Research area

AI search, skill retrieval, NL2SQL benchmarks, LLM serving systems, and capability governance are treated as system surfaces that should be inspected before they are trusted.

Publication record

Authors

Jiandong Ding, Honglei Ji, Ming Liu, Tao Duan

Venue

arXiv preprint arXiv:2610.11650

Year

2026

Area

LLM Agents

Problem

Similar agent skills can share nearly identical instructions while differing in their conditions of use. Standard query-based text selection tends to keep the shared parts and drop these distinguishing details, so the reranker receives text that cannot separate the candidates.

Approach

SkillContrast is a training-free selector. It compares retrieved candidate skills against each other and keeps the differing text together with local context as the input to a pretrained reranker, so the comparison signal comes from the candidates themselves rather than from the query alone.

Public evidence

On 1,235 requests from SameCapRisk-Bench, SkillContrast yields 54 to 72 more clean hits (retrieving a helpful skill without its risky sibling) than TF-IDF query selection at equal input length, across two retrievers and two reranker sizes. Cutting model-input tokens by 51.1 to 58.8 percent versus full skill bodies costs 10 to 18 clean hits at a 0.6B reranker while matching or beating query selection at 4B. These are benchmark measurements, not production deployment results.

Takeaway

Candidate-relative differences complement query relevance when forming compact reranking inputs, extending the skill-retrieval line from measuring same-family risk exposure to reducing it before reranking.