acceptodds
Withdrawn from ICLR 2027

Rule-Conditioned Attribute Comparisons for Visual Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual rules can preserve a collection of attribute values across panels even when objects within each panel disagree. With exact perception, the appearance consensus summary maps every heterogeneous attribute inventory to one inconsistency state and can erase distinctions between answers. We introduce a rule-conditioned comparison layer for fixed probabilistic reasoners. Under eligible Constant rules, occupancy-weighted histograms retain the values that should persist. Under scalar appearance rules, conditioning candidate beliefs on ordinary values changes both their inconsistency penalty and categorical weighting. Perception, inferred rules and other score terms remain matched. On 2,200 fresh high-count I-RAVEN questions in a nine-slot layout, inventory comparison raises PrAE accuracy from 94.50% to 99.68% and NVSA accuracy from 93.77% to 99.50% with shared perception and prespecified low-count guardrails met. Reusing this cohort with three published visual checkpoints retains average gains of 4.74 and 4.23 points. A separately frozen 3,300-question low-count study confirms a 0.45-point conditioning gain. Exact information-loss analysis and matched controls identify retained values as the main inventory benefit, while crossed scalar comparisons separate the effects of candidate shape and consistency mass.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
visual abstract reasoning, probabilistic inference, object-centric representations, distribution comparison

Related papers

Discussion (0)

Sign in to comment.