Skill Injection Attribution
Accepted to Findings of EMNLP 2026 · Public Artifact

When a function-calling gain is really an interface gain.

This project tests whether prompt-prepended skills improve tool-use procedures or simply teach the output contract expected by the scorer.

这个项目检验:prompt 注入的 skill 到底提升了工具调用程序,还是只是教会模型评测器要求的输出格式。

Wanyi Chen · Daoyuan Chen · Fang Kong
苏州大学 · 江苏重点实验室 · 阿里巴巴通义实验室Soochow University · Jiangsu Key Lab of Language Computing · Tongyi Lab

Heatmap summarizing standard and canonicalized function-calling gains

Figure from the paper: standard scores can move for reasons that canonicalized scoring exposes.

One score, two mechanisms.

A correct function call can still fail when its wrapper keys do not match the evaluator contract.

同一个分数可能混合了任务决策和接口合规两种机制。

Task policyChoose the right tool and arguments for the user request.根据用户请求选对工具并生成正确参数。
Interface alignmentEmit the JSON wrapper and key names expected by the scorer.输出评测器要求的 JSON 外层结构和 key 名。
01 · OutputModel call
02 · ParseStandard AST scorer
03 · ProjectCanonicalize aliases
04 · ControlsFormat-only & repaired
05 · ClaimTransfer or alignment?

What the diagnostics show.

The artifact is designed for auditability: each result is tied to a scorer, control, or portability check.

每个结论都对应一个评分器、控制实验或跨契约检查。

35 / 48

standard-score regressions become correct after canonicalized rescoring in the highlighted Qwen3-14B case.

+21.3 pp

format-only gain on Qwen3-8B, exceeding the +11.3 pp full-skill gain in the same cell.

0.5 pp

maximum gap between target-native and length-matched generic procedural prompts in the reported API-Bank check.

DiagnosticWhat it isolatesHow to read it
CanonicalizedWrapper-key aliasingSeparates outer-format failures from call structure; not a complete semantic executor.
Format-onlyExplicit interface contractIf it matches the full skill, the gain is not uniquely procedural.
Repaired / balancedInduction-pool sensitivityTests whether the result survives cleaner extraction and demonstration selection.

Small public method package.

The repository keeps the reusable scorer and hand-written controls public without bundling raw model outputs or generated skills.

仓库公开可复用的 scorer 和手写 controls,不捆绑原始模型输出或自动生成 skill。

Install

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Rescore

python src/rescore_canonical.py \
  --results-dir /path/to/results \
  --output /tmp/summary.json

The paper reports the final numerical matrices; this repository is the compact method release for scorer and control inspection.论文中给出最终数值矩阵;本仓库是用于检查 scorer 和 controls 的精简方法包。

Cite the paper.

Use the arXiv record for the current preprint.

当前预印本请使用 arXiv 记录。

@misc{chen2026attributingstructuredoutputgainsfunction,
  title         = {Attributing Structured-Output Gains in Function Calling: Interface Alignment versus Procedural Transfer},
  author        = {Wanyi Chen and Daoyuan Chen and Fang Kong},
  year          = {2026},
  eprint        = {2607.02595},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2607.02595}
}