Dumping every skill into context scores worse than giving the agent none — SkillsInjector recovers it to 58.7%

Notes on SkillsInjector: Dynamic Skill Context Construction for LLM Agents (arXiv:2605.29794) — Yanchao Li, Wanhao Liu, Ben Gao, Jiaqing Xie, Zhehong Ai, Na Zou, Yuqiang Li, Tianfan Fu · May 2026

Note published July 24, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

Skill injection is usually treated as a retrieval problem: rank candidates, keep the top K, paste their descriptions in unchanged. That static treatment turns out to be the bottleneck, not the skills themselves. On one tau2-bench airline task with an 82-skill library, per-skill benefit over a no-skill baseline ranges from -0.20 to +0.40: an insurance-policy skill that shares surface keywords with the request actively hurts, while four unrelated policy-reference skills each help substantially. Feed the whole library in at once and pass rate collapses to 24.8% on average across benchmarks — worse than injecting nothing at all (40.2%).

SkillsInjector fixes this with two trained components instead of a fixed shortlist. A context planner — a small MLP scorer trained on execution-grounded benefit rather than semantic similarity — scores each candidate and admits everything clearing a per-domain threshold, so the injected set size floats task by task instead of being fixed in advance. A set-aware renderer then rewrites each selected description in light of its co-injected neighbors (appending lines like 'Not for: refund_policy_reference, compensation_eligibility_criteria' to disambiguate overlapping skills), distilled from a 235B-parameter teacher into an 8B model to keep rewriting cheap at inference. Across tau2-bench, SkillsBench, and ALFWorld, the combined method posts a 58.7% average pass rate versus 53.6% for the strongest baseline (Graph of Skills) — with the largest single-domain jump on ALFWorld, 82.7% vs. 75.4%.