acceptodds
Desk rejected at ICLR 2027

Dependency Hierarchy Governs Curvature Gains in Structured Latent Repair

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Dependency hierarchy conditions when curved geometry improves structured latent repair: branching traces provide metric expansion for selective updates, whereas flat connectivity offers little structure to exploit. We test this prediction with CurvRepair through a confound-closed audit that fixes graph source, support retrieval, locator, pullback, update norm, trigger, and selection rule while changing only the latent metric. On the strict released-only ProofWriter-OWA+GSM-Hard+MATH-500 endpoint (), CurvRepair gains percentage points over its edit-rule-matched Euclidean control (95% CI , paired-bootstrap ) while satisfying the prespecified pp harm tolerance. Matched rewiring separates the structural regimes: depth-preserving rewiring retains 85.2% of the whole-set gain, hierarchy destruction removes 68.4%, and random edges leave an 11.6% residual whose interval includes zero. The effect replicates across the complete 23-task BIG-Bench-Hard release (; pp), the complete MMLU-Pro release (; +1.46 pp, 95% CI [+1.19,+1.73]), and a fixed-anchor repair workflow (+0.94 pp). At equal end-to-end token budgets, CurvRepair leads Self-Consistency-Rewrite, Reflexion-Plus, and Constrained-Regeneration, with all three paired intervals above zero. Confound closure and structural intervention therefore turn hierarchy into an actionable criterion for choosing curved geometry in latent repair.

Primary area:
foundation or frontier models, including LLMs
Keywords:
adaptive reasoning and inference, Dependency Hierarchy Governs Curvature Gains in Structured Latent Repair

Related papers

Discussion (0)

Sign in to comment.