The affective gradient hypothesis: an affect-centered account of motivated behavior

The paper addresses a central question in motivation and decision-making: what drives motivated behavior, and what is the role of feelings versus abstract value computations? Contemporary models typically foreground cold value estimates and treat affect as secondary (as an output, input, or modulatory influence). The author hypothesizes instead that affect is the sole form of value that drives all motivated behavior. People continuously represent possible future states and their associated feelings, and behavior emerges from moving toward anticipated positive affect and away from anticipated negative affect. This reframing aims to resolve puzzles about the origins of values and goals and to reconcile apparent conflicts between emotion and cognition by centering affect as the core driver of action.

The paper reviews dominant accounts of motivated behavior and the role of feelings within them: (1) feelings as epiphenomena of value (e.g., regret as counterfactual value, happiness as aggregated prediction errors, mood as average expected reward); (2) feelings as sources of learned value (experienced utility shapes decision utility); (3) feelings as inputs to decision value at choice (integral affect about outcomes or incidental mood states influencing decisions); and (4) feelings as drivers of a separate hot or Pavlovian control system that competes with a cold, goal-directed system. These perspectives all presuppose a valuation process that can be independent of affect. The review highlights unresolved issues these models face, such as how non-affective value is assessed, how multiple value systems are reconciled, and how feelings are abstracted into a common value signal robust across contexts. This motivates consideration of a feelings-only approach that could account for the same behavioral phenomena without positing a separate cold valuation system.

This is a theoretical and integrative framework rather than an empirical study. The author develops the Affective Gradient Hypothesis (AGH) by embedding affect within an internal state-space model of cognition: (1) every represented state (current, remembered, or imagined) has affective features with specific identity and intensity along dimensions such as valence and arousal; (2) affect can be evoked by experience, recall, or prospection; (3) states and their affective features can be activated bottom-up (cues, associative spread) or top-down (directed search, deliberation); (4) multiple affective features can be co-activated in parallel; and (5) the salience of affective features scales with the accessibility and vividness of their associated states (e.g., immediacy, probability, psychological distance). On this basis, the mind maintains a landscape of currently accessible states that together define an affective gradient. Actions are characterized by how they change the likelihood of reaching states with more positive affect and avoiding states with more negative affect. AGH proposes that behavior and control states are optimized by moving along this gradient to maximize expected positive affect and minimize expected negative affect. The framework connects to optimization and control theory: both the action/control space and the affective objective are multivariate, yielding a many-to-many mapping and a form of multi-objective optimization. The theory is contrasted with standard value-based models: instead of starting from options and computing decision values with affect as an add-on, AGH starts from fluctuations in expected affect that determine whether, how, and which decisions are made, as well as levels of engagement and control policy (e.g., attention focus and response thresholds). Boxes and figures elaborate gradient principles (Box 1), the goal selection problem and how AGH resolves the homunculus problem by treating goals as emergent from affective priorities (Box 2), and sources of within- and between-person variability in motivated behavior (Box 3).

• Proposal: Affect is the only form of value driving all motivated behavior; there is no affect-free valuation process underlying choices and control.
• Affective state-space: Every represented state carries affective features that can be accessed automatically or deliberately; multiple affective expectations can be active in parallel with salience scaling by accessibility and vividness.
• Affective gradient: The set of accessible affective consequences defines a gradient that guides action selection and control settings to approach anticipated positive affect and avoid anticipated negative affect.
• Predictions for choice and control: Measures of expected affect should better predict choices and decision engagement than objective reward metrics; expected affect determines not only which option is chosen but whether to engage, how to allocate attention, and when to switch tasks. Stakes and their affective salience should increase effort and persistence; reduced salience should encourage disengagement.
• Reframing dichotomies: Hot versus cold processing reflects differences in how affect is engaged (e.g., intensity, arousal), not whether affect is involved; both automatic and deliberate evaluations rely on affective representations that may share neural codes.
• Self-regulation conflicts: Intrapersonal conflict can be modeled as competition among coactive affective expectations (e.g., immediate enjoyment vs. social or long-term consequences) rather than competing selves or systems; outcomes depend on the relative strength and salience of these affective representations.
• Common currency not required: Adaptive behavior can arise without integrating into a scalar common currency; affective features can influence actions and control in parallel. This predicts difficulties with direct comparison of qualitatively different options and with strictly marginal counterfactual evaluations, with neural signals reflecting option- and task-level affective representations rather than pure scalar comparisons.
• Scope and applicability: The framework generalizes across tasks and real-world settings, offering a unifying account for goal selection, mind-wandering versus task engagement, and dynamic control adjustments (e.g., after errors).

Centering affect as the driver of motivated behavior addresses unresolved questions in value-based and control theories: it explains where values and goals originate and how goal priorities shift over time without invoking a separate cold valuation system or an explicit, stable goal hierarchy. By embedding affect in a rich state-space that can be accessed via bottom-up cues or top-down deliberation, AGH unifies reflexive and goal-directed actions under a single affect-driven optimization process. This reframing dissolves the hot versus cold dichotomy, reconceptualizes self-regulatory conflict as competition between affective expectations, and mitigates the need for a neural common currency of value. The theory yields testable predictions about when and how affect should shape choice, control allocation, task engagement, and task-switching. It also connects to neural and computational literatures on state representations, control optimization, and affect decoding, suggesting that shared neural substrates could represent anticipated affect for both automatic and deliberative evaluations and that control policies are tuned to the salience and valence of anticipated outcomes.

The paper advances the Affective Gradient Hypothesis: affect is the sole value signal underpinning motivated behavior. Represented states carry affective features that, when accessible, define gradients guiding actions and control to approach better and avoid worse anticipated feelings. Goals are emergent from currently salient affective associations rather than fixed controllers, offering a parsimonious solution to goal selection and prioritization. The framework suggests new empirical directions: measuring accessible affective features throughout tasks, including task-extraneous states; modeling many-to-many mappings between affect and actions; and probing neural codes for anticipated affect across automatic and deliberative contexts. Future work should clarify how affective features are learned and organized in state space, how affect-driven optimization is implemented neurally, and how individual differences in affective goals and expectations shape motivation across development and psychopathology.

As a theoretical proposal, AGH lacks direct empirical tests across the full range of predictions. Key open questions include: (1) how affective features are learned and abstracted from interoceptive and exteroceptive signals, and the roles of emotion categories and appraisals; (2) the extent to which motivated behavior can persist without affective input, including value-free habits, reflexes, or recurrence-based task engagement; (3) the structure and boundaries of the internal state space that scaffolds affective representations and what determines state accessibility; and (4) the computational and neural implementation of many-to-many, multi-objective optimization linking affective features to action and control, including whether agents maximize along specific affective axes or maintain set-points and how individual affective goals modulate gradients.