Why Representational and Behavioral Alignment Diverge in Neural Networks

24 concepts13 papers1 proposal

When we build deep learning models that achieve human-level performance on language or vision tasks, it is tempting to assume they have become fundamentally "human-like." We often expect that a model with lower perplexity or higher classification accuracy will naturally mimic the internal mechanics of the human brain and exhibit human-like behavioral patterns.

However, a growing body of research reveals a profound dissociation between three distinct axes of evaluation:

  1. Task Performance: Standard engineering metrics like perplexity or benchmark accuracy.
  2. Representational Alignment: How closely a model's internal activations mirror biological neural activity (such as fMRI or EEG).
  3. Behavioral Alignment: How closely a model's outputs or processing dynamics match human actions (such as reading times or psychophysical error patterns).

Understanding why these dimensions diverge is crucial for both building better artificial intelligence and using neural networks as valid models of human cognition.


The Performance-Representation Disconnect

Why doesn't a better task score mean a more brain-like model? The core reason lies in the optimization landscape. Deep neural networks can find many distinct mathematical solutions to solve the same objective, often exploiting statistical shortcuts that biological systems ignore.

In language modeling, for example, standard metrics like perplexity evaluate a model's next-word prediction capabilities. Yet, Pasquiou et al. demonstrated that a model's proficiency in next-word prediction does not reliably correlate with its ability to mirror human neural activations. A model can dramatically improve its text prediction performance without its internal states becoming any more aligned with the human language system.

Similarly, Hale et al. found that scaling up out-of-domain training data improves a parser's statistical text prediction but fails to improve its ability to predict human brain activity (EEG) during language comprehension. This divergence is also visible in computer vision: Muttenthaler et al. noted that older architectures like AlexNet, despite poor performance on modern object recognition benchmarks like ImageNet, can show remarkably strong correlations with fMRI responses in the higher visual cortex.

Even more strikingly, this relationship can be completely decoupled via model compression. Oota et al. showed that compressing large language models can degrade their performance on complex linguistic benchmarks (like syntax or reasoning) while leaving their capacity to mirror human cortical representations entirely intact. This implies that the neural geometry required to match human brain activity saturates early and does not inherently depend on the mechanisms required to solve complex downstream logic puzzles.


The Many-to-One Mapping of Behavior and Representation

If task performance doesn't guarantee representational alignment, does a model's external behavior at least reflect its internal representations? The answer is frequently no, due to a "many-to-one" mapping: entirely different internal representational structures can produce identical outward behaviors, and vice versa.

As highlighted by Cheung et al., relying solely on behavioral alignment to assess understanding can be deeply misleading because networks can achieve identical task outcomes using fundamentally non-human internal strategies. This insight is supported by Philippsen, who demonstrated using recurrent neural networks that a model can maintain adequate behavioral performance even when its internal understanding of underlying patterns is severely compromised.

Conversely, models can have highly divergent internal structures but yield identical behavioral predictions. Zhou et al. showed that internal representation similarity between neural networks is an unreliable proxy for their behavioral alignment with human perception; models with distinct internal architectures can make identical predictions about stimulus discriminability.

                  ┌───────────────────────────┐
                  │     Task Performance      │
                  │  (Perplexity, Accuracy)   │
                  └─────────────┬─────────────┘
                                │
                    Diverges during scaling
                                │
                                ▼
┌───────────────────────────┐       ┌───────────────────────────┐
│ Representational Alignment│ ◄───► │   Behavioral Alignment    │
│    (fMRI, EEG Matching)   │ Disconnect│ (Reading Times, Errors)   │
└───────────────────────────┘       └───────────────────────────┘

This divergence becomes explicitly visible when we try to align models to the human brain. In a pair of studies, Aw et al. (iclr 2024) and Aw et al. (COLM 2024) discovered that instruction-tuning significantly improves a large language model's internal representational alignment with human fMRI data. However, this internal alignment did not translate to better behavioral alignment with human reading times. The factors that drive a model to represent world knowledge like a human brain do not necessarily govern its surface-level processing dynamics, such as reading pacing.


The Illusion of Brain Scores and Idiosyncratic Invariances

If a model's internal representations correlate highly with human brain data (yielding a high "brain score"), does that mean it processes information like us? Not necessarily. Current alignment metrics often rely on linear regressions or canonical correlation analyses to map model activations to fMRI or EEG signals. These methods can mask deep-seated differences in processing logic.

To expose this, Feather et al. introduced "model metamers"—stimuli synthesized to produce identical activations in a model's internal layers as a natural target stimulus. They found that while these models could accurately predict biological brain activity, their metamers often appeared as completely meaningless noise to humans. This proves that the models rely on idiosyncratic invariances and features that the human brain does not share, despite achieving high brain-matching scores.

Furthermore, Merlo demonstrated that the internal similarity spaces learned by language models fail to correlate with human experimental findings regarding long-distance syntactic dependencies. Similarly, Garcia et al. identified a "social vision gap": while large language models can accurately predict human behavioral ratings of social interactions from text descriptions, they fail to align with the brain's actual visual processing activity when viewing dynamic social scenes. The semantic knowledge captured by language models does not replicate the underlying biological mechanisms of perception.


Moving Toward Multidimensional Alignment Profiles

The fact that task performance, behavioral alignment, and representational alignment can move in completely opposite directions suggests that "alignment" cannot be treated as a single, monolithic metric.

For example, Oh et al. found that language models predict human reading times best when trained on roughly two billion tokens; as training continues and perplexity improves further, their behavioral alignment with humans actually declines, as they develop "superhuman" predictive capabilities that diverge from natural human processing.

To address this, Ahlert et al. argue that different metrics for measuring alignment—such as neural activity matching and behavioral error patterns—often show very low correlation with one another. Consequently, evaluating the cognitive fidelity of artificial intelligence requires constructing a multidimensional profile rather than relying on a single aggregated score, ensuring that a model's success along one dimension does not obscure its failure along another.

Key Foundational Work

Text Genre and Training Data Size in Human-like Parsing (2019) — contributed 2 concept elements to this direction.
Deficits in Prediction Ability Trigger Asymmetries in Behavior and Internal Representation (2020) — contributed 2 concept elements to this direction.
Syntactic Parsing in Humans and Machines (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_representation_behavior_divergence_6b972a9d,
  title = {representation-behavior divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/representation-behavior-divergence/rdr_6b972a9de26e4274be115bae386fe1fe},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_representation_behavior_divergence_6b972a9d_v1,
  title = {representation-behavior divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/representation-behavior-divergence/rdr_6b972a9de26e4274be115bae386fe1fe/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake