Reverse Needle: Training LLMs to Attend to Absence via Contrastive Learning

TL;DR: What if we teach language models to notice what's missing by showing them pairs of documents—one complete, one with gaps—and explicitly training them to spot the differences? An initial experiment could use contrastive learning to align representations of original and edited documents, hypothesizing that models trained this way will outperform standard LLMs on AbsenceBench tasks.

Research Question: Can contrastive learning on paired original and edited documents improve LLMs' ability to detect missing information, bridging the gap identified by AbsenceBench?

Hypothesis: Contrastive pretraining, using positive pairs (unaltered documents) and negative pairs (documents with deliberate omissions), will enable LLMs to develop internal representations that are sensitive to absences, leading to significant performance improvements on missing information detection tasks.

Experiment Plan: - Data: Use AbsenceBench datasets (numerical sequences, poetry, GitHub PRs).

• Method: Pretrain LLMs using a contrastive loss, where the model is tasked with distinguishing original from edited versions.
• Baseline: Compare to standard supervised fine-tuning and existing SOTA models.
• Metrics: F1-score on AbsenceBench; qualitative analysis of attention maps to see if the model attends to "gaps."
• Expected Outcome: Models trained with contrastive objectives should show higher sensitivity to missing content and better F1-scores than baseline models.

References:

• Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature.
• Chuang, Y.-S., Qiu, L., Hsieh, C.-Y., Krishna, R., Kim, Y., & Glass, J. (2024). Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Maps. Conference on Empirical Methods in Natural Language Processing.