Summary of Controllable Context Sensitivity and the Knob Behind It, by Julian Minder et al.
Controllable Context Sensitivity and the Knob Behind It
by Julian Minder, Kevin Du, Niklas Stoehr, Giovanni Monea, Chris Wendler, Robert West, Ryan Cotterell
First submitted to arxiv on: 11 Nov 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 A novel method is introduced to control the sensitivity of language models to their context or prior knowledge. The approach involves designing a task that requires the model to answer questions based on either its contextual or prior knowledge. Fine-tuned versions of popular language models, such as Llama-3.1, Mistral-v0.3, and Gemma-2, are able to solve this task with high accuracy (85-95%). By analyzing these models, a novel linear time algorithm is used to identify the layers that may be important for context sensitivity. A 1-D subspace in a single layer is found to encode whether the model follows context or prior knowledge. This subspace serves as an effective knob in both fine-tuned and non-fine-tuned models of the same family, indicating a simple fundamental mechanism that controls this behavior. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Language models can be trained to predict answers by choosing between their context and prior knowledge. The new approach is designed to control how much the model relies on its context versus its prior knowledge. By creating a task where the model has to answer questions based on either its contextual or prior knowledge, researchers have found that fine-tuned versions of popular language models like Llama-3.1, Mistral-v0.3, and Gemma-2 can solve this task with high accuracy. The results suggest that there is a simple fundamental mechanism controlling how the model chooses between context and prior knowledge. |
Keywords
» Artificial intelligence » Llama
