[Paper Review] Are Language Models More Like Libraries or Like Librarians? Bibliotechnism, the Novel Reference Problem, and the Attitudes of LLMs

Harvey Lederman, Kyle Mahowald|arXiv (Cornell University)|Jan 10, 2024

Digital Humanities and Scholarship9 citations

TL;DR

The paper defends bibliotechnism, showing how LLMs can produce derivatively meaningful novel text via input causality, but raises a Novel Reference Problem that challenges this view and motivates agency-based interpretations.

ABSTRACT

Are LLMs cultural technologies like photocopiers or printing presses, which transmit information but cannot create new content? A challenge for this idea, which we call bibliotechnism, is that LLMs generate novel text. We begin with a defense of bibliotechnism, showing how even novel text may inherit its meaning from original human-generated text. We then argue that bibliotechnism faces an independent challenge from examples in which LLMs generate novel reference, using new names to refer to new entities. Such examples could be explained if LLMs were not cultural technologies but had beliefs, desires, and intentions. According to interpretationism in the philosophy of mind, a system has such attitudes if and only if its behavior is well explained by the hypothesis that it does. Interpretationists may hold that LLMs have attitudes, and thus have a simple solution to the novel reference problem. We emphasize, however, that interpretationism is compatible with very simple creatures having attitudes and differs sharply from views that presuppose these attitudes require consciousness, sentience, or intelligence (topics about which we make no claims).

Motivation & Objective

• Assess whether LLM outputs are meaningfully derivative of human inputs under bibliotechnism.
• Demonstrate how novel yet derivatively meaningful text can arise from n-gram and higher-order models.
• Introduce and analyze the Novel Reference Problem where LLMs invent references not grounded in training data.
• Evaluate potential responses (RLHF, creator intentions, prompts, reader interpretation) and their impact on meaning.

Proposed method

• Employ n-gram toy models to illustrate derivative meaning and causal connections to PrimaryData.
• Extend from unigram to higher-order n-grams to show derivatively meaningful complex expressions.
• Argue that intelligibility is a high-level feature that can transfer meaning to novel GeneratedText.
• Discuss how LLMs can produce novel reference and the implications for bibliotechnism.
• Evaluate potential defenses of bibliotechnism via RLHF, user/creator intentions, and reader-centric semantics.

Experimental results

Research questions

• RQ1 Can LLMs produce inscriptions that are derivatively meaningful, including complex expressions?
• RQ2 Can LLMs generate novel reference not grounded in their training data, and what does this imply for derivative meaning?
• RQ3 Does the presence of novel reference support attributing beliefs, desires, and intentions to LLMs?
• RQ4 What responses (RLHF, creator/user intentions, reader interpretations) can preserve or challenge bibliotechnism?

Key findings

• Unigram models can yield derivatively meaningful words but struggle to produce derivatively meaningful complex expressions.
• Higher-n models can generate derivatively meaningful long text by copying from PrimaryData and coupling with intelligibility.
• LLMs can produce novel references (e.g., Marion Starlight) not grounded in PrimaryData, challenging purely derivative accounts.
• The Novel Reference Problem suggests bibliotechnism may be weaker than agency-based explanations for certain outputs.
• Reader-oriented metamSemantic approaches may help but face challenges in distinguishing intended meaning from intelligibility.