A Decade of Thalamus: Where We Are in 2025

What a decade of science taught us about how the brain organizes thought

Earlier this week, I highlighted two sister papers that were published in Nature Communications. One study involved the main thesis project of my Ph.D. student Brabeeba Wang, who is an incredible theoretician. The other led by a newly minted assistant professor, Bin Wang, who just started his experimental lab. The papers continue a scientific journey that started over a decade ago and focused on a simple question: why is the brain wired the way it is?

The brain is an extraordinarily complex machine, but it’s best known for its outer folded layer: the cerebral cortex. In most neuroscience classes, the cortex is treated as synonymous with the brain itself. Yet, every neuroscience student knows that the cortex is deeply interconnected with a central, egg-shaped structure called the thalamus. But why is that so? Perhaps the tractable questions are: why is the forebrain organized into thalamocortical loops and what does the thalamus actually do?

The conventional answer is the following: The thalamus is a relay station, passing sensory information from the periphery to cortex. The sensory pathways from the retina, cochlea, and skin all connect in thalamic nuclei before reaching primary sensory cortex. The lateral geniculate nucleus relays vision, the medial geniculate relays audition, the ventral posterior nucleus relays touch. Each has orderly topographic maps that preserve spatial information. The function appears obvious: get sensory signals to the cortex, where the real computation happens. In fact, the following figure from Hubel and Wiesel’s work in vision (which won the Nobel Prize) illustrates exactly this point.

Figure from Hubel and Wiesel’s work, LGN (lateral geniculate nucleus) neurons encode visual inputs in contrast patches (center-surround). V1 (primary visual cortex) neurons sum LGN inputs to construct oriented edges, which are then used to construct simple visual features and ultimately visual form in higher visual areas.

This relay framework dominated thinking about thalamus for most of the twentieth century. Come to think of it now, this is puzzling since most of the thalamus (particularly in primates) isn’t even connected to sensory organs. For example, our favorite area, the mediodorsal (MD) thalamus which connects reciprocally with the prefrontal cortex (PFC) was understood through this lens (Mitchell, 2015, “The mediodorsal thalamus as a higher order thalamic relay nucleus important for learning and decision-making,” Neurosci. Biobehav. Rev.).

Early hints that this picture was incomplete existed but failed to amount to a coherent framework. Thalamic lesions in humans produced cognitive deficits that seemed inconsistent with the relay picture and some recordings seemed hard to reconcile with that framework, too. Yet, the relay model held because it was simple, anatomically sound, and mostly consistent with available data.

What emerged over the past decade was different. Rather than relaying information, the thalamus performs computations that shape how cortex processes information. It compresses high-dimensional cortical activity into low-dimensional contextual states. It decomposes uncertainty into separable components. It enables rapid reconfiguration of cortical networks to match changing task demands. The story of how we got here, from mice learning attention tasks to computational models of human reasoning, involved contributions from multiple labs, species, and methodological approaches. This is where we are in 2025.

2015: Building the Task

The story began earlier but was ultimately published in Ralf Wimmer’s 2015 paper (Wimmer et al., 2015). Ralf and I were inspired by work on attention and how it was flexibly deployed (e.g. the cocktail party problem). We wanted to establish a paradigm that would allow for dissecting this process at the level of its circuit implementation. The mouse was the perfect model as causal tools (such as optogenetics) had just taken off and offered a way to directly test whether neural codes were actually driving behavior.

But there were many challenges; we weren’t sure that mice could solve problems in a setting akin to the cocktail party problem. Even if they did, we weren’t sure we’d be able to have the right design that would dissociate the multiple cognitive operations required for the task more generally and isolate attentional control as the object of study. People in the primate world were adept at that, so what we really needed to do is try to make a mouse behave like a monkey. Work from Bob Wurtz’s lab had developed a really nice set of attentional control tasks that we felt could be a good starting point (e.g. McAlonan et al., 2009 Nature).

The paradigm that we developed required mice to flexibly switch attention between visual and auditory target stimuli on a trial-by-trial basis. On each trial, a mouse received an instruction cue indicating which modality to attend. After a delay period requiring holding the instruction in mind (akin to working memory), the animal was presented with simultaneous visual and auditory stimuli that, on many trials, were spatially conflicting. Correct performance required maintaining the rule through the delay, applying it at the moment of choice, and ignoring the irrelevant modality despite its salience.

The trial structure provided a behavioral state clamp: defined epochs where we could interpret neural activity in relation to specific cognitive operations rather than a mixture of sensory, motor, and decision variables. The instruction period isolated rule encoding. The delay period isolated rule maintenance. The stimulus period combined sensory processing with rule application.

This was the first demonstration that mice could express the kind of flexible, rule-based behavior that allows mechanistic insight into executive function. The achievement was not just showing that mice could learn the task, but that they performed it in a way that engaged the same prefrontal mechanisms implicated in primate studies of cognitive control.

Using optogenetics, we could inactivate prefrontal cortex during specific task phases. Inactivation during the instruction or delay period impaired performance down to chance level. Inactivation during target stimulus delivery made no difference. Prefrontal cortex was necessary for rule encoding and maintenance but not for attentional selection itself. This temporal specificity meant we could go beyond whether a brain region was involved, but precisely ask when and how it contributed to a cognitive operation.

Figure 1 of Wimmer et al. 2015

2017: The Surprising Discovery

With the task established, we could ask how prefrontal cortex and mediodorsal thalamus interact during attentional control (Schmitt et al., 2017). Decades of anatomical work had established dense reciprocal connectivity between these regions. Human and primate lesion studies implicated both in cognitive flexibility. But the computational relationship remained unclear.

Prefrontal cortex represented task rules through high-dimensional neural sequences. These patterns were stable across trials with the same rule but distinct between rules. Many neurons were required to encode each sequence, creating a distributed code across the task-relevant prefrontal population.

The mediodorsal thalamus also represented task information, but differently. Its neural representation was lower-dimensional, suggesting a more compressed code. A single MD neuron responded similarly to whether the animal was attending to vision or audition. What’s going on, we wondered.

Optogenetic experiments confirmed that the source of these representations was the prefrontal cortex and that the flow of information starts cortically, not from the thalamus. The thalamus was not relaying cue information to cortex as silencing it did not prevent cortex from receiving information. Instead, mediodorsal input amplified local prefrontal connectivity. This amplification allowed rule-specific neural sequences to emerge and stabilize within prefrontal circuits. Without thalamic input, prefrontal neurons still responded to task events, but they could not organize into the coherent spatiotemporal patterns that constituted stable rule representations.

The thalamus functioned as a temporal scaffold for working memory, not as a source of the content being remembered. It was not delivering information in the way sensory thalamus delivers visual or auditory signals. Instead, it was enabling cortex to construct and maintain its own representations by modulating circuit dynamics.

Figure 1 of Schmitt et al. 2017

2018: Context and Control

The 2017 findings raised as many questions as they answered. We knew the mediodorsal thalamus amplified prefrontal connectivity and compressed cortical representations. But what was the computational logic? And what role did this play in cognitive flexibility, which both human and animal work consistently implicated in thalamocortical interactions?

We next developed a more complex task where mice had to track not just which modality to attend, but which set of learned cues was currently active (Rikhye et al., 2018). This hierarchical structure allowed us to dissect how the two regions represent information at different levels of abstraction.

Prefrontal cortex encoded both individual cues and their meaning as task rules. Neurons responded to different cues that signaled “attend to vision,” abstracting across low-level sensory features in a way we weren’t certain mice could manage.

Mediodorsal thalamus encoded neither the cues nor the rules. Instead its neurons robustly signaled which context the animal was in, which game it was playing. The context could be decoded from prefrontal cortex, but required distributed patterns across many neurons. In MD thalamus, a single neuron was on in one context and off in another. The thalamus was compressing along the dimensions of task context.

Recording from hundreds of thalamic neurons revealed two distinct functional populations. One activated task-relevant prefrontal populations to sustain their patterns. The other suppressed prefrontal populations when the context switched to prevent interference. Suppressing the second population prevented task switching, consistent with human and primate evidence for mediodorsal thalamus involvement in cognitive flexibility. We now had mechanistic specificity: two populations with distinct dynamics implementing complementary functions, one for stability and one for flexibility.

Figure 5 of Rikhye et al., 2018

2021: Genetic and Functional Specificity

Were these two populations transient functional states or distinct cell types with different anatomical and genetic identities? The question mattered because it would tell us whether these computational roles reflected circuit architecture itself.

Arghya Mukherjee combined anatomical tracing with genetic line screening to show that the two functional cell types were genetically and anatomically distinct (Mukherjee et al., 2021). One expressed dopamine D2 receptors, the major target of traditional antipsychotic medications. The other expressed GRIK4, a kainate receptor. The first activated prefrontal cortex through disinhibition. The second suppressed it by innervating parvalbumin-positive interneurons, another schizophrenia-relevant circuit component.

The two projections solved different computational problems related to uncertainty. The suppressive projection activated when incoming information was conflicting, engaging prefrontal inhibition proportional to the amount of conflict. The activating projection engaged when incoming information was sparse, boosting prefrontal activity to amplify weak signals.

This finding provided a mechanistic entry point for understanding decision-making abnormalities in psychiatric disorders. We followed up with the same task in patients and healthy controls, identifying a putative biomarker in schizophrenia (Huang et al., 2024). This was the first direct connection between the circuit mechanisms we’d identified in animals and the cognitive phenotypes we see clinically.

The work also demonstrated that even within mediodorsal thalamus, and even within populations projecting to the same cortical area, distinct cell types implement distinct computations.

Graphical abstract from Huang et al., 2024, based on task developed in Mukherjee et al. 2021

Early 2025: Communication and Efficiency

A critical question remained: how much of this framework applied to complex decision making that humans engage in?

Tree shrews (Tupaia belangeri) offered the right model. They’re basal primates with more differentiated prefrontal cortex than mice, including granular layer 4 that receives thalamic input in a pattern similar to primates.

We developed a hierarchical decision task where tree shrews tracked two forms of uncertainty: cueing uncertainty (is the instruction signal clear or ambiguous?) and rule uncertainty (have the rules changed?) (Lam et al., 2025). These are computationally distinct and could engage different neural mechanisms. They are also closer to real-world complexity that we constantly face.

The thalamus independently represented both types of uncertainty. Different thalamic populations tracked cueing versus rule uncertainty, and these representations were more separated than in prefrontal cortex where the two forms were more mixed. This suggested the thalamus was decomposing mixed cortical activity into separable uncertainty estimates that could then regulate cortical processing.

During rule reversals, animals had to infer whether errors resulted from noisy cues (stick with the current rule) or rule changes (update the rule). This is a classic credit assignment problem. Mediodorsal thalamus tracked this inference process. The population representing rule uncertainty increased activity specifically when errors were likely due to rule changes, not when they reflected cue noise.

The paper also showed that this population receives error signals from cingulate cortex and mediates efficient transthalamic communication across prefrontal areas. The finding was consistent with Murray Sherman’s proposed transthalamic route for cortico-cortical communication, but added computational efficiency through low-dimensional transmission as a potential driver.

The conservation across species separated by over 70 million years of evolution strengthened the argument. Context compression, uncertainty decomposition, and transthalamic communication appeared to be fundamental organizing principles of mammalian thalamocortical systems, not rodent-specific adaptations.

A Tupaia subject playing in cage. Extremely cute animal.

Late 2025: Back to Humans and Theory

Which brings us to this week’s Nature Communications papers. Papers that continue to extend the evolving framework into human cognition and formalize it into a coherent computational theory.

Brabeeba Wang’s paper tackles the theoretical integration (Wang et al., 2025). The challenge was to build a model that could simultaneously explain data from multiple experiments using realistic neural architectures while remaining interpretable at the computational level. The paper provides insight into how distributional reinforcement learning architecture contributes to directed exploration in uncertain environments, how the prefrontal-thalamic network segments ongoing experience into discrete contexts allowing for savings and generalization, and how changes in these processes lead to aberrant beliefs in schizophrenia.

The companion paper, led by Bin Wang in collaboration with Burkhard Pleger and colleagues in Germany and China, tested the model’s predictions in humans (Wang et al., 2025). Volunteers performed a probabilistic reversal learning task during fMRI scanning. The task required learning stimulus-response associations through trial and error, then adapting when the rules reversed unpredictably.

Subjects showed natural variation in performance that was well-fit by model-free and model-based reinforcement learning strategies. Model-free learning caches action values based on experienced outcomes. Model-based learning builds internal models of task structure and uses them to plan. The key question was: what arbitrates between these strategies?

The fMRI results were striking. Activity in mediodorsal thalamus tracked strategy use. When subjects relied predominantly on model-based strategies, an executive prefrontal-MD network was active. When subjects used more model-free strategies, an evaluative prefrontal-MD network was active.

During behavioral transitions, interactions between prefrontal areas appeared to require the thalamus. This pattern matched the model prediction: the thalamus enables rapid cortical reconfiguration by providing a pathway for compressed communication. The finding paralleled what we observed in tree shrews, now confirmed in humans.

Looking Back, Looking Forward

From mice switching attention between conflicting sensory cues in 2015 to computational models explaining human reasoning in 2025, the thalamus transformed from a mysterious relay structure into a central player in how the brain builds and revises internal models of the world.

The work suggests a different way of thinking about brain architecture. Information flow matters, but so does information geometry. The brain does not simply pass signals around. It actively reshapes representational spaces to make different computations possible at different times. The thalamus appears central to that process: compressing high-dimensional cortical activity along task-relevant dimensions, decomposing uncertainty into independent components, and enabling efficient communication across distributed cortical networks.

The path over the last decade involved collaboration across multiple labs, species, and methodological approaches. It required combining animal physiology, human neuroimaging, computational modeling, and neural theory into a single coherent framework. None of this would have been possible without talented trainees, supportive collaborators, and sustained funding that allowed us to pursue questions that took years to answer.

The work happening in the lab right now goes beyond anything I imagined in 2015. We’re applying these principles to understand how human cognition is generally implemented and how thalamic computations play specific roles. This also has implications for psychiatry with the goal of moving towards mechanism-based diagnosis and treatment. How far we get remains to be seen, but the framework is in place and I am optimistic. And… you’ll have to stay tuned to learn more!

Acknowledgements

This work would not have been possible without the contributions of past and present Halassa lab members, and collaborators including Sabine Kastner, Murray Sherman, Marty Usrey, Matt Nassar, Guoping Feng, Sage Chen, Kai Hwang, Francisco Clasca, and Burkhard Pleger. Special thanks to Matt Wilson for shaping how I think about science (another topic altogether for another post).

Bibliography

Huang, A.S., Wimmer, R.D., Lam, N.H., Wang, B.A., Suresh, S., Roeske, M.J., Pleger, B., Halassa, M.M., & Woodward, N.D. (2024). A prefrontal thalamocortical readout for conflict-related executive dysfunction in schizophrenia. Cell Reports Medicine, 5(11), 101802.

Lam, N.H., Mukherjee, A., Wimmer, R.D., Nassar, M.R., Chen, Z.S., & Halassa, M.M. (2025). Prefrontal transthalamic uncertainty processing drives flexible switching. Nature, 637(8044), 127-136.

McAlonan, K., Cavanaugh, J., & Wurtz, R.H. (2008). Guarding the gateway to cortex with attention in visual thalamus. Nature, 456(7220), 391-394.

Mitchell, A.S. (2015). The mediodorsal thalamus as a higher order thalamic relay nucleus important for learning and decision-making. Neuroscience & Biobehavioral Reviews, 54, 76-88.

Mukherjee, A., Lam, N.H., Wimmer, R.D., & Halassa, M.M. (2021). Thalamic circuits for independent control of prefrontal signal and noise. Nature, 600(7887), 100-104.

Rikhye, R.V., Gilra, A., & Halassa, M.M. (2018). Thalamic regulation of switching between cortical representations enables cognitive flexibility. Nature Neuroscience, 21(12), 1753-1763.

Schmitt, L.I., Wimmer, R.D., Nakajima, M., Happ, M., Mofakham, S., & Halassa, M.M. (2017). Thalamic amplification of cortical connectivity sustains attentional control. Nature, 545(7653), 219-223.

Wang, B.A., Wang, M.B., Lam, N.H., Liu, M., Li, S., Wimmer, R.D., Paz-Alonso, P.M., Halassa, M.M., & Pleger, B. (2025). Thalamic regulation of reinforcement learning strategies across prefrontal-striatal networks. Nature Communications, 16, 9095.

Wang, M.B., Lynch, N., & Halassa, M.M. (2025). The neural basis for uncertainty processing in hierarchical decision making. Nature Communications, 16, 9096.

Wimmer, R.D., Schmitt, L.I., Davidson, T.J., Nakajima, M., Deisseroth, K., & Halassa, M.M. (2015). Thalamic control of sensory selection in divided attention. Nature, 526(7575), 705-709.