May 2026

There is a lack of a unifying computation model capable of connecting diverse forms of visual attention, such as spatial orientation and object binding. The primate ventral stream offers a bidirectional recurrent gating model that explains this behavior. The model is defined by two different core network components.

• Feedforward feature pathway ($X$): The bottom-up hierarchial processing sequence extracts visual features.
• Attention pathway ($Z$): The top-down pathway that combines task signals and lateral connections to generate an attention map.
• Multiplicative Gating ($X \times Z$): Before passing through a convolutional layer, the feed-forward feature map $X$ is multiplied by the attention map generated through $Z$.

A diverse set of attention and binding phenomena naturally emerge when error backpropogation is combined with these background constraints. When applied to synthetic (MNIST composites) and naturalistic (COCO) scenes, attention model improves contrast sensitivity, and produces perceptual-load effects.

Conceptually, the paper argues that binding does not require a separate symbolic mechanism where objects are stored. Instead, object coherence can arise when recurrent top-down and lateral signals selectively amplify features that belong together while suppressing irrelevant or competing features.

How does the brain represent actions to support flexible, compositional problem solving? Two macaques were trained to perform a drawing-like touchscreen task where the animals copied geometric figures without instructions on stroke trajectory and order. Over training, each monkey learned a set of reusable stroke primitives. The authors demonstrate that these primitives constitute true action symbols by testing them against three defining properties.

• Motor invariance: The trajectory generalizes across irrelevant motor parameters, such as drawing size or spatial location.
• Categorical structure: When confronted with morphing visual stimuli, the neural states switch discretely between two distinct learned primitives.
• Recombination: When drawing novel multi-stroke characters, the subjects assemble the complex drawings by reusing their own learned primitives.

These symbolic properties were obtained using simultaneous multiarea frontal-cortex recordings. A core conceptual insight is that symbolic structure does not require language; rather it can appear as distributed, reusable neural states.

Standard neuroscience protocols typically utilize minimal reward sizes to maximize the number of trials an animal can perform per session. However, this paper investigates how redistributing the same total reward volume into fewer, much larger portions impacts the efficiency of reinforcement learning in mice.

The authors demonstrate that overall learning efficiency is driven by a combination of three critical factors.

• Learning Rate: The speed of initial behavioral improvement within a single training session.
• Across-Session Capture: The animal's ability to retain and build upon optimal performance gains from prior sessions.
• Sustained Engagement: The ability to avoid abrupt, state-like drops in task performance, which the authors term "disengagement"

Compared to standard small rewards, introducing very large rewards dramatically accelerated task acquisition across across all three factors.

Measurements in the nucleus accumbens revealed that larger rewards generate more sustained dopamine response during the consumption of the reward. To test causality, the researchers used optogenetics to artificially sustain dopamine responses. While prolonged dopamine stimulation increased learning rate and decreased task engagement, it failed to replicate the across-session capture.

How does the brain decide whether to rely on just your most recent experience, or rather a long history of past outcomes? This paper explores how animals flexibly decide how far back in time past rewards should influence current decisions.

The first step was to see if mice could actually adapt their behavioral strategies to different environments (both “short-task” and “long-task” environments). Mice successfully adapted their decision-making strategies to match these respective environments. Importantly, motivation and basic engagement metrics remained similar across both tasks, indicating that the observed differences were driven by distinct temporal strategie.

To uncover where this flexible history integration is represented, the authors performed calcium imaging of several dorsal cortical areas, including the retrospenial cortex (RSC). The RSC exhibited the most robust task-dependent shift. By longitudinally tracking the exact same RSC neurons, the authors discovered that integration timecales are actively reorganized at the single-cell level. Furthermore, optogenetic inactivation of the RSC significantly reduced the mice's reliance on rewarded-choice history in both the short- and long-timescale tasks.