Entering into a self-regulated learning mode prevents detrimental effects of feedback removal on memory

The study examines how external incentives (informative feedback) and internal motivational states (self-regulated, intrinsically motivated learning) interact to affect both immediate performance and longer-term memory in a word-learning paradigm. Prior work shows extrinsic rewards can undermine intrinsic motivation (the undermining effect), with performance dropping when rewards are withdrawn. In educational contexts, feedback is often introduced after unsupervised behavior, yet the effects of adding or removing feedback on longer-term memory are less clear. Learning word meanings from context provides an ecologically valid way to study memory trace formation. Earlier studies without explicit rewards linked successful extraction of congruent meanings to increased pleasure, physiological arousal, and activation of the dopaminergic SN/VTA–hippocampal loop, with enhanced long-term memory; incongruent trials did not show this DA-related pattern. Both external feedback and intrinsic learning can modulate the SN/VTA–HP loop, but whether they interact or compete, and how timing of feedback presentation or withdrawal affects consolidation, remains debated. The authors hypothesized that immediate performance and subsequent memory would be differentially affected by feedback depending on trial type (congruent/DA-dependent vs. incongruent/less DA-dependent) and that timing would matter: feedback provided from the outset might impair later performance upon removal, whereas adding feedback after an initial no-feedback phase might boost performance. These hypotheses were tested in two independent samples (laboratory and online).

Design and task: Participants learned meanings of pseudo-words from two-sentence contexts. For congruent trials, both sentences supported a single coherent meaning; for incongruent trials, no single meaning fit both sentences. Trials were organized in blocks; first sentences for eight pairs were shown, then corresponding second sentences in randomized order. After each second sentence, participants indicated either the inferred congruent meaning or that meanings were incongruent. Informative trial-by-trial performance feedback (smiley/frownie) was provided on half the trials; scrambled images served as no-feedback placeholders on the other half. Feedback order was counterbalanced between subjects: one group received feedback in the first half of the session and then had it removed (feedback-first), while the other started without feedback and then received it (no-feedback-first). Day 2 involved a surprise memory test (free recall or forced-choice recognition). Subjective ratings (pleasantness, arousal; and confidence on day 2) were collected after trials.

Experiment 1 (laboratory): Participants: N=60 healthy right-handed native German speakers (Mage=24.27, SD=3.8; 40 females), recruited at Otto-von-Guericke University; ethics approval 212/19; informed consent obtained. Materials: 80 pseudo-words embedded in 80 sentence pairs (40 congruent, 40 incongruent). Stimuli via Psychtoolbox in MATLAB R2012b. Feedback: happy/sad smileys (informative) vs. phase-scrambled placeholders (no-feedback); each scrambled image shown once. Procedure: Learning session (~2 h): ten blocks of eight sentence pairs (four congruent, four incongruent per block). After responding, participants received feedback (or placeholders) and rated pleasantness/arousal on 9-point scales. Two between-subjects groups manipulated feedback order. Day-2 surprise memory test (~30 min; mean delay ~23 h): 31 participants completed cued free-recall (very low performance; not analyzed); 29 completed a 4-alternative forced-choice recognition test with options: (1) correct congruent meaning (or lure for incongruent), (2) lure meaning, (3) “incongruent,” (4) “I don’t know.” After each recognition response, participants rated pleasantness, arousal, and confidence (9-point Likert).

Experiment 2 (online): Participants: N=73 healthy native German speakers (age/gender not collected per local regulations); recruited from university participant pool; ethics approval 212/19; instructed to use desktop/laptop. Completed two sessions ~24.8 h apart. Materials/Implementation: Same content as Experiment 1; implemented with custom PHP and jsPsych v7.0. Block congruence ratios varied (1/7 to 4/4 congruent/incongruent). Procedure: Day 1 (~45 min): typed responses required; for incongruent trials, participants typed “different” (German “unterschiedlich”); no “I don’t know” option. Answers lowercased and string-matched to expected meanings. Feedback manipulation and order as in Experiment 1. Day 2 (~25 min): 3-alternative forced-choice recognition (two specific meanings or “different”). After each response, sliders captured confidence, pleasantness, and arousal (0–100).

Statistical analysis: In R (v3.6.1/4.2.1). For Day-1 performance (correct/incorrect) and Day-2 recognition (remember/forget), generalized linear mixed models (logistic link) with random intercept per subject were estimated. Fixed effects: congruence (congruent/incongruent), feedback (feedback/no-feedback), order (feedback-first/no-feedback-first), and their interactions. For recognition, primary analyses focused on items answered correctly on Day 1 to assess consolidation. Chance-level tests used Bonferroni-corrected two-sided t-tests. Packages: afex (mixed), lme4 for model fitting; emmeans for post hocs. A reanalysis of prior no-feedback data assessed potential fatigue/boredom effects using Bayesian mixed models (brms) to compute Bayes factors.

Experiment 1 (laboratory):

• Day 1 immediate performance: Participants performed above chance, with higher accuracy for incongruent than congruent trials (congruent: 57.7% ±1.7; incongruent: 79.4% ±1.4; main effect of congruence: χ²(1)=248.6, p<0.0001). Informative feedback improved overall performance by ~4.7% (main effect of feedback: χ²(1)=10.9, p=0.001). Critically, introducing feedback in the second half (no-feedback-first group) boosted performance substantially (from 67.7% to 75.1%; OR=1/0.696≈1.44; p=0.0003), whereas withdrawing feedback (feedback-first group) produced a small, non-significant drop (68.4% to 66.7%; OR=0.923, p=0.37). Feedback-by-order interaction: χ²(1)=4.42, p=0.04. Effects did not differ by congruence (all interactions with congruence ns).
• Day 2 recognition (recognition subgroup N=29; items correct on Day 1): Congruent items were remembered better than incongruent (62.4% vs 38.9%; main effect of congruence: χ²(1)=78.32, p<0.0001). For congruent items, withdrawing feedback caused a significant drop in recognition (62.8% with feedback vs 50.4% without; OR=1.66±0.4, p=0.034) in the feedback-first group, whereas adding feedback later did not significantly change recognition (OR=0.75±0.18, p=0.24). Interaction feedback×order (congruent only): OR=2.21±0.75, p=0.021. Post hoc: in the absence of feedback, recognition differed by whether items were learned before vs after feedback onset (OR=0.42, p=0.013); no order difference when feedback was provided (p=0.79). Reanalysis of prior no-feedback data found no half×congruence interaction (BF01=4.94) and no main effect of half (BF01=7.22), arguing against time/fatigue explanations. For incongruent items, no detrimental effect of feedback withdrawal; there was a trend for feedback improving memory (OR=1.62±0.45, p=0.08), independent of order (interaction OR=0.85±0.23, p=0.55). Recognition was above chance in all four incongruent conditions (all p<0.044). Free-recall (N=31) showed floor performance (overall 22.1%; congruent 4.4%), so not analyzed further.

Experiment 2 (online):

• Day 1 immediate performance: Accuracy higher for incongruent than congruent (congruent 41.5% ±1.4; incongruent 67.2% ±1.3; main effect of congruence: χ²(1)=380.95, p<0.0001). A three-way interaction (congruence×order×feedback: χ²(1)=4.25, p=0.039) showed that adding feedback in the second half improved congruent meaning extraction (from 40.5% to 47.8%; OR=1.34±0.14, p=0.005) for the no-feedback-first group; starting with feedback produced no significant change. No reliable effects for incongruent trials online.
• Day 2 recognition (all participants; 33% chance): Three-way interaction (congruence×order×feedback: χ²(1)=9.28, p=0.002). For congruent items, recognition dropped when feedback was removed (OR=1.63±0.30, p=0.035) but did not change when feedback was added later (OR=0.69±0.12, p=0.13); interaction feedback×order for congruent: χ²(1)=12.04, p<0.001. Incongruent recognition did not differ by feedback/order (all p>0.53).

Overall: Introducing feedback after an initial self-regulated phase improved immediate performance, especially for congruent trials. However, removing feedback impaired next-day recognition for congruent (DA-dependent) learning only; incongruent (less DA-dependent) learning showed no such detrimental effect and sometimes trended toward benefit from feedback.

External feedback and intrinsic self-regulated learning interact in a time- and content-dependent manner. Adding feedback after a period of self-regulated learning boosts immediate performance, likely by strengthening already established task routines rather than providing item-specific information (feedback followed the response). In contrast, for consolidation and next-day recognition, removing feedback harms memory specifically for congruent trials, which prior work links to dopaminergic SN/VTA–hippocampal loop engagement. This pattern resembles the undermining effect and suggests that when learning initially relies on external DA-related reinforcement (feedback), subsequent withdrawal leaves insufficient intrinsic DA-driven support for hippocampal LTP, impairing consolidation. Incongruent trials, being less DA-dependent in this paradigm, benefit modestly from feedback without showing undermining upon withdrawal. Differences between lab and online settings (e.g., weaker, less social feedback online) may reduce feedback efficacy, particularly for DA-independent processes. The findings align with theories of schema/congruency advantages in memory and with evidence that curiosity/insight and reward modulate hippocampus-dependent consolidation. Educationally, results support incorporating self-regulated, feedback-free phases when DA-dependent learning is expected, and judiciously using feedback to support DA-independent learning. Clinically, informative feedback may particularly aid patients in DA-independent learning contexts.

The study demonstrates that intrinsic and extrinsic reward signals interact differently for immediate performance versus long-term memory, and that effects depend on the nature of what is learned. Adding feedback after self-guided learning improves immediate meaning extraction, but withdrawing feedback selectively impairs next-day recognition for DA-dependent (congruent) word-meaning learning, while DA-independent (incongruent) learning is not harmed and can benefit from feedback. These findings bridge undermining-effect literature with dopaminergic memory consolidation models, emphasizing the importance of establishing self-regulated learning modes before introducing external feedback. Future research should combine neuroimaging to directly measure DA-related circuit engagement under different feedback timings, manipulate feedback source/social context, and explore optimization of feedback schedules in educational and clinical interventions.

• Free-recall performance on Day 2 showed floor effects (especially for congruent items), limiting interpretability of recall results.
• Online experiment feedback may have been weaker (automated, less social) than lab feedback, potentially reducing effects, particularly for incongruent trials.
• Age and gender were not collected in the online sample due to data protection, limiting demographic analyses.
• Although reanalyses argue against fatigue/boredom explaining effects, residual time-on-task or strategy differences cannot be entirely ruled out.
• Dopaminergic involvement is inferred from prior work and behavioral patterns; no direct neural or pharmacological measures were collected here to confirm DA mechanisms.