Beyond Human-Like Processing: Large Language Models Perform Equivalently on Forward and Backward Scientific Text (2411.11061v1)

Abstract: The impressive performance of LLMs has led to their consideration as models of human language processing. Instead, we suggest that the success of LLMs arises from the flexibility of the transformer learning architecture. To evaluate this conjecture, we trained LLMs on scientific texts that were either in a forward or backward format. Despite backward text being inconsistent with the structure of human languages, we found that LLMs performed equally well in either format on a neuroscience benchmark, eclipsing human expert performance for both forward and backward orders. Our results are consistent with the success of transformers across diverse domains, such as weather prediction and protein design. This widespread success is attributable to LLM's ability to extract predictive patterns from any sufficiently structured input. Given their generality, we suggest caution in interpreting LLM's success in linguistic tasks as evidence for human-like mechanisms.

Analysis of LLMs Trained on Reversed Text for Scientific Understanding

The paper "Beyond Human-Like Processing: LLMs Perform Equivalently on Forward and Backward Scientific Text" explores how LLMs perform when trained on backward tokenized text—a syntactic structure inconsistent with natural human language. The investigators employed transformer-based LLMs, specifically variations of GPT-2, trained on two decades worth of neuroscience literature, testing their performance on a neuroscience benchmarking tool, BrainBench.

Insights and Methodology

1. Training Setup:
   • The authors employed the GPT-2 architecture, training models from scratch on 20 years of neuroscience data. They used three different sizes of GPT-2 models (124M, 355M, and 774M parameters).
   • Two separate datasets were prepared: one with forward tokenization and the other with character-level reversed tokenization.
2. Benchmarking with BrainBench:
   • BrainBench was used to evaluate the prediction accuracy of trained models. This benchmark includes neuroscience abstracts with altered outcomes, challenging models to differentiate between the original and altered abstracts.
   • Model decisions were guided by perplexity scores, with lower perplexity showing higher confidence in the abstract's correctness.
3. Key Findings:
   • Despite the syntactic inconsistency presented by reversed text, backward-trained models' performance was statistically indistinguishable from forward-trained models.
   • Larger models, particularly GPT-2 774M, either paralleled or exceeded human expert performance.
   • Backward-trained models exhibited higher perplexity relative to forward-trained models when validating against unseen data, yet this did not manifest as a substantial performance decrement on the predictive tasks.

Implications and Theoretical Observations

The findings of this research contribute to ongoing discussions regarding the nature of LLMs as models for human language processing. The equivalence in performance between forward- and backward-trained models on BrainBench tasks underscores a significant insight: transformers possess a generalized ability to process patterns and predict outcomes, irrespective of input alignment with human cognitive constraints.

• Generalizable Pattern Recognition: The models are adept at identifying and leveraging predictive patterns through their sizable datasets. This general pattern extraction capability is manifested in their robust performance across varied inputs, an indicator that these architectures can process abstract constructs broadly—beyond mere linguistic mimicry.
• Cognition and LLMs: The human cognitive system evolved to process language linearly, optimizing linguistic parsing in real-time communication. Despite LLMs' high performance on tasks and alignment-independent processing, their operation should not be conflated with human cognition due to these evolutionary and operational divergences.
• Training and Tokenization Considerations: The paper's methodological shift—training backwards models with a character-reversed tokenizer—diverges from past efforts that only reverse text at word/token boundaries. This approach underlies its nuanced examination of the flexibility of LLM tokenization and their ability to engage with non-standard input formats.

Speculation on Future Developments

This paper exposes the versatility of LLM architectures in handling structurally unconventional data, inviting further exploration into their limits and applications. Future work could delve into applications of reversed text training in other domains, exploring whether cognitive load or efficiency varies across other transformer variables and task designs. Additionally, the implications for cognitive modeling and the continued separation of human language comprehension from artificial pattern recognition warrant deeper investigation to unravel the depth of cognitive parallels, or lack thereof, between humans and LLMs.

This research highlights the necessity of maintaining a clear distinction between machine processing capabilities and human cognitive phenomena to understand better the potential and limitations of advanced artificial intelligence in linguistic and non-linguistic performances.