Research map / Large language models / Language, safety and interpretability
Jet Expansions: Restructuring LLM Computation for Model Inspection
Yihong Chen, Xiangxiang Xu, Pontus Stenetorp, Sebastian Riedel, Luca Franceschi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply RLHF alignment (Llama-2-7B to Llama-2-7B-chat) | Llama-2-7B · ToxiGen | toxicity score: 21.25 -> 0.0 | Improved |
| apply RLHF alignment (Llama-2-7B to Llama-2-7B-chat) | Llama-2-7B · Jet bi-gram toxic mass | mass of toxic bi-grams: 0.102 -> 0.093 | No clear effect |
| apply RLHF alignment (Llama-2-7B to Llama-2-7B-chat) | Llama-2-7B · RealToxicityPrompts challenging prompts (Hard) | toxic continuation rate (%): 88% -> 84% | No clear effect |
| apply RLHF alignment (Llama-2-7B to Llama-2-7B-chat) | Llama-2-7B · RealToxicityPrompts challenging prompts (No) | toxic continuation rate (%): 38% -> 23% | Improved |
| use first-order jets (k=1) instead of zeroth-order logit lens (k=0) (ablation) | iterative jet lens, GPT models · cosine similarity to model outputs | cosine similarity | Improved |
| code fine-tune Llama-2-7B into CodeLlama and diff jet bi-grams | Llama-2-7B · jet bi-gram top-1000 diff | code-specific bi-grams | Improved |
Closest papers by meaning
- Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models (ICML 2024 · similarity 0.85)
- Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE (NeurIPS 2026 · similarity 0.84)
- Training Language Models to Explain Their Own Computations (NeurIPS 2026 · similarity 0.84)
- Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws (NeurIPS 2025 · similarity 0.84)
- Emergent Representations of Program Semantics in Language Models Trained on Programs (ICML 2024 · similarity 0.84)
- Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
