TL;DR
This paper investigates whether scaling up Large Language Models (LLMs) or fine-tuning them with instructions better aligns their internal representations with human brain activity during natural language processing. The study finds that simply increasing model size (scaling) significantly improves this alignment, while instruction tuning does not. This suggests that scaling is a more crucial factor for LLMs to mimic human cognitive processes in language comprehension.
What Problem Does This Paper Address?
As LLMs become increasingly powerful, there's a growing interest in understanding how well they reflect human cognitive processes, particularly in language comprehension. While LLMs have shown impressive performance on various NLP tasks, their sheer size and ability to process vast contexts raise questions about their plausibility as cognitive models. Specifically, this paper addresses whether two key LLM development techniques – scaling (increasing model size and training data) and instruction tuning (fine-tuning with specific instructions) – contribute differently to aligning LLM representations with human brain activity during naturalistic reading.
Methods
The researchers compared the self-attention mechanisms of various LLMs (both base and instruction-tuned versions) against human brain data. They used two primary methods:
- Eye-tracking: Analyzing regressive eye saccades, which are thought to reflect how readers revisit and process previous information, similar to how attention weights function in LLMs.
- fMRI: Analyzing functional magnetic resonance imaging data to capture brain activity patterns during naturalistic reading.
They employed ridge regression to model the relationship between LLM attention matrices and human behavioral (eye-tracking) and neural (fMRI) data. To assess the impact of scaling and instruction tuning, they systematically compared models of different sizes and their fine-tuned counterparts. They also investigated the models' sensitivity to instructions and "trivial patterns" (e.g., focusing on preceding words) to understand their internal workings.
Datasets
The primary dataset used was the Reading Brain dataset (Li et al., 2022), which contains concurrent eye-tracking and fMRI data from 50 participants reading five English STEM articles.
For LLMs, they used:
- GPT-2 models: Including base versions of various sizes.
- LLaMA models: Including base versions of sizes 7B, 13B, 30B, and 65B.
- Fine-tuned LLMs: Alpaca (7B and 13B), Vicuna (7B and 13B), Gemma-Instruct 7B, and Mistral-Instruct 7B.
An additional fMRI dataset from a Chinese audiobook listening task was used to test the generalizability of findings.
Key Insights and Ideas
- Scaling is King: The most striking finding is that increasing LLM size consistently leads to better alignment with human brain activity, reinforcing the "scaling law" observed in other areas of deep learning. This suggests that larger models inherently capture more human-like linguistic processing.
- Instruction Tuning Falls Short (for this task): Despite its success in improving LLM performance on specific tasks, instruction tuning did not significantly enhance the alignment with naturalistic reading brain data. This implies that instruction-following behavior might not directly translate to better mirroring of fundamental language comprehension processes.
- LLMs and Human Reading: The study highlights that regressive eye saccades, a key aspect of human reading behavior, can be predicted by LLM attention patterns, especially in larger models. This provides a behavioral link to the neural findings.
- Trivial Patterns vs. Cognitive Plausibility: While LLMs exhibit tendencies to focus on "trivial patterns" (like preceding words), larger models showed less reliance on these, potentially contributing to their greater cognitive plausibility.
Key Results
- Model-Brain Alignment: As LLM size increased from 774M to 65B parameters, alignment with human eye movement and fMRI activity patterns significantly improved. This improvement followed a clear scaling law.
- Scaling vs. Fine-tuning: For models of the same size, base and fine-tuned versions showed no significant difference in their alignment with human reading behaviors and brain patterns.
- Eye-Tracking: Larger LLMs showed a stronger correlation with regressive eye saccade patterns.
- fMRI: Larger LLMs showed significantly more activation in bilateral temporal-parietal brain regions when predicting fMRI activity.
- Instruction Sensitivity: Fine-tuned models showed a greater divergence in attention matrices when processing instructed versus plain text, suggesting they are more sensitive to instructions than humans might be in naturalistic settings.
Limitations and Future Work
- Task-Specific Data: The authors acknowledge that while instruction tuning might improve alignment on task-specific brain data (e.g., when participants are actively answering questions), they lacked such datasets to directly test this. Future research could explore this in controlled experimental settings.
- Fine-tuning Techniques: The study examined specific fine-tuning methods. Future work could explore whether other fine-tuning techniques might yield different results.
- Dataset Size: While the study used a relatively small dataset compared to some prior LLM-brain studies, it still found consistent scaling effects, suggesting the findings hold even with limited data.
- Generalizability: The study used English text. While an additional Chinese audiobook dataset was used to test generalizability, further cross-lingual and cross-modal studies could be beneficial.
Summary updated