Existing pedestrian navigation systems often provide direction-only instructions with limited visual confirmation at complex intersections or along long straight paths. We study coordinate-free landmark extraction from dashcam image sequences. The model returns a set of landmarks with structured attributes (e.g., category and description), or abstains when visual evidence is insufficient. Although general MLLMs can recognize street-scene objects, they do not reliably follow expert rules for filtering ambiguous candidates. Moreover, a false landmark can mislead a pedestrian, making precision more important than exhaustive coverage in this task. We therefore develop a progressive alignment framework, supported by a landmark annotation and augmentation pipeline that converts navigation experts’ selection criteria into staged supervision signals. We further propose a hard-gated bipartite matcher for box-free evaluation and a navigation-aware risk-sensitive reward that penalizes false positives more heavily than abstentions without inducing blanket silence. On the held-out benchmark, our model achieves 92.29% precision while maintaining practical coverage, outperforming strong MLLM baselines. A 30-day A/B test with 2M users on the Amap platform further shows that landmark-augmented instructions reduce departure-confirmation latency by 23.5%. Together, these results demonstrate that risk-aware alignment enables reliable, abstention-capable landmark extraction for navigation.

Multi-agent LLM systems can outperform single-agent baselines, yet practitioners still cannot predict which design works for a new task or diagnose why one fails. We argue this gap persists largely because the field lacks a diagnostic framework with measurable primitives and testable predictions. We introduce DIANOIA, a three-channel decomposition of multi-agent reasoning gain into coverage, fidelity, and synthesis, each of which is empirically measurable. From this decomposition, we propose a diagnostic protocol that forms qualitative hypotheses about bottleneck channels for tasks with measurable quality signals. We instantiate the protocol as a multi-agent system whose three components mirror the channels: role-diverse proposers for coverage, execution- or pseudo-verification-grounded review for fidelity, and iterative synthesis. On GSM8K, MBPP, and BFCL-SP, our method outperforms strong multi-agent baselines under matched standard configurations; on MBPP, it matches the strongest baseline with approximately 5× fewer tokens and dominates the measured accuracy–token Pareto frontier. Full-scale MBPP replications on Qwen, DeepSeek, and Gemini consistently improve over the corresponding single-model baseline. A prospective held-out evaluation on VitaBench, whose bottleneck hypothesis was recorded before running DIANOIA, improves pass rate from 29.5% to 45.0%. DIANOIA reframes multi-agent design as channel-aware resource allocation: diagnose the likely bottleneck, then invest tokens accordingly.