Community search over short user-generated posts, questions and comments. Focus: relevance, nDCG, CTR, query understanding, retrieval and ranking.
The method calibrates on-policy distillation with information gained after retrieval, adjusting both turn-level supervision and query preferences. The 2B student gains 3.4 accuracy points on average across six multimodal search benchmarks.
Key insight: Train against what a query retrieves, rather than its appearance to the teacher. Token volume can underweight short but informative search turns.
The study separates finding current evidence from using it in the final answer. Across six open-weight configurations and a proprietary deployment, models sometimes rely on snippets and retain stale answers despite available evidence.
Key insight: Track evidence discovery, page opening and final answer use separately. Better retrieval alone may leave a stale-answer failure unresolved.
TRACER indexes observed tool-call sequences in a trie and combines empirical and contextual priors to rank continuations. In a mixed retail and telecom setting, tool-selection accuracy improves by 4.7 points.
Key insight: Use action history as retrieval context. Measure coverage of rare trajectories as well as average selection accuracy when shrinking the candidate registry.
SSR uses structural lexical signals from Agent Cards and an inverted index to preselect agents without training. At 1,000 agents, orchestration latency falls by 70.9% relative to passing the complete catalog.
Key insight: Field structure can make lexical retrieval useful before orchestration. The evidence concerns capability matching, rather than relevance ranking of community posts.
GACC-Bench tests whether compressed context preserves information for a goal before the exact questions are known. It contains 500 goals. Goal conditioning helps, but even the best tested compressor falls short of full context.
Key insight: Evaluate compression against several withheld questions per goal. A good answer to one question can hide the loss of other relevant details.
CARA evaluates evidence through addition and removal tests, then combines these signals with path consistency for graph exploration. It is tested on three STaRK datasets; the abstract does not give a single aggregate improvement.
Key insight: Test whether evidence changes answerability rather than treating semantic similarity as utility. Account for the extra evaluation calls when measuring retrieval cost.