Pith. sign in

REVIEW 5 cited by

Know When To Stop: A Study of Semantic Drift in Text Generation

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2404.05411 v1 pith:TLP66XNL submitted 2024-04-08 cs.CL

classification cs.CL
keywords generationwhenaccuracydriftfactsfactualinformationsemantic
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

In this work, we explicitly show that modern LLMs tend to generate correct facts first, then "drift away" and generate incorrect facts later: this was occasionally observed but never properly measured. We develop a semantic drift score that measures the degree of separation between correct and incorrect facts in generated texts and confirm our hypothesis when generating Wikipedia-style biographies. This correct-then-incorrect generation pattern suggests that factual accuracy can be improved by knowing when to stop generation. Therefore, we explore the trade-off between information quantity and factual accuracy for several early stopping methods and manage to improve factuality by a large margin. We further show that reranking with semantic similarity can further improve these results, both compared to the baseline and when combined with early stopping. Finally, we try calling external API to bring the model back to the right generation path, but do not get positive results. Overall, our methods generalize and can be applied to any long-form text generation to produce more reliable information, by balancing trade-offs between factual accuracy, information quantity and computational cost.

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Illocutionary Explanation Planning for Source-Faithful Explanations in Retrieval-Augmented Language Models

    cs.CL 2026-03 conditional novelty 7.0 of 10

    Chain-of-illocution prompting improves source adherence in RAG explanations for programming education by up to 63% over baselines.

  2. Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook

    cs.CL 2026-03 unverdicted novelty 6.0 of 10

    DOVE constructs a value codebook via rate-distortion variational optimization from 10K documents and measures LLM-human cultural alignment through unbalanced optimal transport, showing 31.56% correlation with downstre...

  3. Lethe: Purifying Backdoored Large Language Models with Knowledge Dilution

    cs.CL 2025-08 conditional novelty 6.0 of 10

    LETHE uses parameter-level model merging plus prompt-level word definitions to dilute backdoor behavior in LLMs, cutting attack success to below 7% in most tested settings.

  4. Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook

    cs.CL 2026-03 unverdicted novelty 5.0 of 10

    DOVE measures LLM cultural value alignment via a rate-distortion value codebook and unbalanced optimal transport between human and model open-ended text distributions.

  5. Generative AI and Digital Ecosystem Resilience: A Proactive Lifecycle-Based Survey

    cs.LG 2026-05 unverdicted novelty 3.0 of 10

    This survey introduces the C5 Interaction Model as a unifying taxonomy to synthesize proactive detection methods for GenAI-enabled adversarial narratives across socio-technical and computational research streams.

Pith tools