Research map / Large language models / Reasoning and chain of thought

Consistency-based Abductive Reasoning over Perceptual Errors of Multiple Pre-trained Models in Novel Environments

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes test-time consistency-based abduction over error-detection rules from multiple pre-trained object detectors, solved exactly via integer programming or greedily via heuristic search, evaluated on an extended simulated aerial MDS-A dataset under weather shifts.

Problem

Pre-trained perception models degrade under distribution shift in novel environments; existing metacognitive error-detection rules improve precision at the cost of recall, and abductive learning only operates at training time.

Approach

Encode per-model error-detection rules and predictions in a logic program, then at test time select a subset of model predictions that maximizes assigned predictions subject to an inconsistency budget on domain constraints, solved via IP or a greedy heuristic with a confidence tie-breaker.

Key results

• IP+TB achieves average relative improvements of about 13.6% in F1 and 16.6% in accuracy over the best individual model across 15 test sets (per abstract).
• On AM_1, IP+TB reaches F1 0.21 vs 0.18 for best individual model and 0.05 for majority vote.
• Removing the tie-breaker drops HS F1 by 10% to over 17% on various datasets, while IP is unchanged.

Surprising findings

• Removing the tie-breaker has zero effect on IP results across all metrics and datasets.
• Majority vote performs worse than the best individual model on most test sets, sometimes far worse (e.g., F1 0.05 on AM_1).

Limitations

• Evaluated only on a single simulated aerial dataset with six detectors.
• IP worst-case NP-hard; runtime grows with objects and models; heuristic is suboptimal in principle.
• Rules rely on unique name assumption and IoU matching (>90%) that may not hold generally.

Experimental setup

Datasets: MDS-A (Multiple Distribution Shift - Aerial, AirSim-generated)

Benchmarks: MDS-A extended test suite (15 test sets)

Metrics: F1-score accuracy execution time

Baselines: Best individual model Average of models Majority vote Tie-breaker only (TB)

Builds on: abductive learning error detection and correction rules (EDCR) PyReason DETR

Closest papers by meaning

• Ambiguity-Aware Abductive Learning (ICML 2024 · similarity 0.88)
• AbductiveMLLM: Boosting Visual Abductive Reasoning Within MLLMs (AAAI 2026 · similarity 0.88)
• Safe Abductive Learning in the Presence of Inaccurate Rules (AAAI 2024 · similarity 0.86)
• Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events (CVPR 2025 · similarity 0.84)
• NL-Eye: Abductive NLI For Images (ICLR 2025 · similarity 0.84)
• Analysis for Abductive Learning and Neural-Symbolic Reasoning Shortcuts (ICML 2024 · similarity 0.84)