Multimodal large language models (MLLMs) have made rapid progress, yet their reasoning ability often lags behind strong text-only LLMs. Bridging this gap typically requires large-scale multimodal reasoning data or reinforcement learning, incurring substantial cost. An appealing alternative is parameter-space model merging between reasoning-enhanced LLMs and MLLMs, but we show that naive merging is fragile: its effectiveness varies widely across model families and can significantly degrade performance (e.g., for Qwen-based MLLMs). We propose Directional Reasoning Injection for Fine-Tuning (DRIFT), a lightweight method that transfers reasoning knowledge in the gradient space while preserving multimodal alignment. DRIFT precomputes a reasoning prior from the parameter differences between text-only reasoning experts and multimodal models, and uses it to bias gradients during supervised fine-tuning. This design retains the simplicity of standard SFT pipelines while enabling efficient and stable reasoning transfer. Experiments on multimodal reasoning benchmarks, including MathVista and MathVerse, show that DRIFT consistently outperforms naive merging and standard SFT, and matches or surpasses training-intensive methods with substantially lower data and compute.

“This paper introduces a novel crowdsourcing worker selection algorithm, enhancing annotationquality and reducing costs. Unlike previous studies targeting simpler tasks, this study con-tends with the complexities of label interdependencies in sequence labeling. The proposedalgorithm utilizes a Combinatorial Multi-Armed Bandit (CMAB) approach for worker selec-tion, and a cost-effective human feedback mechanism. The challenge of dealing with imbal-anced and small-scale datasets, which hinders offline simulation of worker selection, is tack-led using an innovative data augmentation method termed shifting, expanding, and shrink-ing (SES). Rigorous testing on CoNLL 2003 NER and Chinese OEI datasets showcased thealgorithm’s efficiency, with an increase in F1 score up to 100.04% of the expert-only base-line, alongside cost savings up to 65.97%. The paper also encompasses a dataset-independenttest emulating annotation evaluation through a Bernoulli distribution, which still led to animpressive 97.56% F1 score of the expert baseline and 59.88% cost savings. Furthermore,our approach can be seamlessly integrated into Reinforcement Learning from Human Feed-back (RLHF) systems, offering a cost-effective solution for obtaining human feedback. All re-sources, including source code and datasets, are available to the broader research community athttps://github.com/blcuicall/nlp-crowdsourcing.”