Large reasoning language models (LRLMs) achieve strong performance on complex reasoning tasks by generating long chain-of-thought trajectories, but such long-form reasoning also increases token consumption and latency. Existing inference-time methods that do not fine-tune the target LRLM optimize long reasoning through budgeting, early exit, uncertainty control, or reflection suppression, but they often treat reflection as a local signal rather than as a factor shaping the reasoning trajectory. In this paper, we identify unregulated reflection as a source of inefficient reasoning and characterize two failure modes: Trajectory Drift, where intermediate reflection opens weakly related branches that restart, reorient, or detour from the current path, and Verification Saturation, where answer-adjacent reflection repeatedly checks or restates an already-emerged provisional answer. To address these issues, we propose Regulated Reflection Decoding (RRD), an inference-time framework that requires no fine-tuning or parameter updates to the target LRLM and treats reflection events as online control points. RRD combines Trajectory Drift Suppression (TDS), which guides intermediate reasoning toward validated continuation branches, and Redundant Verification Suppression (RVS), which suppresses redundant verification when the provisional answer is sufficiently reliable. Experiments on multiple reasoning benchmarks show that RRD reduces reasoning length while preserving or improving accuracy, yielding a stronger accuracy-efficiency trade-off than direct inference and representative inference-time baselines.

Spatial relation reasoning is a crucial task for multimodal large language models (MLLMs) to understand the objective world. However, current benchmarks have issues like relying on bounding boxes, ignoring perspective substitutions, or allowing questions to be answered using only the model’s prior knowledge without image understanding. To address these issues, we introduce SpatialMQA, a human-annotated spatial relation reasoning benchmark based on COCO2017, which enables MLLMs to focus more on understanding images in the objective world. To ensure data quality, we design a well-tailored annotation procedure, resulting in SpatialMQA consisting of 5,392 samples. Based on this benchmark, a series of closed- and open-source MLLMs are implemented and the results indicate that the current state-of-the-art MLLM achieves only 48.14% accuracy, far below the human-level accuracy of 98.40%. Extensive experimental analyses are also conducted, suggesting the future research directions. The benchmark and codes are available at https://huggingface.co/datasets/liuziyan/SpatialMQA.