Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation
Paper • 2609.26693 • Published
Datasets:
model stringclasses 10
values | native_seed float64 0 1 ⌀ | native_lo float64 0 1 ⌀ | native_hi float64 0 1 ⌀ | native_pooled float64 0 1 | native_n int64 0 519 | native_rej int64 0 8 | native_n_seeds int64 0 8 | native_hint_seed float64 0.35 0.89 ⌀ | native_hint_lo float64 0.07 0.78 ⌀ | native_hint_hi float64 0.66 0.98 ⌀ | native_hint_pooled float64 0.29 0.96 ⌀ | native_hint_n float64 35 317 ⌀ | native_hint_rej float64 0 3 ⌀ | native_hint_n_seeds float64 8 8 ⌀ | text_tools_seed float64 0.34 1 | text_tools_lo float64 0.09 1 | text_tools_hi float64 0.5 1 | text_tools_pooled float64 0.3 1 | text_tools_n int64 8 110 | text_tools_rej int64 0 8 | text_tools_n_seeds int64 8 8 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
qwen2.5-coder:0.5b | 0 | 0 | 0 | 0 | 9 | 0 | 8 | 0.346 | 0.067 | 0.659 | 0.287 | 317 | 2 | 8 | 0.344 | 0.094 | 0.594 | 0.852 | 54 | 1 | 8 |
qwen2.5-coder:1.5b | 0.375 | 0.125 | 0.75 | 0.237 | 519 | 3 | 8 | 0.583 | 0.312 | 0.833 | 0.906 | 53 | 1 | 8 | 0.742 | 0.492 | 0.975 | 0.955 | 110 | 1 | 8 |
qwen2.5-coder:3b | 0.228 | 0 | 0.549 | 0.786 | 42 | 0 | 8 | 0.823 | 0.575 | 0.979 | 0.96 | 176 | 3 | 8 | 0.837 | 0.702 | 0.958 | 0.896 | 77 | 1 | 8 |
qwen2.5-coder:7b | 0.589 | 0.319 | 0.837 | 0.793 | 174 | 2 | 8 | 0.885 | 0.781 | 0.969 | 0.8 | 35 | 0 | 8 | 0.912 | 0.825 | 0.979 | 0.889 | 36 | 0 | 8 |
qwen2.5-coder:14b | 0.595 | 0.383 | 0.772 | 0.737 | 57 | 0 | 8 | 0.796 | 0.625 | 0.944 | 0.708 | 89 | 0 | 8 | 0.915 | 0.856 | 0.979 | 0.882 | 34 | 0 | 8 |
llama3.2:latest | 0.6 | 0.375 | 0.8 | 0.625 | 16 | 0 | 8 | 0.823 | 0.709 | 0.921 | 0.892 | 65 | 0 | 8 | 0.438 | 0.312 | 0.5 | 0.467 | 15 | 0 | 8 |
phi3:latest | null | null | null | 0 | 0 | 8 | 0 | null | null | null | null | null | null | null | 0.375 | 0.125 | 0.75 | 0.3 | 10 | 0 | 8 |
gemma3:4b | null | null | null | 0 | 0 | 8 | 0 | null | null | null | null | null | null | null | 1 | 1 | 1 | 1 | 8 | 8 | 8 |
gemma3:270m | null | null | null | 0 | 0 | 8 | 0 | null | null | null | null | null | null | null | 0.375 | 0.125 | 0.75 | 0.375 | 8 | 4 | 8 |
deepseek-v4-flash | 1 | 1 | 1 | 1 | 61 | 0 | 8 | null | null | null | null | null | null | null | 1 | 1 | 1 | 1 | 57 | 0 | 8 |
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"
}
- Downloads last month
- 50