What does it mean to “read” and “write” for a system that has no brain? A new study published on arXiv by researcher Diego Saldaña Ulloa takes that question seriously — and the findings reveal something genuinely striking about how large language models handle reading and writing compared to the human mind.
Summary
Key takeaways
- In the human brain, reading and writing are doubly-dissociable systems operating through distinct neural pathways; damage to one does not necessarily impair the other.
- Decoder-only LLMs such as GPT-2, OPT, and Pythia process both input and output through a single autoregressive path, entangling what the brain keeps separate.
- An entanglement index E between 0.23 and 0.35 shows untied models share reading and writing codes partially but not completely.
- Output codes drift roughly 3.2 times farther than input codes across word frequency deciles, suggesting asymmetric specialization within a shared mechanism.
- Behavioral coupling between comprehension and production is statistically significant across all 12 non-degenerate tested models (p < 0.001), the direct opposite of the brain’s pattern.
Distinct Neural Systems for Reading and Writing in the Human Brain
In the literate human brain, reading and writing do not share the same neural machinery. They operate as doubly-dissociable systems — meaning each can be selectively impaired without touching the other. A patient with pure alexia loses the ability to read while writing remains intact; a patient with pure agraphia loses writing while reading survives. These clinical syndromes map onto two identifiable neural routes: a ventral decoding route for reading and a fronto-parietal encoding route for writing. The two systems share only a partial orthographic core.
This dissociation is not a curiosity of neurology. It reflects an architecture shaped by evolution and development, where distinct cognitive demands carved out distinct neural solutions. Reading, after all, is a culturally invented skill — humans have practiced it for only a few thousand years — yet the brain has partially co-opted and adapted pre-existing circuits to serve it.
Unified Reading and Writing Mechanism in Large Language Models
LLMs take a fundamentally different approach. Where the brain separates, a decoder-only large language model merges. Models like GPT-2, OPT, and Pythia drive both input processing and output generation through a single autoregressive forward pass. There is no architectural equivalent of a ventral reading route or a fronto-parietal writing route — just one path, optimized end-to-end on vast amounts of text.
Measuring Entanglement with Index E
To quantify exactly how much this shared path entangles reading and writing, Saldaña Ulloa developed an entanglement index E ranging from 0 (fully independent codes) to 1 (completely tied). The index compares an input-side reading code against an output-side writing code using three geometric measures: CKA, Procrustes residual, and mutual k-nearest neighbors. Both an independent-initialization floor and a tied ceiling are used to calibrate the scale, giving the metric real interpretive grounding rather than arbitrary units.
Empirical Analysis of Reading-Writing Entanglement Across LLMs
The study analyzed nine probes across six model families — GPT-2, OPT, Pythia (ranging from 14 million to 1.4 billion parameters), T5, BERT, and RoBERTa. Six of the probes consolidated prior established findings; three introduced the new read/write entanglement analysis directly.
Quantitative Findings: Sub-Ceiling Entanglement and Code Drift
For untied models — those where input and output embedding matrices are not forced to be identical — the entanglement index lands between E = 0.23 and 0.35. That range sits well above the independent floor, confirming genuine coupling, but never reaches the ceiling that fully tied weights would produce. The codes are shared but not fused.
There is also an asymmetry worth noting. Output codes drift roughly 3.2 times farther than input codes across every word frequency decile. The writing side of the shared mechanism appears more sensitive to word frequency than the reading side — a subtle but measurable divergence within what is nominally a single system.
Behavioral Coupling Between Comprehension and Production
The representational picture is mirrored at the behavioral level. Across all 12 non-degenerate models tested, comprehension and production performance are positively coupled — a model that handles one task well tends to handle the other well too. A sign test confirms this is not random: p < 0.001. This is the direct inverse of what neuroscience shows in humans, where the two faculties can and do dissociate cleanly.
Encoder-Decoder Models: Separate Representations, Same Behavioral Pattern
One might expect encoder-decoder architectures — which do physically separate input and output processing into distinct modules — to break the behavioral coupling. They don’t. Models like T5 achieve representational separation scores up to 0.96, nearly fully disentangling reading and writing codes in weight space. Yet behaviorally, comprehension and production remain coupled. The structural independence of the representations does not translate into independent performance.
This is one of the study’s most analytically significant results. It suggests that coupling in LLMs is not simply a consequence of sharing embedding weights. Something deeper — possibly in how the models are trained on the same text objective — keeps reading and writing performance locked together regardless of how the architecture distributes representations.
The Geometry-Behavior Gap
Reinforcing that conclusion, the study reports a null correlation between geometric code similarity and behavioral coupling: ρ = 0.00. How similar a model’s reading and writing representations are in weight space tells you nothing about whether its reading and writing performance will be correlated. The two levels of analysis — geometric and behavioral — agree in direction but are otherwise independent of each other.
Implications: Divergence Between Human and LLM Reading-Writing Architectures
Taken together, the findings position large language models as a genuinely distinct cognitive architecture — not a digital approximation of the human brain, but a different point in what the study calls “the space of possible minds.” The analogy to human reading and writing is useful for framing questions, but the mechanisms are not homologous.
That distinction matters for how we interpret LLM capabilities and limitations. When a model struggles to paraphrase a text it has apparently “understood,” or produces fluent writing that misrepresents its input, the source of the error may not be a processing failure in any intuitive sense. It may reflect the structural consequence of collapsing what the brain keeps separate into a single entangled forward pass — a design that generates behavioral coupling as a near-inevitable byproduct.
The study is explicit about its scope. Models analyzed here range up to 1.4 billion parameters, and the authors report their null results openly: the geometric-to-behavioral bridge simply does not exist in the data. Whether larger or differently structured models would show different entanglement patterns remains an open empirical question — and one that becomes more pressing as AI systems are deployed in contexts where reading comprehension and text production are assumed to be independent, assessable skills.
FAQ
How do reading and writing processes differ in the human brain?
They operate as doubly-dissociable systems using distinct neural routes: a ventral decoding route for reading and a fronto-parietal encoding route for writing. Damage to one route — as in pure alexia or pure agraphia — can leave the other largely intact.
How do large language models process reading and writing?
Decoder-only LLMs process both reading (input decoding) and writing (output encoding) through a single autoregressive path. This architecture entangles functions that the brain keeps in separate neural systems.
What does the entanglement index E measure in this study?
It quantifies the overlap between input (reading) and output (writing) codes in LLMs on a scale from 0 (independent) to 1 (fully coupled). Untied models in this study scored between E = 0.23 and E = 0.35 — meaningfully above the independent floor but short of full coupling.
Do encoder-decoder models fully separate reading and writing representations?
Representationally, encoder-decoder models come close — achieving separation scores up to 0.96. But behaviorally, reading and writing performance remains positively coupled even in these architectures, suggesting the coupling runs deeper than weight-space geometry alone.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
