NOTE-1134: Mancinelli & Bach — A systematic review of human avoidance learning

Contribution

Earlier reviews of human avoidance learning were narrative and each covered one aspect (paradigms, cognitive theories, computational models). This one searches Web of Science, Scopus and PubMed systematically, codes 116 articles, rates each recurring phenomenon for robustness (A to C, Fig. 4), and scores five theories against them (Fig. 1). It is the first attempt to say which empirical regularities any theory of avoidance learning must reproduce, and which of them actually discriminate between theories.

Key insight

Whether a theory survives depends on which phenomena it is tested on. The phenomena that discriminate say that avoidance in these tasks is inferential: it is learned without terminating the warning signal, it transfers to stimuli that predict the same harm without having been trained on them, and it can arise from anticipation alone. “As a whole, these findings support an inference-based view” (p. 25). One robust phenomenon cuts the other way: some avoidance persists after the harm is visibly removed, which inference does not predict. “Yet, the evidence also contradicts a purely inference-based account” (p. 25).

Assumptions

• Inclusion (p. 7). Published, peer-reviewed, in English; “(1) experimental procedures involving healthy humans”, individual acquisition of an avoidance response, a primary reinforcer (electric shock, loud or noxious sound; money and points excluded), no appetitive reinforcer during acquisition, and at least 20 participants. Healthy control groups from clinical studies were included; patients were not.
• Search (p. 6). From August 2023, last updated 12 December 2024. One author screened; the other was consulted in uncertain cases (p. 7).
• Theories compared (pp. 4–6; Fig. 1). Mowrer’s two-factor theory (2FT), safety-signal theory, Seligman and Johnston’s hybrid model, the negative-occasion-setter account, and Lovibond’s expectancy theory. Computational instantiations exist only for 2FT (Moutoussis et al. 2008; Maia 2010), both model-free; “there is no computational counterpart of expectancy theory” (p. 6).
• Robustness ratings (Fig. 4). By diversity of settings, number of studies, confounds and known boundary conditions. A or B marks a candidate benchmark for models.

Key results

• The sample (pp. 9–10). 116 articles, 1939–2024, 15 preregistered. Median sample 50 (20–189, plus one of 502); mostly undergraduates, median age 21.25, 58.8% women. Electric shock in 70 studies, noise in 25. 59 studies used a single avoidance response, usually instructed. “The main concern was ecological validity” (p. 10): the correct avoidance response was typically instructed, “which qualitatively changes the problem setting”.
• Avoidance without CS termination (p. 12; rated A, 8 studies). 2FT predicts no learning if the response cancels the US but not the CS. Eight studies falsified this; “We did not identify any evidence confirming this prediction of the original 2FT.”
• Conditioned responses (pp. 10–11; A, 11 studies). Skin conductance falls during avoidance acquisition, but it also falls with unavoidable shocks, and the comparisons are mixed. Two experiments (Ng & Lovibond) show avoidance intention lowering skin conductance: “The latter findings are not predicted by 2FT, but are in line with expectancy theory” (p. 11). Fear responses are not needed to maintain avoidance (p. 25).
• Generalisation and transfer (pp. 12–15). Perceptual generalisation of avoidance (A); latent transfer through learned stimulus equivalence (A, 5 studies); and US-based transfer, an untrained response used for a new CS predicting the same harm (B/C, 3 studies, holding only for those who understood the contingencies).
• Spontaneous avoidance after Pavlovian learning only (p. 16; A). In virtual-reality studies participants kept their distance from a conditioned stimulus never trained for avoidance; in one, approach was slower even after the electrodes were visibly detached. “This suggests that the anticipation of a US is sufficient to elicit an AR” (p. 17).
• Partial effectiveness (pp. 15–16). Responses that only sometimes work are emitted almost as often as fully effective ones unless nearly useless. People also keep responding to a CS whose US is unavoidable, declining over trials, which no theory explains “unless they incorporate some notion of outcome uncertainty” (p. 16).
• Protection from extinction (pp. 18–19; A, 9 studies). With the response available during extinction, “42% (Malloy & Levis, 1988) and 47% (Williams & Levis, 1991) of participants did not extinguish a unary and costless AR by the 500th would-be-unpunished trial” (p. 18). In two studies, those allowed to avoid “continue to do so and retain higher US expectancies in a follow-up test where avoidance is precluded” (p. 18). Less effective or more costly responses extinguish faster (A, 6 studies). Monetary incentives reduce avoidance during extinction, but when avoidance is then prevented, US expectancy rises in every group regardless of intervention (pp. 18–19).
• Return after response prevention (pp. 19–20; A/B, 7 studies). When the response is made available again after extinction with response prevention, it is used, and removing it again raises threat expectancy. Every such study signalled response availability, which may itself signal danger, and none extinguished to criterion.
• Avoidance spreads to the safe stimulus (pp. 19–20, 26; B, 3 studies). Signalling a costless response for a stimulus never paired with shock (the CS−) leads people to use it, and removing it then raises their threat expectancy for that stimulus, though skin conductance and startle were unaffected in one study. This “must invoke an inference mechanism that deduces CS–US contingencies without direct experience” (p. 26), which a model-free actor-critic cannot do, since its critic “only updates values for states visited by the actor” (p. 26).
• Signalled US omission (pp. 21–23; A, 8 studies). Avoidance persists, at a reduced rate, after the US is visibly removed. Response distributions are zero-inflated: a few participants carry the group effect. No study found the value-by-training interaction that would show habits forming with overtraining. Hence “avoidance learning can not be fully inferential in the first place, directly challenging expectancy theory” (p. 23).
• Neuroimaging (pp. 23–24). Thirteen fMRI and six EEG studies, small (fMRI median 25). Three findings replicated at least once: caudate activity in threat against motor control; periaqueductal gray and anterior insula up and vmPFC down as threat comes closer; PAG activity consistent with an aversive prediction error.
• The authors’ synthesis (pp. 25–28). Expectancy theory is best supported, tempered by Pavlovian influences. Latent-cause inference could explain context effects and generalisation. On persistence: “more effective ARs are harder to extinguish when the response is not prevented, putatively because effective ARs provide fewer opportunities to detect changes in contingency (i.e., the absence of punishment when the AR is not performed), causing slower learning” (p. 28). One explanation for avoidance raising threat expectancy for safe stimuli is that people “might counterfactually infer danger for the safe CS, through the artificial belief that if they had not avoided, when they had the chance, they would have been punished” (p. 26).
• Methods (pp. 28–31). Current paradigms are “structurally less complex than real-world settings” (p. 2); the authors want multi-step tasks, continuous rather than button-press responses, quantified response costs, trial-by-trial individual modelling, and analysis of participants who misunderstand the task: “It is indeed possible that in real-world scenarios, excessive avoidance might stem from a misinterpretation of these contingencies” (p. 31).

Claims

id	claim	strength	support
C1	Avoidance is acquired without the response terminating the CS	strong	8 studies, no counterevidence; rated A
C2	Avoidance transfers by inference to stimuli never trained with it	strong for latent transfer; moderate for US-based transfer	5 and 3 studies; A and B/C
C3	Available avoidance persists through extinction and keeps US expectancy higher; more effective responses persist longer	strong	9 and 6 studies; A
C4	Avoidance offered for a safe stimulus raises threat expectancy for it when withdrawn	moderate	3 studies; B; confounded by signalling
C5	Some avoidance persists after signalled US omission without habits forming by overtraining	moderate	8 studies, zero-inflated; A
C6	The evidence contradicts associative (two-factor) theories and favours inference-based ones, but not a purely inferential account	moderate	the authors’ scoring of Fig. 1 against the phenomena
C7	Effective avoidance persists because it leaves fewer chances to detect a change in contingency	weak: the authors’ “putatively”	interpretation of C3

Concepts

• AR, CS, US, CR: the avoidance response; the conditioned stimulus that warns of the unconditioned (aversive) stimulus; and the involuntary conditioned response, often called fear (p. 4).
• RPE (response prevention extinction): extinction in which both the US and the avoidance response are removed (p. 18), the laboratory analogue of exposure with response prevention.
• Extinction with AR: extinction in which the response stays available and simply is not punished if omitted (p. 18).
• Signalled US omission: removing the outcome visibly (unplugging the electrodes) to test whether responding depends on it (p. 21).
• “Cognitive” against “computational” theory: a theory naming cognitive operations, against an algorithm instantiating them (p. 5).

Connections

It updates Krypotos et al.’s narrative reviews (2015, 2018, 2020) and complements Wong et al. (2023) on individual differences. Its framing pairs two-factor and expectancy theory with model-free and model-based reinforcement learning. It names Moutoussis et al. (2008) and Maia (2010) as the only implemented avoidance models, and points to latent-cause inference (Gershman, Niv) and Bayesian treatments of controllability (Huys & Dayan 2009; Zamfir & Dayan 2022) as where theory should go. It does not cite Zorowitz et al. (LIT-1496).

Bearing on the record

• CLAIM-446 (governance-maintained anxiety). The review is laboratory evidence in healthy people. Within that limit it bears on three of the claim’s steps, and it sharpens one.
  • Avoidance destroys identifiability: supported. The claim’s central mechanism is that, under avoidance, danger and safety predict the same observation. The review finds that available avoidance protects threat expectancy from extinction (rated A), and that the more effective the response, the longer it lasts, which the authors attribute to fewer chances “to detect changes in contingency”. That is the claim’s point in the authors’ words, from the human literature, with their “putatively”.
  • Precaution as evidence of danger: supported, with a confound. The claim names “if I felt compelled to protect myself, the danger must have been serious” as one way the loop closes. The review’s phenomenon of avoidance offered for a never-punished stimulus raising threat expectancy for it when withdrawn (B, three studies) is the nearest laboratory case, and the authors’ first explanation is a counterfactual inference of the same form. The confound they flag is that offering a response may itself signal danger.
  • Belief correction is not behavioural correction: partly supported. Avoidance persists after the outcome is visibly removed (A), without habit formation by overtraining. So some avoidance outlasts a corrected expectation. But a few participants carry the effect, and the review cannot say whether those participants believed the removal.
  • “Avoidance by itself suppresses learning; it does not strengthen the belief”: complicated. The claim says this to separate suppression from self-maintenance. In the review, return of avoidance after response prevention, and its spread to the safe stimulus, raise threat expectancy, so in these tasks the availability or withdrawal of avoidance does move the belief. The claim should either count that as its own closing inference or drop the sentence.
  • Not anxiety. No clinical sample was included, and the paradigms are by the authors’ account simpler than real avoidance. The review cannot say whether any of this is how anxiety is maintained.
• QUESTION-045. The review’s response-prevention studies are the closest existing design to the question’s test of cutting a maintaining link: removing the response during extinction. Their result is mixed: threat ratings fall during prevention but return when the response is offered again. Monetary incentives to refrain cut avoidance more than instructions, but threat expectancy rose in every group when avoidance was then prevented (pp. 18–19).
• Correction to the exchange. The exchange called it a 2026 review giving the strongest support to “expectancy- and inference-based accounts”. It is a 2025 preprint, in print in 2026. It does support expectancy theory by name (the abstract does not), but it also says the evidence “contradicts a purely inference-based account”. Its studies are of healthy people only, with shock or noise.
• Filed: THEORY-901 (persistence and protection from extinction), THEORY-920 (avoidance raising threat expectancy for a safe stimulus) and THEORY-900 (the verdict on theories).
• No instruction for machine-learning practice; nothing for the anthology.

Limitations

• Healthy volunteers only, mostly undergraduates, with primary reinforcers. Nothing here is about anxiety disorders.
• Ecological validity, the authors’ main risk-of-bias concern: instructed, signalled, single-button responses. They argue that such tasks may themselves elicit inference.
• One screener, with the second author for uncertain cases only.
• No meta-analysis. Robustness ratings are judgements; effect sizes are not pooled.
• Supplement and version of record not read. The study list and risk-of-bias ratings are in an OSF data set; the journal version is closed.
• Slips in the preprint. “needed needed” (p. 4); “Notably, both share the feature of signaling AR availability, someting that is unusual” (p. 26); “specific to the paradigms we exist in the literature” (p. 28); Declercq et al., Trick et al. and Garofalo & Robbins are called “both studies” (p. 14). None changes a finding.

Open questions

• Does avoidance persist longer when it is more effective in people with anxiety disorders, and for the same reason (fewer disconfirming observations)? A task that varies effectiveness with threat expectancy measured trial by trial in patients would tell.
• Is the rise in threat expectancy for safe stimuli an inference from one’s own avoidance, or a reading of the experimenter’s signal? An unsignalled, continuously available response would separate them.
• Who are the participants who keep avoiding after signalled omission, and what do they believe about the outcome?