Feed reasoning confidence into the planner coupling - #193
Conversation
…undation#110). Scale the zero-init reasoning residual by pooled confidence so low confidence yields more conservative modulation, preserve no-op at init, and add ECE. Co-authored-by: Cursor <cursoragent@cursor.com>
ShauryaVM
changed the title
|
Thank you for your contribution. |
|
Under the contribution standard proposed in #198, I consider this PR AI slop in its current form. It changes learned model behavior but provides no training comparison, ADE/FDE results, or analysis of model performance. This is not ready for review; please provide the required real-world validation before resubmitting. |
Compare scaled (autowarefoundation#110) vs unscaled (confidence=1) on the same batch and report ADE@3s. Twelve-step IL shows no regression while the zero-init gate is still closed. Co-authored-by: Cursor <cursoragent@cursor.com>
|
A/B numbers for the confidence→planner change. Trained two mock Combined-style models for 12 steps on the same batch (CPU): scaled (this PR) vs unscaled (confidence forced to 1, i.e. pre-#110). JSON:
Forcing conf=0 vs conf=1 on the scaled model after those steps also moves ADE by ~0. That is the containment property: I am not claiming a KITScenes ADE win. The A/B says this change does not regress open-loop ADE while the coupling is still a no-op, which is the bar for landing the safety-loop wiring. |
2 participants
Follow-up to #110 / the #103 safety loop.
The reasoning band already emits confidence (and we supervise it with Brier), but the planner coupling ignored it. This scales the zero-init residual by pooled confidence so low confidence means weaker reasoning modulation — i.e. stay closer to the reactive baseline. At init
alphais still 0, so any confidence value remains a no-op until the gate opens.Also wires confidence through
ReactiveE2E→ Bezier / FlowMatching, and adds a small ECE helper for eval.How I tested
pytest Model/tests/test_reasoning_coupling.pyCloses #110