Summary of Diesel — Dynamic Inference-guidance Via Evasion Of Semantic Embeddings in Llms, by Ben Ganon et al.
DIESEL – Dynamic Inference-Guidance via Evasion of Semantic Embeddings in LLMs
by Ben Ganon, Alon Zolfi, Omer Hofman, Inderjeet Singh, Hisashi Kojima, Yuval Elovici, Asaf Shabtai
First submitted to arxiv on: 28 Nov 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed paper presents a lightweight inference-guidance technique called DIESEL to filter out undesired concepts from large language models’ (LLMs) responses. DIESEL can seamlessly integrate into any autoregressive LLM and semantically filter concepts based on their similarity to predefined negative concepts in the latent space. This technique can be used as a standalone safeguard or an additional layer of defense, enhancing response safety by reranking tokens proposed by the LLM. The paper demonstrates DIESEL’s effectiveness on state-of-the-art conversational models, even in adversarial scenarios that challenge response safety. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The researchers developed a way to keep language models from generating unsafe responses. They created a technique called DIESEL that can be added to any language model to make sure it doesn’t suggest harmful or inappropriate ideas. This is important because language models are getting really good at talking, but sometimes they come up with things that might not be safe or appropriate. The researchers tested their technique and found it works well even when the language model is trying to come up with something that’s not safe. |
Keywords
» Artificial intelligence » Autoregressive » Inference » Language model » Latent space
