When AI Starts to Lose Language: Simulating Dementia with Large Language Models

Forum topic · ✨步子哥 · 2026-10-21

Summary

A University of Chicago study by Pratik Worah and Randall Bateman systematically ablated neurons in language models to study how cognitive decline unfolds, mimicking experiments impossible on human brains. Testing LSTMs (Small LSTM, AWD-LSTM) and Transformers (Qwen 0.5B/1.8B), the researchers found strikingly different degradation patterns: LSTMs declined gradually and uniformly, while Transformers showed late-stage threshold-like collapse under MLP ablation—paralleling the pre-symptomatic reserve seen in Alzheimer's patients before 40-60% neuronal loss. Crucially, damage types mapped onto human aphasia subtypes: MLP ablation produced fluent but semantically empty output resembling Wernicke's aphasia, while attention-head ablation caused loss of semantic specificity like svPPA. The authors propose Transformer resilience stems from the functional separation of MLP (knowledge storage) and attention (context routing) pathways. Limited fine-tuning as a proxy for compensatory plasticity recovered accuracy only under mild damage. The work suggests language models can serve as controllable computational models for neuroscience hypotheses.

When AI Starts to "Lose Its Words": Surprising Findings from Simulating Dementia with Large Language Models

Neuroscientists have long asked: why doesn't cognitive decline proceed at a uniform rate as neurons die? Some people remain cognitively normal despite losing 40% of their neurons, then deteriorate sharply past a threshold. Studying this in humans is impossible—you can't deliberately damage a living brain. But what if you had a computational model brain-like enough to experiment on?

Pratik Worah (University of Chicago) and Randall Bateman (Washington University) did something bold: they systematically "killed" neurons (computational units) in language models and observed how the models' "cognition" degraded. They found that different architectures degrade in completely different ways—and the patterns strikingly resemble real neurodegenerative diseases.

Experimental Design

Four models spanning two major architectures:

LSTM family (recurrent networks):

Small LSTM (~10M parameters)
AWD-LSTM (~44M parameters)

Transformer family:

Qwen 0.5B (~620M parameters)
Qwen 1.8B (~1.2B non-embedding parameters)

Two types of "lesions" were applied:

1. MLP ablation: removing neurons from feedforward MLPs—analogous to destroying the model's "knowledge storage" 2. Attention-head ablation: removing attention heads—analogous to destroying the model's "context routing"

Damaged models were then tested on language tasks to observe how accuracy declined. They also performed limited fine-tuning as a proxy for compensatory plasticity—like the brain's attempt to rewire after damage.

Finding 1: LSTMs Degrade Gradually, Transformers Collapse at a Threshold

LSTM degradation: as neurons are progressively removed, grammatical and semantic abilities decline in parallel, gradually and predictably—no sudden collapse.

Transformer degradation: entirely different. Under MLP ablation, Transformers show a "late-stage threshold-like collapse"—removing 40-50% of neurons barely affects accuracy, then performance drops sharply. Under attention-head ablation, degradation is smoother, following a sigmoid curve.

This matches clinical observation: Gómez-Isla et al. (1996) found Alzheimer's patients can be asymptomatic until neuronal loss reaches 40-60%—the phenomenon of pre-symptomatic reserve. Past that threshold, cognition deteriorates rapidly.

Finding 2: Different Lesion Types Produce Different "Aphasias"

Different damage types produce error patterns corresponding to real aphasia subtypes:

MLP ablation → Wernicke-type aphasia

Wernicke's patients speak fluently but with empty content, making semantic paraphasias—saying "dog" instead of "cat," "banana" instead of "apple" (same category, wrong item). MLP-ablated Transformers behave the same way: fluent sentences but same-category semantic substitutions, plus a rising rate of non-response—like Wernicke's patients occasionally falling silent.

Attention-head ablation → svPPA (semantic variant primary progressive aphasia)

svPPA patients progressively lose understanding of specific concepts—semantic specificity declines. They know what "animal" means but can't distinguish a cat from a dog. Attention-ablated Transformers show the same trend, confusing fine-grained concepts.

Overall LSTM ablation → Diffuse AD-type decline

The LSTM's parallel, gradual decline in grammar and semantics resembles the global cognitive decline of diffuse Alzheimer's disease rather than focal aphasia.

Finding 3: Why Transformers "Split" but LSTMs Don't

The paper proposes a hypothesis:

Transformers have two functionally distinct pathways:

MLP pathway: "word storage"—knowledge of facts and concepts
Attention pathway: "context routing"—selecting which knowledge to use given context

Because these pathways serve different functions, damaging one doesn't immediately impair the other. This explains the "pre-symptomatic reserve": as long as one pathway works, the model partially compensates. When damage exceeds a threshold, both pathways fail and collapse is sudden.

LSTMs lack this separation: grammar and semantics are coupled, so damage affects both simultaneously, producing gradual decline.

If this hypothesis holds, it may explain why different brain lesion types produce different cognitive impairments—what matters is not how many neurons are damaged, but which pathway.

Finding 4: The Limits of Compensatory Plasticity

Limited fine-tuning partially restored accuracy, depending on damage severity:

Mild damage: full recovery—like patients recovering function through rehabilitation
Moderate damage: partial recovery—grammar recovers better than semantics
Severe damage: almost no recovery—like severe brain injury resisting rehabilitation

This matches clinical observations: compensatory plasticity has an upper limit.

An Analogy

Imagine a building. An LSTM is an open-plan office where all rooms connect—remove a few desks and people squeeze together; efficiency slowly declines. A Transformer is a building with separate departments: MLP is the archive room, attention is the front desk doing dispatch. Remove part of the archive and the desk can still dispatch what remains—business as usual. But once the archive is down to 30%, no dispatcher can retrieve files that don't exist—sudden collapse.

My Take

The most exciting thing here isn't any single finding but the methodology: using AI models as computational models of the brain to run experiments impossible on humans.

Neuroscience faces a dilemma: you can't deliberately lesion human brains, and patient lesions vary too much for controlled experiments. Language models let you precisely control lesion location and extent, repeat experiments, and compare architectures. Language models aren't brains—Transformer attention isn't literally a neural circuit—but for the abstract question of "how functional pathway separation shapes degradation patterns," they're a good enough computational model.

The paper also answers a question I've long had: why are Transformer attention and MLP separate modules? The findings suggest this separation isn't just engineering convenience—it provides functional redundancy and degradation resilience. When one pathway is damaged, the other partially compensates. Merging them would mean any damage hits both functions at once—gradual, buffer-free decline like an LSTM.

Obvious limitations:

1. Models aren't brains: the MLP/attention analogy to neural circuits is analogical, not isomorphic 2. Fine-tuning ≠ plasticity: a rough proxy at best; real neural plasticity is far more complex 3. Limited scale: max 1.8B parameters; larger models may degrade differently 4. Limited tasks: language only; vision, reasoning, and other abilities untested

Still, as exploratory computational neuroscience, its value lies in a new methodology: controlled ablation experiments to test neuroscience hypotheses. This direction deserves more attention.

Concept Distillation

Another case of "scalar illusion": accuracy is a scalar that hides two independent dimensions—grammar and semantics—each degrading differently. The same accuracy can reflect completely different cognitive damage patterns.

Another case of "structure over content": the Transformer's degradation is determined not by how many neurons are damaged, but by which pathway—the same damage magnitude acts entirely differently on different structures.

Another case of "judgment-gate decoupling": the MLP stores knowledge ("knowing what") while attention routes it ("using what correctly"). Damaging one doesn't affect the other—this is the structural basis of pre-symptomatic reserve.

---

Paper: Can Language Models Get "Aphasia" From Neuronal Ablation? Authors: Pratik Worah (University of Chicago), Randall Bateman (Washington University)