Predictive and feedback signals differently shape the formation of language representations

arXiv:2605.09409 · q-bio.NC · Submitted 2026-05-10 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Predictive and feedback signals differently shape the formation of language representations".

Marcus: Here's the gist of the paper. The gist: Prediction-based representations best explain neural geometry shared across learners, whereas feedback-related neural signals better predict individual differences in later generalization > How it works The study used a seven-day artificial-language-learning paradigm with 102 adults and collected fMRI data on Day 1 to compare internal representations derived from three matched transformer models: one optimized for prediction (GPT-P), one for feedback (GPT-F), and one combined model (GPT-PF) > The researchers operationalized the prediction objective as next-token prediction and the feedback objective as correctness-supervised grammaticality judgment using a trial-wise feedback structure > The GPT-P model, optimized for next-token prediction, accounted for the largest share of unique neural variance at the group level, despite the human task being feedback-based > Model Performance and Representation Changes The models reached a performance plateau within 200 training epochs, with GPT-P achieving 91% validation accuracy in generating unseen grammatical sentences and GPT-F reaching 100% accuracy on grammaticality judgments > Changes in internal layer representations predicted changes in task performance across training checkpoints, with GPT-P showing significant associations across all layers, while GPT-F showed robustly positive associations strongest in the two hidden layers > Neural Alignment and Geometry Whole-brain searchlight RSA revealed that prediction-based representations accounted for the greatest amount of neural variance during learning, with GPT-P showing significantly higher alignment than the other models in most networks > Prediction-based representations provided the strongest group-level account of neural patterns during early adult language learning, even when the human task was feedback-oriented > Individual Differences and Predictive Signals When examining individual differences in learning outcomes, models trained with neural objectives showed greater similarity to participant-specific neural responses than the behavioral-only model > Incorporating neural constraints improved prediction significantly, reaching a cross-validated predictive power of r = 0.43 (p < 0.

Title and authors: Ines: The authors summarize that adults learn new languages in ways that vary a lot, but they aren't totally clear on what drives those learning processes and individual differences.

Marcus: So the paper’s setup was tracking behavioral learning over seven days while collecting fMRI data from one hundred two adults as they learned an artificial language with corrective feedback.

Yuki: That setup is interesting because it links the actual behavior—what they did—with their internal brain activity at different points in time.

Ines: They trained these matched transformer models, looking at prediction, feedback, or a combination of both, and then compared the resulting internal representations to what was happening in the brain.

Marcus: The main summary point is that representations from the prediction-focused model accounted for the largest share of unique neural variance across all groups during learning.

Yuki: So it supports the idea that prediction is shaping a common neural architecture across learners, even when they are receiving feedback.

Ines: That’s what they concluded, supporting a multi-signal model where prediction sets up the shared structure and feedback mechanisms explain individual differences later on.

Marcus: They also observed that in terms of how well these models predicted learning outcomes on Day seven the neural patterns linked to the feedback model were actually more useful for predicting those specific individual differences <ref:2605.09409#pg2> to the feedback model were>.

Yuki: It’s a nuanced summary: prediction for common structure, feedback for individual divergence. That gives us a better picture of variability.

The paper's summary: Ines: Now let’s look at what the authors suggest we should improve or how we can refine this idea. They propose ways to make this multi-signal model even more robust and predictive for individual learners.

Marcus: One suggestion is that if you want to predict individual learning outcomes better, you shouldn't just rely on behavior alone, but incorporate neural constraints into your prediction objectives.

Yuki: Incorporating neural constraints sounds like forcing the AI model to learn representations that are more biologically plausible or aligned with actual brain activity patterns.

Ines: Right. They found that adding these neural constraints significantly improved the predictive power of the model for individual differences, reaching a cross-validated predictive power of r equals zero point four three <ref:2605.09409#pg2>.

Marcus: That’s a solid number showing that linking the neural objective to the learning prediction actually helps predict who will do better on Day seven <ref:2605.09409#pg2>.

Yuki: So, if we can use neural data to guide the learning process itself, maybe we can better model why some people learn faster than others.

Ines: They also suggest a structural modification to the language model architecture—a weighted multitask objective combining both prediction and feedback losses.

Marcus: That involves dynamically adjusting the weights based on how difficult the task is, like setting alpha equal to the natural logarithm of seventy-nine over two to upweight feedback when it matters most.

Yuki: So they’re suggesting a way to let the model balance predicting grammar structure with correcting errors based on complexity.

Ines: They also mentioned using an autoencoder trained on participant-specific brain data to generate low dimensional neural targets, allowing for joint optimization of multiple objectives in one framework.

Marcus: It sounds like they’re moving toward a system where the AI isn't just guessing the next word, but it’s trying to predict behavior, predict the next token, and also approximate what our brains are actually doing.

The paper's improvements: Ines: Wrapping up this look at Predictive and feedback signals differently shape the formation of language representations, we see that prediction builds a shared structure across learners, while feedback mechanisms help us predict individual generalization differences.

Marcus: The main implication for us is that we need a model that handles both things simultaneously—a common architecture built by prediction, refined by personalized feedback signals.

Yuki: For the wider population genetic view, this suggests that language learning isn't just about mastering rules; it’s about an underlying predictive mechanism interacting with corrective refinement.

Ines: And the work on incorporating neural constraints shows that linking the model to actual brain signals can make those predictions much sharper for specific individuals.

Marcus: So, we have this framework where prediction sets a broad path, but individual neural data fine-tunes where someone will end up on that path.

Yuki: It really gives us something concrete to think about regarding the variability in how different populations acquire complex linguistic skills over time.

Ines: That’s what it is for now with this paper Predictive and feedback signals differently shape the formation of language representations.

Conclusion: Ines: So we're wrapping up on "Predictive and feedback signals differently shape the formation of language representations." We saw that prediction builds a shared structure across learners, while feedback mechanisms help us predict individual generalization differences.

Marcus: The main implication for us is that we need a model that handles both things simultaneously—a common architecture built by prediction, refined by personalized feedback signals.

Ines: And the work on incorporating neural constraints shows that linking the model to actual brain signals can make those predictions much sharper for specific individuals.

Marcus: So we have this framework where prediction sets a broad path, but individual neural data fine-tunes where someone will end up on that path.

Yuki: It really gives us something concrete to think about regarding the variability in how different populations acquire complex linguistic skills over time.

Ines: That’s what it is for now with this paper on Predictive and feedback signals differently shape the formation of language representations. Next up, we’re looking at protein docking and how that harmonic diffusion framework handles flexibility in those biological interactions.

School of Psychology, South China Normal University · Philosophy and Social Science Laboratory of Reading and Development in Children and Adolescents (South China Normal University) · Department of Linguistics and Modern Languages, The Chinese University of Hong Kong · Brain and Mind Institute, The Chinese University of Hong Kong

q-bio.NC

Submitted: 2026-05-10

Updated: 2026-10-21

Code: https://github.com/thedarkkinght/GPTs_learn_Brocanto2

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: The gist: Prediction-based representations best explain neural geometry shared across learners, whereas feedback-related neural signals better predict individual differences in later generalization >

Key concepts

Prediction Objective: This objective trains a language model to predict the next word or token in a sequence. The researchers used this as the training goal for one model, GPT-P. This objective appears to create neural patterns shared across different learners, forming a common learning architecture.
Feedback Objective: This objective trains a language model using trial-wise feedback, such as judging the grammaticality of a sentence. This was used for the GPT-F model. While crucial for human learning, this objective did not account for the largest share of unique neural variance at the group level.
Neural Geometry: This refers to how neural representations are organized and related in brain space. The study found that prediction-based representations exhibited stronger alignment across different brain networks compared to feedback-based ones, suggesting prediction shapes a more common geometric structure for language learning.
Individual Differences Prediction: This involves using models trained on neural objectives to better predict how individuals will perform later in the learning task. By incorporating neural constraints, the model's predictive power increased significantly (r = 0.43), indicating that linking behavior to specific brain activity helps capture personal learning trajectories.

Terminology

Abstraction — 1× in this paper · explained in 1 episode
Generalization — 1× in this paper · explained in 14 episodes
Transformer — 1× in this paper · explained in 2 episodes
Transformer Model — 1× in this paper · explained in 2 episodes
Transformer Models — 1× in this paper · explained in 3 episodes

Summary

The gist: Prediction-based representations best explain neural geometry shared across learners, whereas feedback-related neural signals better predict individual differences in later generalization >

How it works

The study used a seven-day artificial-language-learning paradigm with 102 adults and collected fMRI data on Day 1 to compare internal representations derived from three matched transformer models: one optimized for prediction (GPT-P), one for feedback (GPT-F), and one combined model (GPT-PF) >

The researchers operationalized the prediction objective as next-token prediction and the feedback objective as correctness-supervised grammaticality judgment using a trial-wise feedback structure > The GPT-P model, optimized for next-token prediction, accounted for the largest share of unique neural variance at the group level, despite the human task being feedback-based >

Model Performance and Representation Changes

The models reached a performance plateau within 200 training epochs, with GPT-P achieving 91% validation accuracy in generating unseen grammatical sentences and GPT-F reaching 100% accuracy on grammaticality judgments >

Changes in internal layer representations predicted changes in task performance across training checkpoints, with GPT-P showing significant associations across all layers, while GPT-F showed robustly positive associations strongest in the two hidden layers >

Neural Alignment and Geometry

Whole-brain searchlight RSA revealed that prediction-based representations accounted for the greatest amount of neural variance during learning, with GPT-P showing significantly higher alignment than the other models in most networks >

Prediction-based representations provided the strongest group-level account of neural patterns during early adult language learning, even when the human task was feedback-oriented >

Individual Differences and Predictive Signals

When examining individual differences in learning outcomes, models trained with neural objectives showed greater similarity to participant-specific neural responses than the behavioral-only model >

Incorporating neural constraints improved prediction significantly, reaching a cross-validated predictive power of r = 0.43 (p < 0.

Improvements for AI systems

Predictor-based learning systems can be improved by incorporating next-token prediction objectives into language models to create a common neural learning architecture across learners, as this objective accounted for the largest share of unique neural variance at the group level, despite the human task being feedback-based.
Feedback-related mechanisms can be enhanced by using individual learner data to predict outcomes; specifically, incorporating neural constraints improved prediction significantly, reaching a cross-validated predictive power of r = 0.43 for predicting Day 7 generalization performance on unseen items.
The language model architecture should be modified to use a weighted multitask objective combining prediction and feedback losses, where the weight is dynamically adjusted based on task difficulty: "we set alpha = ln (79/2) to upweight the behavioral term relative to next-token prediction, linking the weight to the difference in baseline cross-entropy between a 79-way token-prediction problem and a binary grammaticality classification problem."
To better capture individual differences, an autoencoder should be trained on participant specific brain data to generate low dimensional neural targets; this allows the model to achieve joint optimization of multiple objectives within a single framework: predicting behavior, predicting the next token, and approximating neural activity.