Follow the Gradient: Tracing Safety Degradation Back to Training Data

27 Oct 2026Reported by arXiv21

Follow the Gradient: Tracing Safety Degradation Back to Training Data. It centres on Fine-tuning, and also names Qwen and Llama. Reported by arXiv. Bharat Hunt files it under AI Models and AI Coding — the section covering a new or updated model, its capabilities, benchmarks or availability.

Written by Bharat Hunt from the headline and the coverage below. The original reporting is the source of truth.

Read original coverage

Companies, models and people in this story

Reported by arXiv

First seen 27 oct 2026

arXivFollow the Gradient: Tracing Safety Degradation Back to Training Data27 Oct 2026Saket Reddy

AI Tools24

Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation

Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation. It centres on Benchmarks, and also names Ollama and Phi. Reported by arXiv. Bharat Hunt files it under AI Tools and AI Models — the section covering applications people use — assistants, editors, search, productivity.

arXiv22 Sept 2026

AI Models23

Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment

Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment. It centres on Meta, and also names Llama and Nothing. Reported by arXiv. Bharat Hunt files it under AI Models — the section covering a new or updated model, its capabilities, benchmarks or availability.

← All AI stories

Follow the Gradient: Tracing Safety Degradation Back to Training Data

Companies, models and people in this story

Reported by arXiv

Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation

Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment

Fine-Tuning LLMs for Translation: General Forgetting Mitigation Does Not Preserve MT-Specific Instruction Following

Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers

Companies, models and people in this story

Reported by arXiv

Related AI coverage

Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation

Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment

Fine-Tuning LLMs for Translation: General Forgetting Mitigation Does Not Preserve MT-Specific Instruction Following

Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers