State-modulated low-light object detection using a continuous thought machine

Low illumination reduces signal-to-noise ratio, compresses contrast, and obscures object boundaries and textures, thereby weakening detector representations. Independent pixel enhancement can optimize perceptual quality rather than detection accuracy and introduce task-domain shift, whereas conventional feature enhancement is generally a memoryless feed-forward mapping. This study investigates state modulation for spatially nonuniform darkness and distinguishes the darkness guidance available during the first tick from detector-derived feedback available only after an initial prediction. CTM-YOLO inserts a lightweight feature enhancer between the early and deep stages of a YOLOv8n backbone. For a 640×640 input, three backbone features with 32, 64, and 128 channels are projected to a shared 128-channel space and aligned to an 80×80 grid. A learnable darkness estimator provides first-tick feature injection and spatial gating. Two structurally identical but separately parameterized convolutional branches form a subtractive excitation--inhibition feature. A 128-dimensional state is initialized from globally pooled aligned features, updated through gated context, a synapse multilayer perceptron, and residual layer normalization, and then mapped to feature-wise linear modulation parameters. Three recovery paths restore the modulated representation to the original backbone scales. Training uses one tick and a three-stage checkpoint-continuation procedure. Inference supports horizontal-flip test-time augmentation and an optional second tick that converts the first detection output into an uncertainty map. Experiments use the fixed 5, 890/1, 473 train/test split of ExDark and dataset-specific training on the visible subset of LLVIP and on DARK FACE. AP50 is computed by all-point interpolation with a confidence threshold of 0.001, a unified non-maximum-suppression intersection-over-union threshold of 0.6, and at most 300 detections per image. On ExDark, single-view, single-tick CTM-YOLO obtains an AP50 of 0.6963, 4.24 percentage points above the implemented Vanilla YOLOv8n control result of 0.6539. Horizontal-flip augmentation raises AP50 to 0.7115, 3.82 points above the Vanilla result with the same augmentation. Because the Vanilla run kept the neck frozen whereas the joint CTM stage optimized nearly all parameters, these differences compare the implemented configurations and are not interpreted as strictly controlled net enhancer gains. Dataset-specific training raises AP50 from 0.8348 to 0.8539 on LLVIP but reduces it from 0.5192 to 0.4705 on DARK FACE. A second zero-shot tick reaches 0.8619 on LLVIP but does not improve the ExDark single-view result. Three repeated CTM training runs yield 0.6604±0.0360 for single-view inference. The complete model contains 3.87 M parameters and has a measured single-step latency of 8.88 ms, or 112.7 frames/s, on an AMD Radeon RX 6750 GRE GPU at batch size 1. Under the reported implementations, CTM-YOLO achieves higher ExDark and LLVIP scores than the available controls while preserving real-time desktop-GPU throughput, but the improvement does not extend to DARK FACE and the benefit of a second tick is dataset dependent. Matched trainable-parameter controls, paired multi-seed retraining, and multiscale state alignment for very small targets remain necessary to establish broader applicability.