[Code scan] Scatter plot reports latency under the efficiency field

[njzjz]

This issue was found by a Codex global repository scan of tracked non-test files at commit 8c93925cb10b401b2b83c738bd9263fd74474468.

Relevant code

LAMBench/lambench/metrics/vishelper/plot_generation.py

Lines 35 to 49 in 8c93925

	zeroshot_raw = self.metrics_calculator.calculate_mean_m_bar_domain(model)
	if (
	efficiency_raw is None
	or efficiency_raw["average_time"] is None
	or zeroshot_raw is None
	):
	continue
	results.append(
	{
	"name": model.model_metadata.pretty_name,
	"family": model.model_family,
	"nparams": model.model_metadata.num_parameters,
	"efficiency": np.round(efficiency_raw["average_time"], 2),
	"std": np.round(efficiency_raw["standard_deviation"], 2),
	"generalizability error": np.round(zeroshot_raw, 2),

LAMBench/lambench/metrics/vishelper/metrics_calculations.py

Lines 200 to 216 in 8c93925

	def calculate_efficiency_results(self) -> dict[str, float]:
	efficiency_results = self.fetcher.fetch_inference_efficiency_results()
	# filter out models with missing efficiency results
	efficiency_results = {
	model: metrics
	for model, metrics in efficiency_results.items()
	if metrics is not None and metrics["average_time"] is not None
	}
	if not efficiency_results:
	logging.warning("No inference efficiency results found.")
	return {}
	# extract inference time and calculate efficiency score
	efficiency_results = {
	model: 100 / metrics["average_time"]
	for model, metrics in efficiency_results.items()
	}

LAMBench/lambench/metrics/results/README.md

Lines 109 to 111 in 8c93925

	We define an efficiency score, $M_E^m$, by normalizing the average inference time (with unit $\mathrm{\mu s/atom}$), $\bar \eta^m$, of a given LAM measured over 900 configurations with respect to an artificial reference value, thereby rescaling it to a range between zero and positive infinity. A larger value indicates higher efficiency.
	$$M_E^m = \frac{\eta^0 }{\bar \eta^m },\quad \eta^0= 100\ \mathrm{\mu s/atom}, \quad \bar \eta^m = \frac{1}{900}\sum_{i}^{900} \eta_{i}^{m}$$

Impact

The ranking code defines efficiency as 100 / average_time, and the README describes the same inverse-latency score. generate_scatter_plot() stores efficiency_raw["average_time"] directly under the JSON field named efficiency.

This makes the scatter data use the opposite direction from the ranking score: larger values mean slower models in the scatter JSON but better models in the ranking metric.

Suggested fix

Either emit 100 / average_time for the scatter efficiency field, or rename the field to latency and update downstream visualization labels and docs accordingly.