On Improving Faithfulness of Podcasts from Documents

Technical Summary: Improving Faithfulness of Podcasts from Documents

Problem Statement
Large Language Models (LLMs) are increasingly utilized to generate long-form, multi-speaker conversational content, such as podcasts, from textual source documents. While these systems produce fluent narratives, they frequently introduce "ungrounded" information—content that is factually plausible or correct in the real world but not supported by the specific source document. This issue is distinct from general hallucination; a statement may be true but unfaithful to the provided context.

Existing research has largely focused on faithfulness in short-form tasks (e.g., summarization, QA) or on podcast attributes like fluency and engagement. However, there is a lack of systematic study regarding faithfulness in document-grounded podcast generation, where models must maintain factual accuracy across extended, multi-turn dialogues. The core challenge lies in evaluating and correcting ungrounded content at the level of individual conversational turns without disrupting the narrative flow or requiring access to model internals (logits), which is necessary for black-box or closed-source models.

Methodology
The authors propose a comprehensive framework involving dataset construction, evaluation, and a mitigation strategy.

1. Dataset Construction:

   • The authors constructed a dataset of 1,520 documents spanning five domains: academic, legal, policy, financial, and medical.
   • Podcast transcripts were generated from these documents using multiple LLMs (including GPT-4o, LLaMA-3.3-70B, and various Qwen-2.5 models) in a sequential, turn-by-turn manner, simulating a two-speaker structure (host and expert).

2. Evaluation Framework (Turn-Level LLM-as-a-Judge):

   • The authors introduce a turn-level evaluation protocol where an LLM (specifically GPT-4o) acts as a judge to assess the faithfulness of each conversational turn against the source document.
   • Instead of binary classification, the judge assigns a Likert-style score (1–5), where 1 is entirely unfaithful and 5 is fully faithful. This allows for the capture of nuanced cases, such as partial grounding.
   • The reliability of this framework was validated against human annotations, showing a Pearson correlation of 0.63, significantly outperforming existing automatic metrics like Knowledge-F1 and FaithCritic.

3. Mitigation Strategy: Catch-n-Repair:

   • To address unfaithfulness, the authors propose catch-n-repair, a model-agnostic framework that operates during the generation process.
   • Detection (Catch): An auxiliary model (a LoRA-fine-tuned LLaMA-3.1-8B) is trained on a synthetic dataset of faithful and unfaithful turns to classify whether a generated turn is grounded in the source document.
   • Correction (Repair): When a turn is flagged as unfaithful, the generation model is re-prompted to rewrite the turn using only facts explicitly stated in the source document, ensuring the correction preserves conversational flow. This process happens sequentially, preventing the propagation of ungrounded content to subsequent turns.

Key Results

• Baseline Performance: Even state-of-the-art models, including GPT-4o, frequently generate ungrounded content. GPT-4o achieved the highest average faithfulness score (3.89/5), followed by Qwen2.5-32B (3.75). The study found that faithfulness varies significantly by domain (e.g., policy domains scored higher than academic ones) and that larger open-source models do not always guarantee better faithfulness.
• Evaluator Reliability: The LLM-as-a-judge framework demonstrated substantial agreement with human annotators (Krippendorff's $\alpha$ = 0.69) and outperformed lexical overlap metrics and learned evaluators in correlating with human judgments.
• Effectiveness of Catch-n-Repair: The proposed mitigation strategy consistently improved faithfulness scores across all tested models and domains.
  • In-domain (ID) improvements ranged from +0.1 to +0.3 points.
  • Out-of-domain (OOD) improvements were more pronounced, with gains up to +0.5 points (e.g., LLaMA-3.3-70B).
  • The strategy proved robust to domain shifts, with average relative increases in faithfulness being higher for OOD sets (10%) compared to ID sets (5%).
  • Qualitative analysis confirmed that the method successfully reverts ungrounded tangential discussions back to the scope of the source document.

Significance and Claims
The paper claims to present the first systematic study of faithfulness in document-grounded podcast generation, addressing a gap in the literature where prior work focused on fluency and structure rather than grounding. The authors emphasize that:

1. Turn-level granularity is critical: Faithfulness must be assessed and maintained at the level of individual conversational turns in long-form generation, as ungrounded content is often localized but can propagate through the narrative.
2. Model-agnostic mitigation is viable: The catch-n-repair framework demonstrates that faithfulness can be improved without requiring access to model internals, making it applicable to both open-source and closed-source (black-box) LLMs.
3. LLM-as-a-judge is reliable for long-form: The study validates that LLM-based evaluation is a more reliable proxy for human assessment in long-form conversational settings than traditional lexical or semantic similarity metrics.

The authors conclude that while significant progress has been made in generative AI, maintaining faithfulness in long-form, multi-speaker conversations remains a distinct and challenging problem that requires specific evaluation and mitigation strategies.