activation-probes-claim-correctness
Linear probes on Llama 3.1 8B hidden states rank claim correctness better than the model's stated confidence; legal QA and biography probes, Modal runs. Write-up: casella.dev/blog_interp.html
- 0
- Python
- MIT
- 8 wk ago
Linear probes on Llama 3.1 8B hidden states rank claim correctness better than the model's stated confidence; legal QA and biography probes, Modal runs. Write-up: casella.dev/blog_interp.html
Replication package for “Illocutionary Explanation Planning for Source-Faithful Explanations in Retrieval-Augmented Language Models” (xAI 2026): code, data, prompts, and evaluation for chain-of-illocution prompting in textbook-grounded RAG.