Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Keisha, Figarri, Wu, Zekun, Wang, Ze, Koshiyama, Adriano, Treleaven, Philip
Format:	Preprint
Published:	2025
Subjects:	Computation and Language
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

Personality as a Probe for LLM Evaluation: Method Trade-offs and Downstream Effects
by: Handa, Gunmay, et al.
Published: (2025)

Mind the Gap: Evaluating Model- and Agentic-Level Vulnerabilities in LLMs with Action Graphs
by: Wicaksono, Ilham, et al.
Published: (2025)

HEARTS: A Holistic Framework for Explainable, Sustainable and Robust Text Stereotype Detection
by: King, Theo, et al.
Published: (2024)

From Text to Emoji: How PEFT-Driven Personality Manipulation Unleashes the Emoji Potential in LLMs
by: Jain, Navya, et al.
Published: (2024)

THaMES: An End-to-End Tool for Hallucination Mitigation and Evaluation in Large Language Models
by: Liang, Mengfei, et al.
Published: (2024)

Control Reinforcement Learning: Interpretable Token-Level Steering of LLMs via Sparse Autoencoder Features
by: Cho, Seonglae, et al.
Published: (2026)

Stereotype Detection in LLMs: A Multiclass, Explainable, and Benchmark-Driven Approach
by: Wu, Zekun, et al.
Published: (2024)

CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features
by: Cho, Seonglae, et al.
Published: (2025)

Eliciting Personality Traits in Large Language Models
by: Hilliard, Airlie, et al.
Published: (2024)

The Confidence Manifold: Geometric Structure of Correctness Representations in Language Models
by: Cho, Seonglae, et al.
Published: (2026)

MPF: Aligning and Debiasing Language Models post Deployment via Multi Perspective Fusion
by: Guan, Xin, et al.
Published: (2025)

How LLMs Fail to Support Fact-Checking
by: Proma, Adiba Mahbub, et al.
Published: (2025)

Tool Calling is Linearly Readable and Steerable in Language Models
by: Wu, Zekun, et al.
Published: (2026)

Assessing Bias in Metric Models for LLM Open-Ended Generation Bias Benchmarks
by: Demchak, Nathaniel, et al.
Published: (2024)

When Facts Change: Probing LLMs on Evolving Knowledge with evolveQA
by: Nakshatri, Nishanth Sridhar, et al.
Published: (2025)

JobFair: A Framework for Benchmarking Gender Hiring Bias in Large Language Models
by: Wang, Ze, et al.
Published: (2024)

LibVulnWatch: A Deep Assessment Agent System and Leaderboard for Uncovering Hidden Vulnerabilities in Open-Source AI Libraries
by: Wu, Zekun, et al.
Published: (2025)

Mind the Gap: Comparing Model- vs Agentic-Level Red Teaming with Action-Graph Observability on GPT-OSS-20B
by: Wicaksono, Ilham, et al.
Published: (2025)

SAGED: A Holistic Bias-Benchmarking Pipeline for Language Models with Customisable Fairness Calibration
by: Guan, Xin, et al.
Published: (2024)

Paying Attention to Facts: Quantifying the Knowledge Capacity of Attention Layers
by: Wong, Liang Ze
Published: (2025)

BERT vs GPT for financial engineering
by: Sharkey, Edward, et al.
Published: (2024)

When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs
by: Li, Xiaomin, et al.
Published: (2025)

When Correct Beliefs Collapse: Epistemic Resilience of LLMs under Clinical Pressure
by: Xiao, Boyu, et al.
Published: (2026)

Decoding by Contrasting Knowledge: Enhancing LLMs' Confidence on Edited Facts
by: Bi, Baolong, et al.
Published: (2024)

Tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge
by: Feng, Yiyang, et al.
Published: (2026)

HyPA-RAG: A Hybrid Parameter Adaptive Retrieval-Augmented Generation System for AI Legal and Policy Applications
by: Kalra, Rishi, et al.
Published: (2024)

All for law and law for all: Adaptive RAG Pipeline for Legal Research
by: Keisha, Figarri, et al.
Published: (2025)

Synthetic Fluency: Hallucinations, Confabulations, and the Creation of Irish Words in LLM-Generated Translations
by: Castilho, Sheila, et al.
Published: (2025)

Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data
by: Gerstgrasser, Matthias, et al.
Published: (2024)

When Retrieval Succeeds and Fails: Rethinking Retrieval-Augmented Generation for LLMs
by: Wang, Yongjie, et al.
Published: (2025)

When Hate Meets Facts: LLMs-in-the-Loop for Check-worthiness Detection in Hate Speech
by: Ocampo, Nicolás Benjamín, et al.
Published: (2026)

Cultural Alignment in Large Language Models Using Soft Prompt Tuning
by: Masoud, Reem I., et al.
Published: (2025)

The Illusion of Certainty: Uncertainty Quantification for LLMs Fails under Ambiguity
by: Tomov, Tim, et al.
Published: (2025)

How to Train Your Fact Verifier: Knowledge Transfer with Multimodal Open Models
by: Lee, Jaeyoung, et al.
Published: (2024)

Beyond Training for Cultural Awareness: The Role of Dataset Linguistic Structure in Large Language Models
by: Masoud, Reem I., et al.
Published: (2026)

SK-VQA: Synthetic Knowledge Generation at Scale for Training Context-Augmented Multimodal LLMs
by: Su, Xin, et al.
Published: (2024)

More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration
by: Yadav, Advait, et al.
Published: (2026)

When Contextual Inference Fails: Cancelability in Interactive Instruction Following
by: Bila, Natalia, et al.
Published: (2026)

Paraphrase-Induced Output-Mode Collapse: When LLMs Break Character Under Semantically Equivalent Inputs
by: Liu, Aofan, et al.
Published: (2026)

When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift
by: Le, Khoi, et al.
Published: (2026)