Answers are grounded in this paper and cite their sources.
A large language model (LLM)-based agent uses text observations to find UI elements. Their content affects token cost and grounding success. Observation-reduction work asks how much to remove; we ask which attribute subsets are sufficient under a fixed success predicate. We compare 6000 events across 24 synthetic pages, varying attributes and serialisations. Removing volatile class and id values increases success by 9.48 pp (95% CI [+4.90, +15.52] pp). The secondary metric P_now (unique correct match) changes little where a stable attribute exists but decreases otherwise. The improvement comes from removing unstable locator candidates, not adding information. Abbreviating attribute names reduces success by 12.08 pp, saving 0.2% of tokens. Removing JSON syntax while keeping attribute names reduces cost without a significant change in success. The cost-optimal profile (COP), best among 21 main configurations, achieves 10.00 pp higher success with 34.0% fewer tokens than full-attribute JSON. In a one-attempt comparison, the accessibility snapshot uses half the tokens but performs worse; neither dominates. The direction persists across three gpt-4.1 generations and four models under non-identical decoding settings (for o4-mini, via the COP). A 33-site archived-page pilot shows a smaller gain in that direction. Results concern predicate-relative sufficiency in HTML DOM with CSS locators.
No takes yet. Share an insight, caveat, or question.
Angelov et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: