Problem solving that becomes repetitive and unproductive.
Future threat representations may keep high priority.
Uncertainty can keep them goal-relevant.
Removal failure leaves them ready to return. Borkovec et al., 1983; Meyer et al., 1990
RNT as a shared process
Section figure slot
Measurement
Measurement separates content from process
Process measures
PTQ, RTQ, and RTQ-10 assess repetition, intrusiveness, difficulty disengaging, and unproductiveness.
These scales align closely with a process account of RNT. Ehring et al., 2011; McEvoy et al., 2010; Topper et al., 2014
Content measures
PSWQ measures worry as generalized uncontrollable future-oriented thought.
RRS measures rumination as repetitive self-focused response to sad mood. Meyer et al., 1990; Nolen-Hoeksema & Morrow, 1991
State measures
Induction tasks, probes, EMA, and diary methods measure recurrence, dwell time, and controllability.
These methods connect trait RNT to real-time persistence. Moberly & Watkins, 2008; Kircanski et al., 2015
Measure family
Primary target
Best use in this model
PTQ, RTQ, RTQ-10
Repetition, intrusiveness, difficulty disengaging
Predict behavioral and neural indices of removal failure.
PSWQ, RRS
Worry and rumination content
Test whether content focus changes removal demands.
EMA and diary
Episode frequency, duration, mood coupling
Index real-world consequences of incomplete removal.
Section figure slot
Internalizing hierarchy
Hierarchy of RNT dimensions and diagnoses
Level 1
Common Internalizing and Broad RNT
Broad RNT captures shared variance across depressive and anxiety symptoms.
Explains why RNT predicts recurrence, comorbidity, and cross-disorder transitions.
Aligns with the p-factor/general internalizing dimension.
Spinhoven et al., 2018; Smolker et al., 2023; Moulds & McEvoy, 2025
Level 2
Specific Dimensions
Rumination is more past or self-evaluative, with emphasis on loss, failure, and meaning.
Anxious apprehension is more future threat and uncertainty focused.
These retain specific content and temporal focus within the shared process.
Stade & Ruscio, 2023; Watkins & Roberts, 2020
Level 3
Diagnostic Categories
MDD and GAD share broad RNT but differ in dominant content features.
Diagnoses load onto both common and specific dimensions.
Removal hypotheses can test whether content dimensions have different control demands.
Snyder et al., 2022; Moulds & McEvoy, 2025
Section figure slot
Persistence evidence
RNT is tied to difficulty discarding no-longer-relevant material
Say: Three converging lines of evidence anchor the clinical claim. Zetsche et al.'s 2018 meta-analysis of 94 studies found that discarding — the ability to remove no-longer-relevant WM content — showed the strongest association with RNT among all cognitive control functions. Shifting, updating, and inhibition were all weaker. This is not a generalized executive deficit; it is a specific impairment in releasing content. Whitmer and Gotlib's attentional scope model gives us a mood-based amplifier: narrowed scope makes the same negative content more likely to be selected again. And Working Memory 2.0 from Miller et al. grounds this neurally: gamma tracks content, beta is the suppression signal, and disrupted beta-gamma coupling is the neural signature of removal failure.
Behavioral convergence
RNT is linked to difficulty disengaging from negative material once it enters attention — not a broad executive deficit.
Zetsche et al.'s 2018 meta-analysis of 94 studies: discarding no-longer-relevant WM content showed the strongest association with RNT (r = -0.20).
Associations with shifting, updating, and inhibition were substantially smaller (r ~-0.10). Pattern remained after controlling for depression and anxiety. Koster et al., 2011; Zetsche et al., 2018
Attentional scope
RNT may narrow the contents held in working memory, reducing the range of competing representations.
Negative mood biases attention toward mood-congruent material; this makes the same negative content more likely to be re-selected.
Positive mood broadens attentional scope and supports disengagement. This model links mood state to the ability to clear cognitive space. Whitmer & Gotlib, 2013
Persistence route from entry to recurrence
Section figure slot
Working memory
Working memory models: foundations and convergence
Say: I am drawing on three working memory frameworks, and I want to be clear about what each one contributes. Baddeley gives us effortful executive maintenance — disengagement is an executive demand, not passive decay. Cowan gives us the most important single insight: content outside the focus is not gone; it remains activated and can re-enter without any new input. Oberauer refines this with priority weight — failure to downweight a representation leaves it competing for re-entry. All three converge on the same point: leaving focal awareness does not eliminate influence.
Baddeley & Hitch
Multi-component model
Phonological loop, visuospatial sketchpad, and central executive — domain-specific maintenance subsystems coordinated by limited attentional control.
Later episodic buffer integrates information across domains and connects working memory to long-term memory.
Active maintenance is effortful and executive; disengagement is equally an executive demand, not passive decay. Dual-task paradigms showed selective disruption by modality — verbal vs. spatial loads do not fully interfere with each other.
Baddeley & Hitch, 1974; Baddeley, 2003
Cowan
Embedded-processes model
Working memory is a subset of activated long-term memory, with a smaller focus of attention (~3-4 chunks) nested within a broader activation zone.
Representations that leave the focus are not gone — they remain activated and can re-enter attention without any new external cue. This is the core mechanism for intrusive return in RNT.
Lapses of executive attention allow previously active content to recapture processing; the thought returns not because something triggered it, but because it was never sufficiently deactivated.
Cowan, 1995; Cowan, 2001
Oberauer
Three-embedded-components model
Three nested regions: activated long-term memory, a capacity-limited direct-access workspace, and a single-item focus of attention for active processing.
Priority weight governs selection — items in direct access influence processing speed even when not focal; negative content may hold chronically elevated priority through salience, self-relevance, and affective value.
Removal requires active downweighting; failure to reduce priority leaves content competing for re-entry — the mechanism most directly relevant to RNT persistence.
Oberauer, 2002; Oberauer, 2009
Representational states: nested zones of accessibility — upload your figure here
Model
Key structural claim
What it explains about persistence
Shared contribution
Baddeley & Hitch
Separate storage buffers coordinated by a central executive with limited control capacity.
Removal is an executive function; capacity limits mean competing demands impair disengagement of negative content.
All three agree: content can remain active or available after leaving focal attention. Returning requires no new input — only a failure of control or a shift in priority.
Cowan
Activated LTM with a nested focus of attention (~3-4 chunks); no separate buffers.
Content in the activation zone can re-enter the focus without any new cue; lapses allow negative content to recapture processing.
Oberauer
Three nested regions differentiated by priority weight and selection probability.
Failure to downweight prior representations leaves them competing for re-entry; negative content holds elevated priority through salience or self-relevance.
Section figure slot
Working Memory 2.0
Working Memory 2.0: oscillations, sparse coding, and layered control
Say: This is the updated neural framework — Miller, Lundqvist, and Bastos 2018. It fundamentally changed how we think about neural implementation. The classic view was persistent sustained spiking during a delay period. WM 2.0 replaced that with sparse burst-mode activity and a critical dissociation between gamma, which tracks content, and beta, which is the suppression control signal. Beta in deep cortical layers is negatively correlated with gamma in superficial layers — that is the neural signature of control actively suppressing content. That is exactly what removal should look like. Zhou et al. then showed that selecting from WM and from perception recruits the same frontoparietal circuits — which means the same attentional mechanisms that orient us externally are also selecting our internal negative representations.
Working Memory 2.0 — Miller, Lundqvist & Bastos, 2018; Zhou et al., 2022
Sparse spiking: WM content is maintained through brief neural bursts rather than persistent sustained firing. Absence of sustained activity does not mean a representation is gone.
Gamma = content signal: Gamma-band oscillations in superficial cortical layers track what is held in mind, increasing with WM load.
Beta = control signal: Beta-band oscillations in deep cortical layers are the suppression signal — beta rises when content is no longer needed. Beta and gamma are negatively correlated across layers.
Relevance for removal: Successful removal should produce increased beta in control regions (DLPFC, FPCN) coupled with decreased gamma in content-bearing regions — the neural signature of control suppressing content.
Zhou et al. (2022): Selecting from WM recruits the same frontoparietal circuits as selecting from perception. The same attentional mechanisms that orient us externally are selecting our internal representations — including intrusive negative thoughts.
Miller, Lundqvist & Bastos, 2018; Zhou et al., 2022
WM 2.0: gamma-beta dissociation — upload your figure here
Why WM 2.0 matters for understanding RNT and removal: The gamma-beta framework gives us the first neural vocabulary specific enough to test removal failure in RNT. If successful removal requires beta to suppress gamma, then people with high RNT may show chronically disrupted beta-gamma coupling during removal attempts — control signals that cannot override elevated content signals for negative material. Zhou et al.'s finding that WM selection and perceptual selection share the same frontoparietal circuits further predicts that RNT-related control deficits should be visible precisely when the system tries to gate out an internal negative representation. Together, WM 2.0 turns removal failure from a behavioral description into a measurable neural signature — one that future EEG and MEG work can directly test. Miller et al., 2018; Zhou et al., 2022
Section figure slot
Removal progression
Working memory removal as a linear progression
Say: This sequence is the empirical backbone of the argument. Banich established that removal is observable and that different operations recruit distinct neural patterns. Kim shifted from where to what — tracking the actual representation through MVPA decoding and showing that operations differ in their effect on proactive interference. My 2024 work showed how those operations are organized across large-scale networks. My 2025 work showed that individual differences in thought-control difficulty predict how distinctly those neural representations are organized — and that this is specific to task performance, not resting state.
Banich to Kim to DeRosa
Banich et al.
Clearing as active control
Removal is not passive decay.
Clearing engages control and interruption systems.
Behavioral interference shows whether previous content still matters.
Kim et al.
Neural evidence and interference
Classifier evidence can track whether a removed item remains decodable.
Lower evidence after removal should predict less later interference.
DeRosa et al.
Operations and individual differences
Maintain, replace, suppress, and clear can be separated as control states.
Thought-control difficulty can be linked to representational and network patterns.
Progression
What it adds
Why it matters for RNT
Interference logic
Old content is not removed if it disrupts the next task state.
RNT can be tested as persistent influence, not only repeated report.
Neural evidence logic
Classifiers test whether old content remains decodable.
Negative content may remain accessible even after disengagement attempts.
Operation logic
Different removal operations can produce different representational outcomes.
RNT risk may reflect which operations fail and under what conditions.
Section figure slot
Task evidence
Task evidence for lingering content
Say: The key distinction Kim established is between displacement and removal. Replace and Clear displace content from focal attention early — classifier evidence drops quickly. But proactive interference still occurs on the next trial, meaning the representation stayed accessible from the direct-access zone. Suppress is different: it reverses proactive interference, indicating a deeper reduction of the representation's accessibility. This distinction — displacement versus genuine reduction — is the heart of the model's clinical prediction.
Accessibility
Content remains available if it can still be retrieved, selected, or decoded after it is no longer task relevant.
This links subjective recurrence to measurable representational persistence.
Interference
Old content remains influential if it disrupts new encoding, retrieval, or selection.
This provides the behavioral bridge from working memory removal to RNT.
Evidence marker
Predicted pattern if removal succeeds
Predicted pattern if removal fails
Behavioral interference
Prior content has little effect on the next trial or state.
Prior content slows, biases, or disrupts the next state.
Classifier evidence
Evidence for the removed content drops toward baseline.
Evidence remains detectable after the removal instruction.
Individual differences
Lower RNT predicts stronger reduction in old content influence.
Higher RNT predicts lingering accessibility or weaker representational change.
Section figure slot
Removal operations
Four operations, distinct representational consequences
Say: This is the core empirical summary from Kim et al. Reading across the rows tells you the dissociation pattern. Look at the Proactive Interference row: Maintain, Replace, and Clear all produce interference on the next trial. Suppress reverses it. That means Replace and Clear displace content from focal attention but do not reduce its accessibility — the representation stays in the direct-access zone. Suppress produces a fundamentally different outcome. Now look at classifier evidence: Replace and Clear drop early, but interference persists. Suppress drops later at the category level but earlier at the item level — it acts on fine-grained representations first. The RNT row at the bottom is my clinical translation: which operation best describes what a person with RNT needs to do, and which is most likely to fail?
Maintain
Replace
Suppress
Clear
Classifier evidence
Preserved throughout — evidence stays high. Item remains decodable from ventral visual cortex.
Early, large reduction — quickly displaced from focus. But category evidence may linger.
Earlier item-level reduction; later category-level reduction. Acts on fine-grained features first.
Early, large reduction — rapid disengagement. Representation quickly absent from focus.
Proactive interference (N+1)
High interference on next trial — item fully accessible, disrupts new encoding.
Interference STILL occurs — prior item accessible despite displacement from focus.
Interference still occurs — content accessible despite absence from focal awareness.
Behavioral access (RT)
Fastest responses to manipulated items — maximum accessibility maintained.
Still faster responses — representation remains accessible.
Slower responses to manipulated items — accessibility reduced below baseline.
No advantage or disadvantage — typical accessibility benefit removed, but not further reduced.
RNT relevance
Analogous to continued engagement with a negative thought — maintaining full accessibility.
Moving on by substituting new content, but the old representation lingers in direct access.
Deepest representational change — most directly relevant to breaking RNT cycles. Most vulnerable to failure in high-RNT individuals.
Empties WM without replacement — relevant when the goal is to clear cognitive space of negative content.
Classifier evidence over time by operation — upload your figure here
Section figure slot
Brain systems
Brain systems: neural implementation of persistence and removal
Say: These five systems are not separate independent contributors — they are interconnected, and their interactions are what the model predicts. The DMN and medial temporal systems support the self-relevant autobiographical content that is hardest to remove. The FPCN is the operation engine — it represents all four operations as distinct control states, and the clarity of that representation predicts thought-control difficulty. The salience network and valuation regions maintain the affective priority that negative content carries. And medial temporal systems provide the mechanism for cue-driven reinstatement — content that was moved out of the focus but not weakened gets reactivated by mood, context, or related cues. The critical question for future work is how FPCN interacts with DMN, salience, and medial temporal systems during attempted removal.
Default mode
Click to expand
mPFC, PCC, precuneus, and angular regions. Relevant to self-referential and internally generated content.
Rumination and worry induction studies find engagement of mPFC, PCC, and medial temporal regions relative to neutral thought. Cooney et al. (2010) found greater default-mode activation in depressed ruminators.
DeRosa et al. (2024) found DMN specifically differentiates removal states — suppress and clear from maintain and replace. DeRosa et al. (2025): DMN representational distinctiveness predicts thought-control difficulty. Andrews-Hanna et al., 2014; Cooney et al., 2010; DeRosa et al., 2024, 2025
Frontoparietal control
Click to expand
DLPFC, lateral parietal cortex, and frontopolar regions. The operation engine of the model.
DeRosa et al. (2024): FPCN showed fully dissociable representations for all four operations — each treated as a separate control state. Distance in representational space predicted classifier accuracy.
DeRosa et al. (2025): less distinct, less stable FPCN representations predicted greater thought-control difficulty. This association was specific to task performance, not resting state. Banich et al., 2015; DeRosa et al., 2024, 2025
Salience network
Click to expand
Anterior insula and dorsal anterior cingulate cortex. Relevant to monitoring, salience, and interruption of ongoing content.
Paulesu et al. (2010): persistent dACC and mPFC activation during worry in GAD, continuing into post-worry rest. Correlated with PSWQ scores.
May explain why negative representations keep winning attentional competition even after an attempted removal — by continuously flagging content as motivationally significant. Menon, 2011; Paulesu et al., 2010; Steinfurth et al., 2017
Valuation regions
Click to expand
Amygdala, orbitofrontal cortex, rostral ACC, and medial PFC. Relevant to threat value, self-relevance, and affective priority.
Cooney et al. (2010): depressed participants showed greater amygdala, rACC, OFC, and parahippocampal activation during rumination than controls — heightened affective significance and self-relevance.
These regions may maintain priority of negative representations after attempted removal, making them resistant to downweighting. Cooney et al., 2010; Steinfurth et al., 2017
Medial temporal / memory
Click to expand
Hippocampal and parahippocampal regions. Enable associative reinstatement — environmental or mood-based cues retrieve prior negative representations even after disengagement.
This is Cowan's mechanism in neural form: content moved from the focus to the activated LTM zone can be reactivated without deliberate retrieval, by a related cue or a shift in mood.
For RNT: a person moves on from a worry or regret, but a mood state, a reminder, or a semantically related thought reactivates it from the accessible zone. Andrews-Hanna et al., 2014; Puccetti et al., 2025
Sensory systems
Click to expand
Ventral visual and superior temporal regions. Used by Banich et al. (2015) and Kim et al. (2020) as compliance checks — these regions stay active when content remains in WM and return to baseline when successfully removed.
In WM 2.0 terms, gamma-band activity in these sensory regions tracks the content signal. Successful removal should show gamma reduction here, paired with beta increase in control regions.
Useful when classifier evidence tracks persistence of perceptual content in ecologically valid RNT paradigms using personally relevant negative images or scenarios. Banich et al., 2015; Kim et al., 2020; Miller et al., 2018
Network / System
Regions
Role in the model
Key evidence
Default mode
mPFC, PCC, precuneus, angular regions
Content-bearing system for self-referential and internally generated material; differentiates removal states.
Cooney et al., 2010; DeRosa et al., 2024, 2025
Frontoparietal control
DLPFC, lateral parietal, frontopolar cortex
Implements all four operations as distinct control states; representational clarity predicts thought-control difficulty.
Contributes affective priority and self-relevance; may resist downweighting of emotionally significant negative content.
Cooney et al., 2010; Steinfurth et al., 2017
Medial temporal
Hippocampus, parahippocampal cortex
Enables cue-driven reinstatement of content from activated LTM zone; mechanism for thought returning without deliberate retrieval.
Andrews-Hanna et al., 2014; Puccetti et al., 2025
Section figure slot
Synthesis
Toward a neurocognitive model of removal and RNT persistence
Say: The top row is the cognitive sequence — negative input, WM representation, control operation, outcome. The middle row shows the three factors that shape whether removal succeeds: self and memory systems providing the content that is hardest to remove, incomplete removal leaving the representation in direct access where it interferes, and affective priority maintaining the representation's competitive advantage. The bottom row shows the consequences that cascade from incomplete removal — observable markers in the lab, RNT episodes in daily life, and internalizing risk across time. The loop arrow is the key: content not sufficiently removed cycles back into working memory when mood, cues, or related representations reactivate it. That cycle, repeated across daily life, is RNT.
Integrated neurocognitive model
Working memory
Negative content can remain active, accessible, or prioritized even when it is not focal — across all three WM frameworks.
The state of the representation determines its later influence. Content outside the focus is not gone.
WM 2.0: gamma tracks content being maintained; burst-mode spiking maintains representations without continuous sustained firing. Cowan, 2001; Oberauer, 2002; Miller et al., 2018
Removal
Removal is successful when old content loses measurable influence — observed through behavior, classifier evidence, and operation-specific brain patterns.
WM 2.0: successful removal should produce increased beta in control regions (DLPFC, FPCN), coupled with decreased gamma in content-bearing regions — the control signal suppressing the content signal.
Replace and Clear displace; Suppress genuinely reduces accessibility and reverses proactive interference. Kim et al., 2020; Miller et al., 2018; DeRosa et al., 2024
RNT and internalizing risk
RNT reflects repeated return of content that was not sufficiently removed — still accessible from the direct-access zone, still elevated in priority through salience and self-relevance.
Less distinct neural representations of operations in DMN and FPCN predict thought-control difficulty. This is task-specific, not a resting-state property.
Trait RNT may reflect an enduring WM control vulnerability that maintains internalizing risk even when acute symptoms remit. DeRosa et al., 2025; Spinhoven et al., 2018; Smolker et al., 2023
Section figure slot
Future tests
Future tests and research questions
Say: These four questions operationalize the model's predictions. The first is the most direct test: add negative stimuli to the paradigm and test whether PTQ or RTQ scores moderate the decay of classifier evidence and the magnitude of proactive interference. The second is the WM 2.0 bridge: does suppress produce a different beta signature than clear? Can we decode the operations from oscillatory signatures? The third is ecological validity — do lab measures of representational persistence predict daily RNT episodes? The fourth tests valence specificity: does negative content specifically resist removal in high-RNT individuals, above and beyond general cognitive load effects?
Research question
Testable prediction
Design
Primary readout
Model implication
Does RNT predict weaker removal?
Higher trait RNT (PTQ/RTQ) predicts stronger residual classifier evidence and greater proactive interference for negative content after suppress and clear.
Negative and neutral stimuli; maintain, replace, suppress, clear; PTQ/RTQ as between-subject moderator.
RNT reflects incomplete reduction of representational influence, not just repeated occurrence.
Do beta dynamics track removal success?
Suppress and clear produce distinct beta increases in FPCN, coupled with gamma reductions in DMN and sensory regions. Beta-gamma coupling predicts trial-by-trial interference.
Combined EEG-fMRI or MEG removal task; time-frequency analysis by operation and valence; WM 2.0 oscillatory framework.
Beta power, gamma power, beta-gamma coupling across cortical regions and layers; oscillatory predictors of N+1 interference.
Grounds removal operations in WM 2.0 neural framework; identifies oscillatory signature of successful vs. failed removal.
Does removal failure predict daily RNT?
Weaker neural removal (less distinct FPCN and DMN operation representations) predicts more frequent, longer RNT episodes and stronger mood-RNT coupling in daily life.
Lab removal fMRI task followed by EMA diary; symptom moderation by PTQ/RTQ; mood x RNT coupling analysis.
Bridges lab mechanism to real-world persistence; tests whether removal ability explains trait RNT across timescales.
Does negative content specifically resist removal?
Suppress is less effective for negative than neutral content in high-RNT individuals — slower classifier evidence decay, more residual interference, weaker beta-gamma dissociation.
Valence-crossed operation design; individually tailored negative stimuli; PTQ/RTQ as between-subject variable; valence x operation x RNT three-way analysis.
Valence x operation x RNT interaction in classifier evidence, proactive interference, and oscillatory signatures.
Tests whether negative valence specifically resists control operations; grounds affective priority in representational framework.
Prediction 1
RNT predicts weaker removal
Higher PTQ/RTQ predicts stronger residual classifier evidence for negative content after suppress and clear.
Effect should be strongest for self-relevant or threat-related material.
Prediction 2
Beta dynamics track removal
Suppress and clear should produce distinct beta increases in FPCN paired with gamma reductions in content regions.