Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 2·5 min read
StateSwap: Probing Support-Elimination Hidden States in Multiple-Choice Questions
TL;DR
StateSwap introduces a method to analyze how large language models respond differently to multiple-choice questions based on framing. By probing internal states, the study reveals that these discrepancies can be systematically manipulated to improve prediction consistency.
✦ Why It Matters
Experiment with dual-framing techniques to enhance the consistency of your model's predictions on multiple-choice tasks.
Key Takeaways
- Implement dual-framing protocols in your model evaluations.
- Use special tokens to probe internal model states effectively.
- Analyze the impact of framing on model predictions.
- Explore mean-difference steering for more stable responses.
How It Works
StateSwap employs a dual-framing protocol that uses two types of prompts—support-oriented and elimination-oriented—to evaluate the same question. By appending an untrained token, [STATE], to the model's input, the researchers can capture the model's internal activations in response to each framing.
The distinct activations observed in intermediate layers indicate that the model processes the two framings differently. By swapping these activations between prompts, the researchers can manipulate the model's predictions, demonstrating that the internal states are behaviorally relevant.