All experiments were approved by the Institutional Animal Care and Use Committee (IACUC) at the Medical University of South Carolina in accordance with the NIH-adopted Guide for the Care and Use of Laboratory Animals (Protocol 2021-01325-1). Adult (8–24 weeks) male and female C57BL/6J wild-type (WT) background mice (8 weeks old/20 g minimum at study onset; Jackson Labs or MUSC DLAR Breeding Colony) were group-housed pre-operatively and single-housed post-operatively under a reversed 12:12-h light cycle (lights off at 8:00 a.m.; all experiments were performed during the dark phase) with access to standard rodent chow and water ad libitum. To facilitate instrumental learning and maintain robust responding for a caloric reward (10% Ensure), mice were placed on a regulated water restriction schedule 3 days prior to the start of acquisition. This is a standard requirement for operant paradigms using liquid reinforcers to ensure that reward-seeking behavior is not prematurely extinguished by satiation (Goltstein et al., 2018; Grant et al., 2021; Guo et al., 2014; Otis et al., 2017; Otis et al., 2019). We employed a tiered supplemental water system to maintain each animal between 85 and 90% of its initial free-feeding body weight. Unlike fixed-volume restriction paradigms, our approach adjusted daily water access based on a combination of the animal’s current weight and its instrumental performance (lever presses) during the session. This method provides a more refined balance between behavioral motivation and animal welfare compared to traditional 24-h restriction protocols. To maintain mice at 85%–90% of their free-feeding body weight, we regulated their daily water intake based on their behavioral performance: 50–100 active presses: 0 mL water, 30–50 active presses: 0.25 mL water, 20–30 active presses: 0.5 mL water, 0–20 active presses: 1.0 mL water. Water was provided at least 2.5 h after the final behavioral session of the day. Mice were weighed daily before each session. If a mouse’s weight exceeded the target, its water allowance was reduced by one category to encourage weight loss. Conversely, if a mouse’s weight fell below the target, a separate set of criteria was applied to ensure adequate hydration: 50–100 active presses: 0.5 mL water, 30–50 active presses: 1.0 mL water, 20–30 active presses: 1.5 mL water, 0–20 active presses: 2.0 mL water. Mice were transitioned to a Teklad Diet (2,918, Inotiv) 1 week prior to training. This fixed-formula diet was maintained throughout the study to minimize batch-to-batch nutrient fluctuations and metabolic variability. This ensured that fluctuations in body weight were a result of the water restriction protocol rather than dietary inconsistency, providing a reliable metric for adjusting water intake.

For intracranial surgeries, all mice were anesthetized with isoflurane (0.8%–1.5% in oxygen; 1 L/min) and placed within a stereotaxic frame (Kopf Instruments). Ophthalmic ointment (Akron), topical anesthetic (2% Lidocaine; Akron), analgesic (Carprofen & Ketorolac, 2 mg/kg, i.p.), and subcutaneous sterile saline (0.9% NaCl in water) were given pre- and intra-operatively for health and pain management. An antibiotic (Cefazolin, 200 mg/kg, s.c.) was given post-operatively to reduce the possibility of infection, and mice were allowed to recover for at least 2–3 weeks prior to behavioral training. All mice received head-ring implantation. A custom-made head-ring (stainless steel; 5 mm ID, 11 mm OD) was implanted onto the animal skull, along with 2–3 screws and dental cement to allow for head fixation.

Lever press and lick data during acquisition and test sessions were collected using a custom Python interface, REACHER, connected to Arduino (Vollmer et al., 2021). Statistical analyses were performed using GraphPad PRISM (Dotmatics, version 10). Behavioral data were analyzed using either analysis of variance (ANOVA; two-way or three-way) or t-tests (paired or unpaired), as appropriate. The independent variables analyzed included: lever type (active or inactive), sex (male or female), virus (mCherry vs. DREADD), training duration (limited vs. overtrained), and day (acquisition, devaluation test, omission test). Sidak’s and Tukey post hoc tests were used following significant main effects and interactions with the ANOVA analyses. The two-stage linear step-up procedure of Benjamini, Krieger, and Yekutieli (BKY) was also used, as appropriate, for controlling the False Discovery Rate in multiple comparisons. Prior to analysis, all datasets were checked for normality and homogeneity of variance.

To ensure findings were robust against between-subject variability and missing values, key longitudinal datasets (e.g., acquisition) were cross-validated using linear mixed-effects models (REML), which yielded results consistent with our reported ANOVA findings. For analyses in Figure 2, mCherry control mice from chemogenetic cohorts were pooled with the original acquisition cohorts, as behavioral protocols were identical. Sample sizes were determined based on established standards in the field for chemogenetic behavioral studies. The adequacy of these sample sizes is supported by the robust effect sizes (partial eta squared) observed across all primary behavioral interactions. Mouse exclusion criteria were defined a priori. Specifically, (1) failing FR1 acquisition within 14 days; (2) loss of head-cap stability (3); retraining exceeding 7 days; or (4) viral expression outside the DLS. Altogether, one mouse was excluded during FR1 training, and another was excluded from omission testing, both due to head-cap instability.

To investigate the role of the DLS in regulating goal-directed and habitual reward-seeking behaviors, we used a site-specific chemogenetic approach. Mice were first trained on the limited 2-day or extended 7-day head-fixed instrumental paradigm, as described above. Upon successful completion of this training, mice received a systemic pretreatment of clozapine N-oxide [CNO, 10 mg/kg; i.p.; (Mahler et al., 2019)] 30 min before each test session, which included devaluation, valuation, omission, and omission probe. During devaluation test days, CNO was administered 15 min after the onset of the pre-feeding period. This timing was selected to ensure that peak DREADD-mediated inhibition coincided with the instrumental devaluation test. Importantly, DLS inhibition did not affect the pre-feeding process itself; no differences in the amount of Ensure consumed were observed between DREADD-expressing mice and controls [t(76) = 1.579, p = 0.1184], nor between the experimental groups and prior pilot cohorts [t(54) = 0.006, p = 0.9949].

After behavioral testing, mice were transcardially perfused, and brains were collected and post-fixed in 4% paraformaldehyde (PFA). For cryoprotection, brains were subsequently immersed in successive 15 and 30% sucrose solutions. Coronal sections (40 μm) were cut using a cryostat, mounted on slides, and imaged on a Leica THUNDER Imager Tissue.

Previous freely moving instrumental models have successfully produced both goal-directed and habitual behavioral control (Adams and Dickinson, 1981; Yin et al., 2006; Gremel and Costa, 2013; Corbit et al., 2012; Jones et al., 2022). However, a head-fixed model capable of capturing each behavioral strategy in isolation has not yet been developed. To this end, here head-fixed mice were trained to press an active lever to receive a 1 s tone cue paired with delivery of 10% (v/v) Ensure (Figures 2A,B). A second inactive lever was present but did not result in a cue or reward. To establish the desired goal-directed or habitual strategies using a paradigm adapted from well-validated models (Malvaez et al., 2018), training progressed across three different schedules of reinforcement: FR1, VI15s, and VI30s. VI schedules were included to promote habitual control, as the unpredictability of reinforcement reduces reliance on immediate A-O associations and promotes the development of stable, outcome-insensitive behavior. To assess initial learning trajectory before the divergence in training protocols, we compared the performance of limited and overtrained groups across their shared acquisition phase. Mice in both groups pressed the active lever significantly more than the inactive, revealing comparable initial lever discrimination and overall lever press rates that reflected the progressive development of robust discrimination between the active and inactive levers (Figure 2C). This acquisition pattern was further validated by a linear mixed-effects model, which confirmed the primary effects of Lever (p < 0.0001) and Day (p = 0.028) while accounting for individual subject variability. Assessment of sex differences revealed no significant main effect of sex on active lever press rates (Supplementary Figure S2A). Overall, these data confirm successful and comparable initial acquisition of instrumental behavior across all animals prior to the manipulation of training duration.

After completing acquisition on their respective schedules, mice received a set of devaluation tests to assess their underlying behavioral strategy. In these tests, non-reinforced lever pressing was measured after a 45-min prefeed period with the Ensure flavor earned during training (devalued condition) and compared to pressing after pre-feeding with the alternative Ensure flavor provided outside the training context (valued condition). The order of the devalued and valued conditions was counterbalanced across subjects and each mouse experienced both conditions, allowing direct within-subject comparison of responding during the valued versus devalued state. Goal-directed actions are based on knowledge of the A-O relationship and are therefore flexible, resulting in a reduction of responding when the outcome is devalued. In contrast, habits are marked by a lack of consideration of action consequences, leading to inflexibility, and persistent responding regardless of outcome’s current value. To ensure devaluation wasn’t simply due to varying Ensure intake during prefeed, we measured prefeed consumption and found no difference across conditions or groups (Supplementary Figure S2B).

During devaluation tests, limited-trained mice demonstrated sensitivity to the changing value of the outcome, exhibiting a significantly lower lever-press rate on the devalued test day compared to the valued test day. In contrast, overtrained mice showed no change in lever press rate across the valued and devalued test days, demonstrating insensitivity to outcome devaluation (Figure 2D). To further quantify this behavioral distinction, a devaluation index was calculated (# of devalued lever presses) / (# of devalued + valued lever presses). This index provides a normalized measure of goal-directed versus habitual control, where 0.5 indicates insensitivity to devaluation (habitual control) and 0.0 indicates sensitivity to devaluation (goal-directed control). Consistent with the overall press rates, the devaluation index in limited-trained mice was reduced as compared with overtrained mice, confirming that limited-trained mice were selectively sensitive to reward devaluation (Figure 2E). The number of licks for the valued and devalued test days did not differ (Supplementary Figure S2C), suggesting consumption did not change and that the observed differences were specific to the instrumental response.

Following devaluation testing and subsequent retraining, mice received an omission test designed to test their behavioral flexibility by changing the A-O relationship. In this procedure, rewards were delivered automatically every 20 s, any lever press reset the timer, thereby delaying reward delivery. This test differentiates goal-directed actions from habits: limited-trained mice, relying on their knowledge of the A-O contingency, should rapidly reduce pressing to maximize the automatic rewards. In contrast, overtrained mice—whose actions are independent of the A-O contingency—are expected to maintain high press rates, demonstrating inflexibility. During the 60-min test, limited-trained mice showed behavioral flexibility by significantly reducing lever pressing in response to the reversed contingency. In contrast, overtrained mice were inflexible, continuing to press at high rates and failing to adjust their behavior despite the absence of reward (Figure 2F). Consistent with their persistent responding, overtrained mice received significantly fewer reward deliveries than limited-trained mice (Figure 2G). The number of licks for the omission test did not differ between the groups (Supplementary Figure S2D), suggesting consumption was not a contributing factor to the difference in lever pressing.

Twenty-four hours after the initial omission test, all mice completed a 10-min, non-reinforced omission probe test to assess memory consolidation of the new, reversed A-O contingency. Mice with limited training maintained low lever pressing rates, consistent with retention of the altered A-O contingency from the previous day. Conversely, overtrained mice continued to press at high rates, indicating persistent insensitivity to the prior omission contingency and a failure to adapt their behavior despite prior experience (Figure 2H). The number of licks during this non-reinforced omission probe test did not differ between the groups (Supplementary Figure S2E). In summary, these results demonstrate that our novel head-fixed instrumental learning paradigm successfully recapitulates the two distinct modes of behavioral control seen in freely moving models. The use of both devaluation and omission tests confirmed a clear divergence in behavioral strategy, with the limited-trained group exhibiting flexible, goal-directed actions and the overtrained group displaying inflexible habits.

To further verify the utility of this head-fixed approach, we assessed whether the habitual behavior would rely on the same canonical neuronal systems that have been described in freely moving studies. Thus, using chemogenetics we inhibited neuronal activity in the DLS during devaluation and omission tests in both limited-trained and overtrained mice. WT mice received bilateral DLS microinjections of either a Gi-coupled (Gi) DREADD (AAV2-hSyn-hM4Di-mCherry) or a control (CTL) virus (AAV2-hSyn-mCherry) (Figures 3A–C). Prior to each testing session (30 min), systemic administration of CNO (10 mg/kg, i.p.) was performed to activate the DREADD. Prefeed consumption was again measured and did not differ across groups or test day (Supplementary Figure S3A).

During devaluation testing, Gi-limited, CTL-limited, and notably, Gi-overtrained mice all demonstrated sensitivity to devaluation by significantly reducing lever presses on the devalued day compared to the valued day. Conversely, CTL-overtrained mice showed no reduction (Figure 3D). An effect of the DLS neuronal inhibition was further seen in Gi-overtrained mice pressing significantly less than CTL-overtrained mice on the devalued day, indicating that chemogenetic inhibition of the DLS prevented habit expression and restored goal-directed control in overtrained animals (Figure 3D, statistical result not shown). To account for individual differences in lever press rates, a devaluation index was calculated. CTL-overtrained mice displayed a higher index—indicating insensitivity to devaluation—compared to Gi-overtrained, Gi-limited, and CTL-limited mice (Figure 3E). Thus, Gi-overtrained mice exhibited sensitivity to devaluation, confirming that chemogenetic inhibition of the DLS restored goal-directed control in overtrained mice. Analysis of lick data showed that only the Gi-limited group displayed a significant difference in the number of licks between the valued and devalued test days, while all other groups showed no difference (Supplementary Figure S3B).

Next, we examined how chemogenetic DLS inhibition affected responding in omission and omission probe tests. In the omission test, we found that Gi-limited, CTL-limited, and notably, Gi-overtrained mice significantly reduced lever pressing and received more reward deliveries than the CTL-overtrained mice (Figures 3F,G). This pattern of responding demonstrates that Gi-overtrained mice, but not CTL-overtrained controls, regained sensitivity to changes in the instrumental contingency following chemogenetic inhibition of the DLS. This distinction persisted into the omission probe test a day later, where Gi-limited, CTL-limited, and Gi-overtrained mice maintained low press rates, while CTL-overtrained mice continued pressing at high rates despite the absence of the cue tone or reward (Figure 3H). In both omission and omission probe tests, lick data showed no change between any of the groups (Supplementary Figures S3C,D), confirming that the observed differences in pressing rates were due to behavioral adaptations and not consumption differences. These results demonstrate that DLS activity is required for the expression of outcome-insensitive responding in overtrained mice, as silencing this region reinstated goal-directed flexibility during both devaluation and omission tests.

Collectively, these results demonstrate that while CTL-overtrained mice continued to exhibit inflexible, habitual responding, the Gi-limited, CTL-limited, and, critically, the DLS-inhibited Gi-overtrained mice all displayed behavioral flexibility, significantly reducing their lever pressing when it no longer resulted in a valued reward. These findings provide strong evidence that chemogenetic inhibition of the DLS suppresses habitual behavior and restores goal-directed control, further validating our paradigm for modeling goal-directed and habitual actions.

Traditionally, studies of habitual behavior have relied on freely moving operant conditioning, which, while valuable, limits the spatiotemporal resolution and longitudinal tracking of individual neurons necessary for fully capturing the habit formation process. Recently, head-fixed setups have been adapted for various drug and reward self-administration tasks, allowing detailed neural monitoring during goal-directed learning (Clarke et al., 2024; Paniccia et al., 2024; Vollmer et al., 2021; Vollmer et al., 2022; Ward et al., 2025). Our work builds on this progress by showing that mice in a head-fixed setup not only learn goal-directed actions but can also transition to and sustain habitual behavior. Behavioral outcomes in our paradigm recapitulate key features of habitual control observed in freely moving animals, including insensitivity to outcome devaluation and persistent responding under altered contingencies. Crucially, this paradigm provides a methodological framework for addressing scientific questions that have remained inaccessible in freely moving models. The mechanical stability afforded by head-fixation allows for the longitudinal tracking of neural ensembles across the multi-day transition from goal-directed to habitual states. This level of access facilitates cell-level interrogation of the specific circuit dynamics and population shifts that occur as behaviors consolidate into outcome-insensitive habits. This establishes our system as a powerful, analogous platform for dissecting the neural circuit adaptations underlying habits.

Our head-fixed paradigm’s compatibility with high-resolution, 2P microscopy provides an opportunity to observe the precise neural adaptations that drive the transition to habitual behavior. 2P microscopy offers superior resolution, deeper tissue penetration, and reduced phototoxicity, enabling prolonged in vivo imaging sessions. This setup provides stable, longitudinal optical access to specific neural populations, allowing us to track cellular and synaptic changes as behaviors consolidate into habits. Crucially, this combined approach enables the direct correlation of single-neuron activity and population dynamics with the moment-to-moment emergence and expression of goal-directed actions and habits. Such detailed, time-resolved tracking of circuit activity during habit formation—including the mapping of dynamic neural codes across learning stages and the identification of specific cells recruited during the transition—has been largely inaccessible with previous methodologies. This level of interrogation will be essential for fully elucidating the underlying circuit mechanisms of behavioral inflexibility and rigidity.

Our functional manipulations provide strong support for the established role of the DLS in habitual behavior and, crucially, validate the fidelity of our head-fixed paradigm. The DLS is widely recognized as a critical hub for habit expression. We found that chemogenetic inhibition of the DLS in overtrained mice disrupted the expression of inflexible habits, restoring sensitivity to outcome devaluation and contingency reversal. While these results—that DLS activity is necessary for maintaining habits and that its modulation can reinstate goal-directed, flexible control—are consistent with findings from freely moving models (Balleine, 2019; Dickinson and Balleine, 1993; Malvaez and Wassum, 2018), our study extends this work by demonstrating that DLS activity is acutely necessary for the real-time expression of habits. Unlike foundational studies that utilized permanent lesions—thereby affecting both the acquisition and maintenance of habits—our results show that DLS activity is required to maintain the outcome-insensitive state even after a habit has been fully established. By showing that DLS modulation can reinstate behavioral flexibility, our work highlights the DLS as a potent therapeutic target for restoring goal-directed control.

A unique feature of this paradigm is the use of a consistent VI30s schedule for both limited-trained and overtrained cohorts, rather than the canonical random ratio (RR) to VI dissociation often employed in freely moving studies (Yin et al., 2005a; Malvaez et al., 2018). This VI → VI design was a deliberate methodological choice to isolate training duration as the sole independent variable. By maintaining a single reinforcement logic, we establish a standardized platform for future longitudinal imaging that avoids the confounding neural and behavioral shifts associated with rule transitions. This approach allows for the direct comparison of early-stage versus late-stage VI performance, facilitating the identification of circuit-level shifts that occur as behaviors consolidate into habits. While VI schedules are traditionally associated with accelerated habit formation, our limited-trained mice remained robustly sensitive to both sensory-specific devaluation and contingency omission. This empirical validation confirms that the limited-trained group maintains an A-O associative structure, making them a true goal-directed comparison for the overtrained habit phenotype. Furthermore, modeling the transition within a single VI schedule may more closely recapitulate the graded erosion of behavioral flexibility observed in naturalistic and pathological states, where habits often emerge from consistent, time-based reinforcement rather than shifts in work-to-reward ratios.

While the head-fixed instrumental learning paradigm presented here offers opportunities for investigating the computational neural dynamics of habitual behavior, it is important to acknowledge certain inherent limitations common to such in vivo models. While head restraint can potentially induce stress, the successful acquisition of flexible, goal-directed actions, comparable to freely moving studies, strongly suggests adequate acclimation and low confounding stress levels (Giovanniello et al., 2025). Future integration of air-lifted platforms could further enhance validity and physiological monitoring. Additionally, while we utilized a consistent VI schedule to isolate the temporal transition from flexible action to habit, we acknowledge that RR schedules classically sustain goal-directed behavior for longer durations. Comparing the neural signatures of our goal-directed VI cohorts to RR-trained animals remains an important future direction to determine how reinforcement logic interacts with the rate of habit consolidation. The simplified motor requirements of head-fixation may also reduce motivational demands relative to freely moving models; future work could introduce high response requirements or more complex task structures to better engage motivated behavior. Furthermore, because the motor requirements of head-fixation are simplified and constrained, this model does not directly assess sensorimotor dimensions of habit related to motor topography, such as the development of automated motor sequences or action chunking. Future iterations of this head-fixed preparation will allow for the integration of high-resolution tracking to investigate how these complex motor features emerge alongside outcome-insensitivity. Despite these considerations, this paradigm provides a robust framework for examining the transition to, and maintenance of, habitual behavior with unparalleled temporal and spatial resolution. Collectively, these identified limitations offer valuable avenues for future model refinement and expansion.