Abstract
Large Language Models (LLMs) are notorious for "hallucinations"—generating fluent but factually incorrect statements. While often attributed to data quality or scaling issues, this paper argues that hallucinations are a fundamental consequence of the standard training objective: maximizing string probability. Because truth is statistically independent of string distribution, LLMs lack an inherent mechanism to ground outputs in reality. We propose a theoretical framework for evidential closure, utilizing a "Learn-Babble-Prune" heuristic to ensure model outputs are synonymous with verified evidence.
1. Introduction: The Statistical Independence of Truth
Current research into LLM hallucinations often focuses on scaling laws or data cleaning. However, this paper posits a more fundamental philosophical and mathematical disconnect. LLMs are trained as "stochastic parrots," optimized to mimic the statistical distribution of strings within a corpus.
The core issue is that truth is an unidentified nuisance parameter in standard training. Because a model observes only strings and not the "state of the world" those strings describe, it has no mathematical basis to prefer a true statement over a false one, provided both are statistically probable. This gap between fluency and factuality necessitates a new approach to grounding.
2. A Modular Architecture for Faithfulness
Drawing from the analytic philosophy of language (notably the work of Donald Davidson and W.V.O. Quine), the paper proposes a "Faithful Multimodal LLM" architecture. This system moves beyond simple next-token prediction by incorporating four distinct learnable components:
- The LLM: Manages linguistic fluency and the conditional probability of strings.
- Perceptual Learner: Interfaces with the external world (e.g., a vision module observing a physical scene).
- Extensional Learner: Maps strings to specific states of the world (referential grounding).
- Intensional Learner: Maps strings to their synonyms, capturing meaning independent of specific reference.
3. The Learn-Babble-Prune Framework
To operationalize these components, the author introduces a rejection sampling heuristic called Learn-Babble-Prune:
- Learn: The agent populates an "Evidence Set" consisting of verified facts and their known synonyms.
- Babble: A standard LLM generates multiple candidate responses based on a prompt.
- Prune: The system filters candidates. A response is only permitted if it is synonymous with a claim in the Evidence Set. If no candidate meets this criterion, the model defaults to "I don't know."
This process ensures evidential closure: the model’s output is strictly limited to the logical and synonymous bounds of its verified evidence. The paper emphasizes that similarity is not synonymy; statistical proximity in a vector space does not guarantee shared truth conditions.
4. Theoretical Results
The paper’s contributions are primarily formal, providing a mathematical foundation for why current models fail and how the proposed framework succeeds:
- The Impossibility Result (Theorem 4.5): Proves that standard probabilistic models cannot distinguish between world structures where a sentence is true versus false, making them inherently non-factual.
- The Data Fallacy (Theorem 4.6): Demonstrates that even training on a 100% verified, "perfect" corpus does not induce factuality, as the model still lacks the mechanism to link strings to world-states.
- Sufficiency of Grounding (Theorems 5.10 & 5.11): Proves that a model integrating perceptual, extensional, and intensional learning will achieve faithfulness—consistency with its own evidence.
5. Discussion and Limitations
While theoretically robust, the "Learn-Babble-Prune" approach faces practical hurdles. Rejection sampling is computationally expensive and inefficient for real-time generation. Future research may involve using Reinforcement Learning (RL) to bake these constraints directly into the model's weights.
Furthermore, the framework relies on a "gold-standard" evidence set, which raises questions about who defines "truth" in contested domains. The current model also treats sentences as primitives, leaving the complexities of compositionality (how word meanings combine) and linguistic vagueness for future iterations.
6. Conclusion
By distinguishing between faithfulness (consistency with evidence) and factuality (consistency with the world), this paper provides a roadmap for moving LLMs beyond mere statistical mimicry toward grounded, reliable communication. Evidential closure ensures that when a model speaks, it does so with the authority of its evidence, rather than the mere probability of its training data.
Summary updated