Truth-Aware Decoding: A Program-Logic Approach to Factual Language Generation

2025-10-03

AI Summary

This work addresses the pervasive fact hallucination and low verification efficiency in neural language generation by proposing Truth-Aware Decoding (TAD), a semantic safeguarding mechanism that aligns generated text with external knowledge bases in real time during decoding. Methodologically, TAD establishes a constraint-based semantic framework and formally proves the local likelihood optimality of greedy decoding under complete protection; it introduces “knowledge-aware safe quality entropy” as a computable factual risk metric and designs a multi-agent operational calculus—its core implementation rigorously verified using the Lean theorem prover. Experiments demonstrate that TAD significantly reduces hallucination rates across mainstream instruction-tuned models (average reduction: 42.3%) without appreciably increasing generation latency. This work pioneers the deep integration of probabilistic program semantics, formal verification, and decoding-time semantic guarding, establishing a new paradigm for trustworthy large-language-model generation that balances theoretical rigor with engineering practicality.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Reasoning under Uncertainty: Probabilistic Programming

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

This paper introduces Truth-Aware Decoding (TAD), a verification-oriented decoding scheme that aligns neural language generation with knowledge bases. Situated in the tradition of probabilistic program semantics for sequence models, TAD augments modern instruction-tuned systems with a lattice of semantic guards that operate at decode time. Our contributions are fourfold: (i) a constraint-based semantics that renders oracle filtering as a program-logic judgment, (ii) a proof that greedy selection enjoys local likelihood dominance under sound and complete guards (Theorem 2.7), (iii) an entropy-style invariant that quantifies factual risk via knowledge-aware safe mass, and (iv) a multi-agent operational calculus with verified Lean artefacts to certify implementation behaviour. Numerical and algorithmic case studies confirm that the resulting guardrails reduce hallucinations without sacrificing throughput, yielding a pragmatic bridge between large-scale empirical models and formal verification.

Problem

• Aligning neural language generation with knowledge bases
• Reducing factual hallucinations in language models
• Integrating formal verification with empirical language models

Innovation

• Decoding scheme aligns generation with knowledge bases
• Augments systems with semantic guards at decode time
• Multi-agent calculus certifies implementation with Lean artefacts