Gearshift Fellowship: A Unified Supertask Platform

• Gearshift Fellowship is a research-grade video game platform integrating neurocomputational models to study adaptive learning and strategy shifts within a unified mission-based environment.
• It employs hierarchical, cross-mission tasks—spanning cognitive flexibility, instrumental learning, and social inference—using computational methods like reinforcement learning and Bayesian inference.
• The platform serves scientists, clinicians, and players by providing experimentally controlled, ecologically valid assessments and adaptive, personalized interventions.

Searching arXiv for the specified paper and closely related work to ground the encyclopedia entry. Gearshift Fellowship (GF) is a research-grade video game and neurocomputational platform introduced as the first prototype of a “Supertask,” a mission-based paradigm for studying how humans and artificial agents adapt to shifting environment demands within one unified game environment and one hierarchical modeling framework. It combines computational neurocognitive modeling with serious gaming, using a multi-mission, story-based setting in which players chase cars, decode visual codes, and interact with other “drivers,” while every click, response time, and choice is modeled with formal computational models such as reinforcement learning, Bayesian inference, and sequential sampling. The platform is designed simultaneously for scientists as an experimental system, clinicians as a phenotype-to-mechanism assessment and intervention tool, and players as a training environment for self-regulated learning, mood, and stress resilience (Ging-Jehli et al., 11 Jul 2025).

1. Conceptual definition and scope

Gearshift Fellowship was proposed to address a central question: how humans and AI systems learn when to persist, when to let go, and when to “shift gears” to a new strategy. In the underlying formulation, a Supertask is defined as a mission-based game environment that studies learning, decision-making, and behavioral regulation across multiple, reconfigurable contexts using one unified core mechanic and one hierarchical modeling framework. GF is the flagship implementation of that idea, and its design is explicitly grounded in cognitive neuroscience, computational psychiatry, economics, and artificial intelligence (Ging-Jehli et al., 11 Jul 2025).

The platform departs from traditional neuropsychological tasks by targeting adaptation as a process over time rather than isolated single-construct performance. The stated contrast is with paradigms that target single constructs in isolation, often have poor ecological validity, offer limited meaningful feedback or training, and make it difficult to understand how behavior changes under shifting goals and contexts. GF instead aims to combine experimental control with ecological validity and engagement while measuring cognitive flexibility, instrumental learning, social learning and avoidance, meta-cognition and strategy shifts, and links to mood, stress, and ADHD-like symptoms within one environment (Ging-Jehli et al., 11 Jul 2025).

The core logic of “shifting gears” is broader than rule switching alone. In GF, it includes switching rules or strategies, switching between self-reliance and reliance on others, and adjusting learning rates in response to volatility or changing latent structure. This suggests that the platform is intended not merely as a battery of embedded tasks, but as an integrated assay of context-sensitive regulation across multiple cognitive and social scales (Ging-Jehli et al., 11 Jul 2025).

2. Supertask architecture and hierarchical organization

GF operationalizes the Supertask paradigm through five key properties: goal-directed behavior, contextual reconfiguration, hierarchical embedding, continuity and memory dependence, and unified computational modeling. Goal-directed behavior refers to multi-step goals requiring sequences of interdependent decisions. Contextual reconfiguration denotes reuse of the same core mechanic under different framing, rules, rewards, and controllability conditions. Hierarchical embedding organizes behavior across trials, blocks, missions, and sessions. Continuity and memory dependence indicate that missions influence each other and that past performance and strategies matter. Unified computational modeling means that the same model family can be fit across missions, enabling trait-level inference (Ging-Jehli et al., 11 Jul 2025).

GF keeps a constant visible loop across missions: chase car, view a coded stimulus on a pedestal, choose how to respond, and receive reward or penalty. Around that invariant loop, the environment is reconfigured to instantiate distinct experimental demands. The hierarchical levels are explicitly described as nano-level for the single trial, micro-level for the block with fixed contingencies, meso-level for the mission with a shared theme or context, and macro-level for sessions, longitudinal play, and real-world outcomes. This hierarchical structure makes it possible to model changes both within a context and across contexts over time (Ging-Jehli et al., 11 Jul 2025).

A concise summary of the three missions emphasized in the prototype is given below.

Mission	Primary function	Core demand
Mission 1 – Cognitive Flexibility / Task Switching	Cued task-switching paradigm	Switch between rules and manage switch costs
Mission 2 – Instrumental Learning / Structure Learning	Hierarchical RL / structure learning situation	Infer which code feature is relevant from feedback
Mission 3 – Social Learning, Avoidance, and Trust	Partner-based delegation and controllability manipulation	Trust calibration, social inference, and avoidance

The prototype contains at least five missions, although the published discussion focuses on the first three. This suggests that the architecture is meant to be extensible rather than tied to a fixed tripartite task set (Ging-Jehli et al., 11 Jul 2025).

3. Missions, mechanics, and adaptive variables

Mission 1 implements cognitive flexibility through a cued task-switching paradigm in which the relevant rule is known: the player must determine which feature of the code, such as letter versus number, matters, cued by the car type. The central demands are rapid rule application, rule switching, and management of switch costs in response time and accuracy. Mission 2 removes the explicit rule and requires the player to infer the relevant code feature from feedback, thereby instantiating instrumental learning and latent structure discovery. Mission 3 introduces “partner drivers” with different reliability and intent profiles—kind, clumsy, and jerk—and allows delegation, mistrust, or avoidance under varying degrees of controllability loss (Ging-Jehli et al., 11 Jul 2025).

Across these missions, GF manipulates a stable set of variables. These include choices about which rule to apply and whether to delegate, rewards and penalties for correct or incorrect responses, response times and later effort-like trade-offs, uncertainty and volatility in contingencies, controllability and agency, and social context through partner type and reliability. In Mission 3, loss of control can be induced when partners “cut in front,” forcing unwanted delegation. Such manipulations allow the same raw interaction loop to instantiate classic task-switching, instrumental learning, and social trust paradigms within one integrated environment (Ging-Jehli et al., 11 Jul 2025).

The constructs measured by the platform are described in terms of persistence, disengagement, and adaptability. Persistence denotes continuing with the same strategy or rule even when evidence suggests it is suboptimal. Disengagement denotes dropping out of a strategy or avoiding effort, such as frequent delegation or avoidance of code engagement. Gear shift or adaptability denotes timely strategy change, including switching attended features after a cue switch, increasing learning rates under new latent structure, or changing trust and avoidance behavior in response to partner performance and controllability changes. The paper notes that closed-form “gear shift indices” are not yet formalized, but that the platform is designed to capture these forms of adaptation quantitatively (Ging-Jehli et al., 11 Jul 2025).

4. Computational modeling framework

The modeling program underlying GF is explicitly hierarchical and cross-mission. The current work extends analyses with hierarchical Bayesian models that combine reinforcement learning models for Missions 2 and 3, sequential sampling models or drift-diffusion models for Mission 1, Bayesian inference or latent structure models for hidden-rule and partner-type inference, and meta-learning models in which parameters themselves adapt across missions (Ging-Jehli et al., 11 Jul 2025).

For feedback-driven learning, the paper presents a standard Q-learning-type update: $Q_{t+1}(s,a) = Q_t(s,a) + \alpha \, \delta_t$

$\delta_t = r_t - Q_t(s,a)$

where $Q_t(s,a)$ is the expected value of action $a$ in state $s$, $r_t$ is observed reward, $\delta_t$ is the reward prediction error, and $\alpha$ is the learning rate. Choice is described with a softmax policy,

$$P(a \mid s) = \frac{\exp(\beta Q(s,a))}{\sum_{b} \exp(\beta Q(s,b))}$$

where $\beta$ is the inverse temperature. In the interpretation given for GF, higher $\delta_t = r_t - Q_t(s,a)$0 corresponds to stronger sensitivity to recent outcomes, whereas lower $\delta_t = r_t - Q_t(s,a)$1 corresponds to slower updating; high $\delta_t = r_t - Q_t(s,a)$2 corresponds to deterministic exploitation and low $\delta_t = r_t - Q_t(s,a)$3 to more random exploration (Ging-Jehli et al., 11 Jul 2025).

For Mission 1, response times and accuracy can be modeled with a drift-diffusion process: $\delta_t = r_t - Q_t(s,a)$4 with drift rate $\delta_t = r_t - Q_t(s,a)$5, noise $\delta_t = r_t - Q_t(s,a)$6, boundary separation $\delta_t = r_t - Q_t(s,a)$7, and non-decision time $\delta_t = r_t - Q_t(s,a)$8. The paper states that drift rates may be lower on switch trials, thereby producing switch costs in response time and accuracy, and that stress or clinical measures can be related to changes in $\delta_t = r_t - Q_t(s,a)$9, $Q_t(s,a)$0, or $Q_t(s,a)$1 (Ging-Jehli et al., 11 Jul 2025).

The hierarchical Bayesian formulation is framed at individual and group levels. Individual players have parameter vectors such as $Q_t(s,a)$2, while group-level priors are expressed with examples including

$Q_t(s,a)$3

$Q_t(s,a)$4

This architecture is intended to support partial pooling across players, estimation of trait-like parameters, and simultaneous analysis of cross-mission similarities and differences. A plausible implication is that GF is structured to move beyond descriptive psychometrics toward shared latent-mechanism estimation across cognitive and social contexts (Ging-Jehli et al., 11 Jul 2025).

5. Empirical findings and construct validity

The initial evidence comes from an online study with $Q_t(s,a)$5, described as ongoing. The reported results indicate construct validity by recovering effects familiar from traditional neuropsychological tasks while also revealing novel cross-context patterns. In Mission 1, participants showed classical switch costs: switch trials were significantly slower than no-switch trials, with mean difference $Q_t(s,a)$6, and accuracy was lower on switch versus no-switch trials, with $Q_t(s,a)$7. These results are stated to replicate canonical laboratory findings for task switching (Ging-Jehli et al., 11 Jul 2025).

Mission 2 showed learning curves in which accuracy increased over repeated presentations of the same cue-code mapping, and errors shifted from “wrong feature” mistakes to perceptual misclassification. Cross-mission analysis revealed that greater flexibility in Mission 1, expressed as smaller switch costs, predicted higher learning rates in Mission 2 as the task became more complex, with $Q_t(s,a)$8. Mission 3 showed that participants learned the social structure of the partner types over time, trusting kind more than clumsy and clumsy more than jerk partners. A further cross-context result was that higher learning rates in Mission 2 predicted greater willingness to trust clumsy partners in Mission 3, with $Q_t(s,a)$9, while Mission 2 and Mission 3 learning rates were not strongly correlated overall, implying partially distinct learning mechanisms for nonsocial and social uncertainty (Ging-Jehli et al., 11 Jul 2025).

The platform also produced behavior-symptom correlations using DASS-21, a self-efficacy scale, and Conners’ CAARS. Higher stress correlated with greater switch costs in Mission 1, $a$0, driven by more out-of-context errors, and with lower learning rates in Mission 2, $a$1. Depressed mood was not strongly linked to Mission 1 performance but was negatively linked to learning in Mission 2, $a$2. Higher self-efficacy was associated with more initial double-checking of jerk partners in Mission 3, $a$3, and under forced partnership with loss of control, with a reduced tendency to follow any partner’s signals and a greater tendency to guess independently, $a$4 (Ging-Jehli et al., 11 Jul 2025).

Engagement metrics were also reported. The dropout rate was less than $a$5 despite the multi-layered task structure, and participants reported high engagement and sense of value. Taken together, these findings support the claim that GF reproduces canonical cognitive effects while also surfacing context-specific associations among stress, depressed mood, self-efficacy, controllability, and social trust that are not accessible from isolated tasks alone (Ging-Jehli et al., 11 Jul 2025).

6. Clinical, AI, and training roles

GF is explicitly positioned as an experimental platform for scientists, a phenotype-to-mechanism tool for clinicians, and a training environment for players. For scientific use, the platform offers a single parametric environment in which multiple constructs can be probed and compared under hierarchical, model-based analyses. For clinical use, it is envisioned as a digital phenotyping system for identifying markers of rigidity, avoidance, intolerance of uncertainty, effort sensitivity, and controllability preference, with the longer-term goal of treatment personalization and longitudinal monitoring. For players or end-users, the training objective is to enhance reflection, meta-learning, self-efficacy, agency, and stress resilience through adaptive feedback and individualized missions (Ging-Jehli et al., 11 Jul 2025).

The human-AI dimension is integral to the project rather than peripheral. The environment can be presented to AI agents such as Q-learners and meta-RL agents, allowing direct comparison of their learning trajectories and failure modes with human behavior. The stated goals include benchmarking AI generalization and meta-learning under context shifts and identifying where humans and AI differ, for example in understanding intent. Planned extensions include embedding RL and meta-RL agents that infer latent player states such as fatigue, disengagement, or high uncertainty and dynamically adjust difficulty, pacing, mission framing, feedback style, and levels of controllability or social challenge. This is described as a closed-loop design in which player behavior drives model inference, which drives adaptive task reconfiguration, which then affects subsequent behavior (Ging-Jehli et al., 11 Jul 2025).

Future versions are also intended to include in-game AI characters that reflect back the player’s behavioral patterns as “mirror agents” and that are taught by the player through learning-by-teaching interactions. The paper links this idea to evidence that teaching a robot can improve metacognitive awareness and learning. This suggests that the platform is intended not only to assay deficits or styles, but also to instantiate adaptive interventions that alter those styles (Ging-Jehli et al., 11 Jul 2025).

The technical implementation supports these roles. GF uses a customized open-source jsPsych backend for precise trial-level timing, event logging, and real-time adaptation of task parameters. The frontend is cross-platform, designed for desktop and tablet use, with a minimalist accessible interface optimized for clinical and non-gamer populations. The data pipeline is structured for hierarchical Bayesian and RL modeling and is designed to integrate future physiological and neuroimaging measures such as EEG and eye-tracking for bio-computational phenotyping (Ging-Jehli et al., 11 Jul 2025).

7. Relation to adjacent “gearshift” literatures and stated limitations

The term “gearshift” in Gearshift Fellowship refers to strategy change, learning adaptation, and controllability shifts rather than vehicular gear selection. This distinction matters because “gearshift” is also used in automotive control for ecological gearshift strategy and EV transmission controller calibration, where it refers to discrete gear selection or controller tuning under energy or drivability objectives (Luo et al., 2024, Beaudoin et al., 2021). It also appears in the later systems paper “GF-DiT: Scheduling Parallelism for Diffusion Transformer Serving,” where “GF” explicitly stands for “Gearshift Fellowship” and denotes dynamic parallelism “gear shifts” and temporary GPU “fellowships,” but that usage concerns diffusion serving rather than the neurocomputational game platform itself (Qiang et al., 11 Jun 2026).

Within its own scope, GF is presented as proof-of-concept. The authors note that full hierarchical modeling and larger clinical studies are ongoing, and they identify several next steps: scaling to larger and clinical samples, including remitted depression; integrating EEG and eye-tracking; and developing and evaluating fully-fledged adaptive interventions. They also note technical and design constraints, especially balancing complexity with accessibility for clinical and non-gamer populations and ensuring transparency and interpretability for clinicians and participants (Ging-Jehli et al., 11 Jul 2025).

These limitations delimit the current evidential status of the platform. The available results support construct validity, engagement, and initial symptom-behavior associations, but not yet definitive clinical efficacy or finalized cross-mission mechanistic models. A plausible implication is that GF occupies an intermediate position between laboratory task battery, computational phenotyping platform, and adaptive intervention environment: more integrated and ecologically structured than standard tasks, but still in the early stages of validation and deployment (Ging-Jehli et al., 11 Jul 2025).