Idea
Reranking retrieved papers by asking an LLM to compare candidates pairwise is effective and expensive: exhaustive comparison is O(N²), which is 190 pairs at K = 20. It is also noisy — the same judge can be inconsistent across pairs.
Both problems have one shape: noisy pairwise preferences under a comparison budget, i.e. a dueling bandit. Applying UCB elimination to the comparison schedule cuts 190 exhaustive pairs to roughly 44, spending comparisons on candidates that are still plausibly the best rather than on pairs whose outcome is already decided.
Citation context as evidence
Titles and abstracts describe what a paper claims about itself. The sentences in which other papers cite it describe what it is actually used for — which is closer to what a searcher wants.
- Built a corpus of 10,395 papers and 497,210 citation contexts by breadth-first traversal of LATEX sources.
- Augmenting each candidate with its in-link citation context raised MRR from 0.485 to 0.572 at K = 5, over reranking on titles and abstracts alone.
Downstream writing pipeline
The retrieval front-end feeds a retrieval-augmented writing pipeline, evaluated for factual consistency against its source material rather than for fluency: 0.64 SummaCConv, close to the 0.65 scored by human-written text, and 0.87 semantic similarity to the human reference.