Re-Ranking Through an Attribution Lens for Citation Quality in Legal QA

Legal QA systems built on retrieval-augmented generation face a fundamental mismatch: semantic similarity, the standard ranking metric, fails to surface passages that language models actually cite. Researchers discovered this gap using attribution methods like C-LIME on the AQuAECHR benchmark, where random retrieval outperformed similarity-based ranking for gold citations. The fix involves training a lightweight cross-encoder on perturbation-based attribution scores to re-rank candidates before generation. This work exposes a critical blind spot in RAG pipelines and suggests that post-hoc explanation techniques can be repurposed as ranking signals, with implications for any domain where citation fidelity matters.

Modelwire context

Explainer

The paper's real contribution isn't just identifying the mismatch between similarity ranking and citation behavior. It's showing that post-hoc explanation techniques (C-LIME) can be inverted into training signals for re-ranking, suggesting a broader pattern: interpretability methods have dual use as performance levers.

This connects directly to the causal tracing work on sparse MoE models (arXiv cs.CL, June 2) and the clinical provenance categorization paper (arXiv cs.CL, June 1). Both those pieces also grapple with the gap between what a model outputs and where that output comes from. Here, the researchers use attribution to close that gap in retrieval ranking. The pattern across all three is the same: understanding model internals (via causal methods or fine-grained categorization) becomes operationally useful, not just interpretatively interesting. For RAG systems specifically, this exposes why citation quality has been a persistent failure mode that semantic metrics alone cannot fix.

If the re-ranked cross-encoder approach maintains its citation improvement on out-of-domain legal corpora (e.g., UK case law or EU directives not in AQuAECHR), that confirms the method generalizes. If it degrades significantly, the signal is dataset-specific and the approach is less portable than claimed.

Coverage we drew on

Expert-Aware Causal Tracing of Factual Recall in Sparse MoE Language Models · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsAQuAECHR · C-LIME · retrieval-augmented generation · cross-encoder

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.