NSDL - LOCUS - ID: leaderboard

NSDL - LOCUS - ID — NSDL v1: VIKI reported/transcribed baselines and LOCUS typed robot-planning protocols. VIKI end-to-end(rep.)rows are transcribed from reference2, not rerun by this study; LOCUS VIKI-R rows use that study weights under the named native/op-level protocols. Scores are fractions of executable,goal-satisfying plans under the task length bound; not judgment accuracy or rejection rate. LOCUS Mean is the equal-weight mean of ID/OOD-S/OOD-M/OOD-X, not success pooled over all test instances. Whole trained-judgment/detector/compiler/planner configurations retain exact native training labels, including SFT/RL and Phi compiler ablations. The trained checkpoint components are real trained artifacts; these end-to-end scores belong to the composite system, not a new alias for the unchanged backbone. Reference GT sketch and section headings are not evaluated models. VIKI op-level/native protocols stay distinct. Metric: Task success (fraction, 0–1).

Metric: Task success (fraction, 0–1). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 1 model tracked. Updated 15 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1NSDL(NSDL run: native=NSDL 7B (SFT + RL) = NSDL-RL, judgment=model)0.839

Interactive version: theaggregate.ai/benchmark?slug=nsdl-locus-id · How It Works · Data refreshed daily, snapshot 2026-11-16.