Removing Hidden Influences from LLM Explanations
π‘A practical inference-time technique to reduce hidden influences in LLM explanations without retraining.
β‘ 30-Second TL;DR
What Changed
Targets incompleteness by removing input concepts that the model did not cite in its explanation.
Why It Matters
The approach could improve auditing and reliability for LLM-assisted decisions without retraining models. However, its effectiveness depends on accurately identifying which input concepts are credited in an explanation.
What To Do Next
Prototype the removal-and-requery loop on your explanation dataset and compare faithfulness against standard prompting using separate completeness and soundness metrics.
Key Points
- β’Targets incompleteness by removing input concepts that the model did not cite in its explanation.
- β’Preserves the influence of concepts mentioned in the explanation while eliminating unmentioned influences.
- β’Improved faithfulness over standard prompting and faithfulness-focused prompting across two datasets, multiple model families, and two metrics.
- β’Requires no model-weight access or parameter updates, making it suitable for inference-time deployment.
Weekly AI Recap
Read this week's curated digest of top AI events β
πRelated Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: ArXiv AI β
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.