NeFut Logo NeFut
Admin Login

[CS.AI] Not All Agreement Counts as Corroboration: Provenance-Conserving Multi-View Fusion for Typed Action Admission in Human-Robot Collaboration

Published at: 2026-09-04 22:00 Last updated: 2026-09-05 12:23
#algorithm #Machine Learning #Artificial Intelligence

For embodied systems, predictive agreement alone is insufficient to justify an action; the provenance of evidence matters equally. Repeated inference on the same observation can artificially inflate agreement without adding genuine evidence.\ \ PACT treats evidence countability as a relational variable. A supplied provenance partition defines countable units; within each unit PACT retains coordinatewise support and accumulates only across units.\ \ Formally, the coordinatewise meet is shown to be the greatest budget satisfying singleton fidelity and insertion non‑amplification, possessing coarsening monotonicity and fixed‑partition stability. Source‑local values cannot reveal countability.\ \ Across 31,200 evaluations in 48 scene clusters, PACT achieves a common‑support normalized risk‑coverage area (ncsAURC) of 0.0861. Removing the constructed adversarial‑consensus arm, provenance‑partition aggregation reduces ncsAURC by 0.0557 relative to singleton aggregation, and the corroboration contrast disappears. On complete‑source records, native scores favor PACT, but a common posterior‑peak score narrows its gap with nested Dirichlet and favors product fusion.\ \ In offline human‑robot collaboration, eightfold within‑camera duplication yields 720 typed responses per checkpoint without change. Camera‑grouped PACT admits 47 of 57 Qwen3‑VL‑32B reference‑consistent candidates, with no observed reference‑inconsistent admission over 60 episodes.\ \ Conclusion: PACT separates computational multiplicity from evidential multiplicity; agreement constitutes corroboration only when provenance permits separate accumulation.\ \ Review

Original Source: https://arxiv.org/abs/2609.01662

[h] Back to Home