Neural architecture search (NAS) is a core technique for automating AI model design, yet its efficiency is limited by the high cost of reliable architecture evaluation. To address this bottleneck, we introduce code-semantic transition modeling for NAS acceleration: before spending evaluator budget, we ask whether a proposal’s semantic change from its parent suggests a worthwhile architecture evolution. We instantiate this idea with ECO-SCE, an Evolutionary Code Semantic Change Estimator placed between the architecture generator and evaluator. Given a parent-to-proposal code transition, ECO-SCE filters likely low-gain proposals and reserves evaluator calls for proposals that are more likely to provide useful evolutionary progress. To train ECO-SCE, we construct NB101-SemPair from NAS-Bench-101 by sampling beneficial, harmful, and marginal transitions with diverse semantic and structural changes. ECO-SCE learns risk-aware transition utility from global semantic shifts, sparse semantic overlaps, and explicit structural edits, allowing it to skip likely low-gain proposals while sending promising or uncertain transitions to the evaluator. Across seven downstream benchmarks including CLRS, ECO-SCE reduces average GPU time by 70.88% and improves task performance in 6 out of 7 cases when inserted zero-shot into OpenEvolve.

Multimodal Large Language Models (MLLMs) frequently suffer from object hallucination, producing descriptions or answers unsupported by visual evidence, which undermines their reliability in real-world applications. This paper begins with an intriguing phenomenon: prefacing queries with statements explicitly contradicting image content often yields more accurate responses than direct questioning. We term this the Contradictory Prompting Effect (CPE). Priming inputs with assertions opposing reality triggers an intrinsic self-verification mechanism within MLLMs, where conflicting claims compel the model to prioritize visual evidence over linguistic priors. Inspired by this discovery, we propose Contradiction-Guided Visual Reasoning (CGVR), a novel training-free framework for hallucination mitigation. CGVR systematically generates deliberate contradictory captions alongside standard ones. These captions drive a dual-channel inference process, in which conflicting signals enforce rigorous cross-checking against visual evidence. Our work reveals MLLMs’ latent self-verification capability triggered by contradictions and establishes a paradigm for lightweight, reliable multimodal reasoning. Extensive evaluations on POPE and MME benchmarks demonstrate that our proposed CGVR significantly mitigates hallucinations across diverse MLLMs. Our code is available at https://github.com/dhr1698/CGVR.