Datasets:
license: mit
language:
- en
pretty_name: Serving-Stack Confounds in Local Tool-Use Evaluation
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- tool-calling
- function-calling
- llm-evaluation
- llm-agents
- ollama
- vllm
- sglang
- llama-cpp
configs:
- config_name: serving_modes
default: true
data_files: serving_modes.csv
- config_name: chain_compare
data_files: chain_compare.csv
- config_name: crossstack
data_files: crossstack.csv
- config_name: humaneval_probe
data_files: humaneval_probe.csv
Serving-Stack Confounds in Local Tool-Use Evaluation
Result tables for Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation (Lijuan Tang, Yuemeng Zheng; REALM Workshop @ EMNLP 2026).
- Paper: ACL Anthology · arXiv:2609.26693
- Code, configs, and reproduction steps: github.com/LijuanTang94/serving-confound-repo
What this measures
Protocol fidelity = the rate of valid, in-schema tool calls over the turns the model
actually produced. Requests the server refused (for example Ollama's HTTP 400
does not support tools for Phi-3 and Gemma-3) are counted separately as rejected and
excluded from the denominator. A naive analysis that counts them as model non-calls
reports a capable model at 0% even though it never ran.
Serving conditions:
| condition | tools= sent |
text tool list in prompt |
|---|---|---|
native (default) |
yes | no |
native_hint |
yes | yes |
text_tools / tt |
no | yes |
All local models run on Ollama 0.30.8 with default (Q4_K_M) quantization;
deepseek-v4-flash is a cloud anchor. A "seed" indexes a task instance; decoding is
sampled at T=1.0.
Configs
serving_modes (default): aggregation task, 8 seeds, three conditions
For each condition prefix (native, native_hint, text_tools):
| column | meaning |
|---|---|
<cond>_seed |
per-seed mean fidelity |
<cond>_lo, <cond>_hi |
seed-level bootstrap 95% CI |
<cond>_pooled |
turn-pooled fidelity |
<cond>_n |
model-produced turns in the pool |
<cond>_rej |
rejected requests (excluded from the denominator) |
<cond>_n_seeds |
seeds with at least one produced turn |
Empty cells mean the condition was not run or the rate is undefined (every request was rejected).
chain_compare: dependency-chain task, 4 seeds
native_* and tt_* (text-tools) with _seed (per-seed mean), _pool (turn-pooled),
_n (turns), _rej (rejected requests).
crossstack: same tools= request on four serving stacks
One row per model; each cell is the outcome on that stack: 200_native (native
tool_calls), 200_text / 200_text_json_block (call returned as text),
200_runs_text (request accepted, model runs), 200_prose_nocall / 200_text_prose
(prose, no call), HTTP400_rejected (Ollama per-model gating), HTTP400_no_parser
(vLLM without --enable-auto-tool-choice --tool-call-parser), not_probed, n/a.
vllm_configured is vLLM with tool parsing enabled. The vLLM and SGLang probes cover
only the models shown.
humaneval_probe: single-turn replication on 6 HumanEval problems
native_gating is the Ollama response to tools= (- = not probed), then
text_tools_valid of n problems produced a valid call under the text-tools protocol.
Citation
@inproceedings{tang-zheng-2026-measuring,
title = "Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation",
author = "Tang, Lijuan and
Zheng, Yuemeng",
booktitle = "Proceedings of the Second Workshop for {RE}search on Agent Language Models ({REALM} 2026)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.realm-1.7/",
doi = "10.18653/2026.realm-1.7",
pages = "86--94"
}