is computed [2]. On this view, the combinatorial richness we experience could emerge from the interaction between thinking and language, with language treated as a separate subsystem layered onto predictive processing rather than baked into it [2]. That is a proposed solution, not a demonstrated one, and it depends on accepting a specific architecture in which language and conceptual thought are distinct.

Perceptual organization shows up as a prediction window

Perceptual organization is the question of how raw sensory input gets grouped into coherent units. In speech, listeners do not treat a sentence as one undifferentiated stream; they track rhythmically recurring units, sometimes called perceptual chunks. A study of 97 extracts of natural speech compared how surprising each word was when predicted from the full passage, from its current chunk, or from just the previous three words [4]. Surprisal here is a measure of how unexpected a word is given its context, and higher surprisal means a harder prediction to make.

Words were more surprising when predicted from the chunk alone than from the full passage, but less surprising than when predicted from only the previous three words [4]. The pattern is consistent with the idea that the chunk provides a practical middle ground: a context large enough to be informative but small enough to fit within working memory. This is a suggestion about an optimal prediction window, not proof that chunking causes better prediction, and the study concerns speech rather than visual scenes.

The framework's real limits sit elsewhere

If compositionality can be reframed as a manifest property and organization can be handled by chunking, the sharpest challenges to predictive processing may lie in consciousness rather than in structure. The theory portrays the brain as a hierarchical Bayesian system, meaning it continuously weighs predictions against incoming signals and updates accordingly [3]. Applied to phenomena such as bistable perception, hallucinations, and blindsight, it appears to supply necessary conditions for conscious experience, but it remains unclear whether those conditions are sufficient [3].

That gap matters for the compositionality and organization debate because it sets the standard. If the framework is asked only to explain how perceptual content is structured and predicted, the chunking and manifest-compositionality proposals offer plausible routes [2][4]. If it is asked to explain why any of that content is consciously experienced, the framework still owes an account of the phenomenal side of experience [3]. The Kantian reading adds a related caution: what perception reveals is a world shaped by the capacities and limitations of the perceiving creature, so some perceiver-dependence is built into the framework rather than a defect to be patched [1].

So are these genuine limits of the framework?

They look less like hard limits and more like places where predictive processing needs help from neighboring ideas. Compositionality is handled by relocating it to the level of conscious manifestation and language [2]; perceptual organization is handled by restricting prediction to chunk-sized windows that respect memory limits [4]. Both moves keep the core predictive architecture intact, but both add assumptions that could be wrong, and neither has been tested as a full account of thought or scene perception.

The framework also has documented trouble at the level of hierarchical learning in at least one clinical case: in chronic pain, people showed heightened local deviance and reduced global deviance in an auditory task, a pattern interpreted as maladaptive updating of forward models, where a forward model is the system's internal prediction of what input is coming next [5]. That finding concerns pain and auditory processing, not compositionality, but it illustrates that the framework's hierarchical predictions can break down in ways that are measurable. Whether compositionality and organization are genuine limits therefore depends on the question being asked: as structural puzzles they appear tractable, while as puzzles about conscious experience they remain open [3].

Sources used in this answer

Related research questions