waldiez/slm-callsite-eval

SLM Call-Site Evaluation Benchmark

The benchmark cases, scored model outputs and live end-to-end judgements for the paper

Not Every Call Needs a Frontier Model: Per-Call-Site Evaluation of Small Language Models in a Deployed Agentic Home-Automation System Panagiotis Kasnesis, Christos Chatzigeorgiou, Lazaros Toumanidis, Amalia Contiero Syropoulou. 1st Workshop on SLMs for Agentic Systems (SLM-Agents), NeurIPS 2026, Paris.

arXiv:2610.09021 · OpenReview · Code (GitHub) · Wactorz

An agentic system makes many different LLM calls, and they are not equally hard. This benchmark evaluates small language models separately at each of the five call sites of Wactorz, an open-source multi-agent home-automation framework, using its production prompts (v0.5.3). The cases are grounded in two live Home Assistant installations, a household and an office.

Call site	Config	Category	Cases	What the model must produce
Intent	intent	intent	36	the routing label for a user message, e.g. ACTUATE or PIPELINE
HA action	ha_action	ha	28	the Home Assistant action type, e.g. create_automation or other
Actuator	actuator	actuator	116	the service calls for a device request, or [] to refuse
Planner	planner	planner	50	a multi-agent plan over the available entities and topics
Codegen	codegen	dynamic	50	Python for a dynamic agent that meets an output contract
All	cases		280

Usage

import json
from datasets import load_dataset

cases = load_dataset("waldiez/slm-callsite-eval", "cases", split="test")
planner = load_dataset("waldiez/slm-callsite-eval", "planner", split="test")

case = planner[0]
key = json.loads(case["expected"])   # the answer key, always JSON-encoded

outputs = load_dataset("waldiez/slm-callsite-eval", "model_outputs", split="test")

Scoring is not a string match: each call site has its own checker (label match, action-set comparison, plan type and entity constraints, static checks on generated code). Use the scorers in harness/ to score new outputs comparably to the paper. To re-run a model with the exact production prompts, use harness/evalharness.py with wactorz==0.5.3.

Configs and fields

cases and the per-site configs

cases holds all 280 cases. intent, ha_action, actuator, planner and codegen hold the same rows, split by call site.

Field	Type	Description
id	string	case id; joins with case_id in model_outputs
category	string	call site: intent, ha, actuator, planner, dynamic (codegen)
subgroup	string	sub-group used in the paper's appendix breakdowns, e.g. a_control, refusal, reusetopic
source_file	string	file the case comes from in the GitHub repo
prompt	string or null	the user message passed to the call site. Grounded sites append the available Home Assistant entities, and actuator prompts are a JSON payload. Null when withheld (see Privacy)
prompt_withheld	bool	true for the 46 key-only actuator cases
expected	string	answer key, JSON-encoded: a label for intent and HA action, a list of service calls for actuator ([] means the request must be refused), and a constraint object or list of plan types for planner and codegen

model_outputs

2,520 records: 9 models × 280 cases, one pass at temperature 0.

Field	Description
model	provider:model. openai: marks local models served through an OpenAI-compatible endpoint, anthropic: hosted models
category, case_id	the case scored
rep, temperature	repetition index (always 0) and sampling temperature (0.0)
ok	the call returned without error
passed	the output passed the call site's checker
output	raw model output
input_tokens, output_tokens, latency_s, cost_usd	usage and cost (cost is 0 for local models)
error, ts	provider error (null when none) and Unix timestamp

e2e_eval

138 rows: 46 requests sent to the live system under three routing configs (all-hosted, gen-hosted, all-local), each judged by a person. Fields record the request (message, kind, group), the verdict (outcome, passed) and per-stage checks (intent_ok, device_ok, action_ok, planner_ok, code_ok, as y, n or n.a.), plus failure_mode, false_claim, note, and cost_usd where it was metered.

Results at a glance

Accuracy (%) per call site from model_outputs:

Model	Intent	HA action	Actuator	Planner	Codegen	All	Cost ($)
Qwen3.5 0.8B	50.0	71.4	33.6	8.0	24.0	33.2	0
Qwen3.5 2B	61.1	78.6	62.1	30.0	36.0	53.2	0
Qwen3.5 4B	91.7	92.9	50.9	52.0	38.0	58.2	0
Gemma4 E2B	83.3	89.3	56.0	46.0	34.0	57.1	0
Gemma4 E4B	91.7	96.4	81.9	50.0	80.0	78.6	0
Gemma4 12B	86.1	89.3	95.7	76.0	84.0	88.2	0
Gemma4 26B	91.7	96.4	94.8	88.0	80.0	90.7	0
Claude Sonnet 5	94.4	92.9	99.1	84.0	100.0	95.4	12.13
Claude Haiku 4.5	80.6	92.9	93.1	82.0	94.0	89.6	3.06

Rankings change from site to site and do not rise steadily with model size. Aggregate accuracy also hides two opposite failure modes on the actuator site: some small models act on devices that do not exist, while others refuse almost everything.

Privacy

The cases come from two live installations, so some material is withheld or redacted:

• Device registries are not included, because they carry network and location identifiers.
• 46 actuator cases (14 from site B and 32 cross-site) are included as keys only, with prompt set to null, because their device payloads describe a person's phone and the building's network. Every scoring table can be reproduced from the keys alone, but these 46 prompts cannot be re-run.
• Private addresses in prompts and outputs are replaced with the RFC 5737 documentation range. Recorded input_tokens count the original text, so re-tokenising a redacted prompt may give a different count by a token or two.

Limitations

Three intent keys and the planner entity block were revised after model outputs were inspected. The paper's Limitations section gives the reasons, and the generators in harness/make_bench_*.py document the changes. The revisions moved the hosted model by +0.08 and three local models by −0.03, which is against the direction of the paper's conclusion. Accuracy weights each case equally, not by how often each call site occurs in a real deployment.

Citation

@inproceedings{kasnesis2026notevery,
  title     = {Not Every Call Needs a Frontier Model: Per-Call-Site Evaluation
               of Small Language Models in a Deployed Agentic Home-Automation System},
  author    = {Kasnesis, Panagiotis and Chatzigeorgiou, Christos and
               Toumanidis, Lazaros and Contiero Syropoulou, Amalia},
  booktitle = {1st Workshop on SLMs for Agentic Systems (SLM-Agents), NeurIPS 2026},
  year      = {2026},
  eprint    = {2610.09021},
  archivePrefix = {arXiv},
  url       = {https://openreview.net/forum?id=H5z1pZCS4E}
}

License

Apache-2.0. Research supported by the NVIDIA Academic Grant Program, using two NVIDIA DGX Spark systems.