Community search over short user-generated posts, questions and comments; relevance, query understanding, retrieval and ranking.
EKLAV provides teacher reasoning as training context while supervising only answers. It wins 41 of 48 model-domain comparisons on BRIGHT reranking.
Key insight: Reasoning traces need not be imitation targets. Changing where supervision applies can reuse a fixed teacher corpus with less training and inference work.
Jev supplies relevance labels and reranks Portuguese search results. On Quati, nDCG@10 rises from 0.3329 to 0.6331 over BM25.
Key insight: Label agreement and agreement on system rankings measure different things. An assessor can preserve comparative evaluation even when individual judgments differ.
PRSR removes language-specific directions estimated from parallel sentences, without training the underlying LLM. Monolingual evaluation spans 18 languages.
Key insight: Parallel-sentence residuals offer a concrete way to separate language identity from meaning before similarity search.
A local bilingual banking search pipeline combines hybrid retrieval, reranking and evidence-based answer gating. Top-1 hit rate rises from 50.48% to 79.05% on 210 queries.
Key insight: Separating related-but-mismatched evidence from sufficient evidence makes retrieval failures more inspectable. The small, paraphrase-based evaluation and automated judging limit the evidence.
Code Skill Server indexes 993 reusable coding skills and evaluates their retrieval through downstream code generation. The best configuration improves Pass@1 by up to 13.91%.
Key insight: Skill retrieval should be evaluated by task completion as well as relevance. This makes downstream execution a useful check on retrieval choices.
ReaLiC trains explicit rewriting strategies with rewards for retrieval effectiveness and contextual consistency. It evaluates two conversational retrieval benchmarks; the accessible preview does not expose numerical gains.
Key insight: Reward both usefulness and intent preservation. Explicit strategy descriptions make rewriting behavior easier to inspect, though numerical results were unavailable in the accessible preview.
SAG2 combines SQLite, BM25 and query-time evidence decisions for multi-hop retrieval. Recall@10 reaches 86.62% on a 1,000-question MuSiQue subset.
Key insight: Moving semantic decisions to query time reduces offline preparation, while shifting work to iterative online evidence selection.
Gap Detection searches embedding space for relevant but uncovered directions. Across five curated topics, mean relative output-collision reduction is 12.8%.
Key insight: Relevance and novelty can be scored separately before generation. Geometric separation still needs task-specific validity checks.
MetaMeta describes power, cooling and capacity planning for its Prometheus training cluster, with a roadmap toward roughly one million GPUs. This is an infrastructure target, not a measured search-quality gain.
Key insight: Compute expansion depends on power and cooling capacity. This industrial coverage exception has limited direct relevance to retrieval ranking.