ARC: a framework for access, reciprocity and conduct in psychedelic therapies

Research into LSD as a potential psychiatric treatment has re-emerged, with theorised benefits centred on effects on learning and neural plasticity. The serotonin 2A (5-HT2A) receptor is considered a key target mediating psychedelic effects and plasticity, and animal studies have shown that 5-HT2A agonists and LSD can influence associative learning. Serotonin and dopamine systems are both implicated in adaptive behavioural flexibility and reversal learning, and prior human studies of LSD and other psychedelics have examined a range of cognitive and affective domains but have not comprehensively addressed the mechanisms of instrumental learning and probabilistic choice under uncertainty in humans. The authors set out to test how LSD alters instrumental conditioning and cognitive flexibility using a probabilistic reversal learning (PRL) paradigm in healthy volunteers. They aimed both to assess overt behavioural changes using standard measures (for example win-stay/lose-stay and perseveration) and to probe underlying learning mechanisms by fitting computational reinforcement-learning (RL) models. Specific questions included whether LSD modulates sensitivity to immediate feedback, changes the rate at which choice values are updated (learning rates), and alters exploratory behaviour (indexed as reinforcement sensitivity and stimulus stickiness).

Nineteen healthy volunteers (mean age 30.6 years; 15 males) completed a single-blind, within-subjects, balanced-order study in which they attended two sessions at least two weeks apart and received intravenous LSD (75 μg in 10 mL saline) or placebo (10 mL saline). One participant from the original sample of 20 did not complete the PRL task and was excluded; all participants had prior experience with a classic psychedelic without adverse reaction and met medical, psychiatric, and drug-use screening criteria. Participants were blinded to condition but experimenters were not. A cannula was used for a two-minute injection; participants reported subjective effects 5–15 minutes after dosing and the PRL task was administered approximately five hours after injection. The study formed part of a larger project; supplementary demographic and procedural details were reported elsewhere. The behavioural task was a three-choice probabilistic reversal learning paradigm of 80 trials (40 acquisition, 40 reversal). On each trial participants chose one of three visual stimuli. During acquisition one stimulus yielded positive feedback on 75% of trials, a second on 50% (neutral), and a third on 25%; after 40 trials the 75% and 25% contingencies were reversed. Positive and negative feedback were signalled by emotive faces. Raw behavioural measures included the number of responses to each stimulus, win-stay and lose-stay probabilities (defined as the proportion of occasions a choice was repeated following a win or loss, respectively), and perseverative errors in the reversal phase (defined as two or more responses to the formerly correct stimulus after reversal, with the first reversal trial excluded). Statistical testing used conventional null-hypothesis methods with α = 0.05 for raw measures. To characterise latent learning processes, the authors fitted three reinforcement-learning models to trial-by-trial choices using a hierarchical Bayesian approach implemented with Hamiltonian Markov chain Monte Carlo sampling in Stan. Models were compared with bridge sampling to estimate marginal likelihoods and posterior model probabilities. Convergence was checked with the potential scale reduction factor R (< 1.2 criterion). The hierarchical structure modelled drug condition at the group level and subjects below that, allowing within-subject comparisons of parameter means between LSD and placebo. Posterior parameter differences were interpreted using highest posterior density intervals (HDIs). The three candidate models comprised: (1) separate reward and punishment learning rates plus a reinforcement-sensitivity parameter; (2) a single learning rate plus a stimulus stickiness parameter capturing outcome-independent tendency to repeat the previous choice; and (3) a full model combining separate reward and punishment learning rates with stimulus stickiness. Key parameters were reward learning rate (α_rew), punishment learning rate (α_pun), reinforcement sensitivity (τ_reinf; akin to value-based exploitation), and stimulus stickiness (τ_stim; value-free tendency to repeat choices).

On overt task performance, repeated-measures ANOVA showed a strong main effect of stimulus type and a stimulus-by-phase interaction, indicating participants chose according to the reinforcement probabilities and adapted after reversal. There was no overall effect of LSD on the number of correct responses during acquisition (paired t test, t18 = 0.84, p = 0.4) or reversal (t18 = 0.23, p = 0.8), and no reliable interactions between LSD and stimulus or phase in the ANOVA. Examining the relationship between initial learning and perseveration, LSD strengthened the coupling between acquisition performance and subsequent perseverative errors: the change in acquisition correct responses (LSD minus placebo) predicted the change in reversal perseverative errors (LSD minus placebo) with a regression coefficient β = 0.56 (p = 0.002). Under LSD fewer acquisition errors predicted more perseverative errors (β = 0.44, p = 0.003), whereas no such relationship was present under placebo. However, perseverative errors considered alone did not differ between conditions (t18 = 0.03, p = 0.98). For immediate feedback sensitivity, participants stayed more after wins than losses (main effect of valence, F1,18 = 37.76, p = 8.0 × 10^-6) but LSD had no main effect and did not interact with valence (no significant drug or valence×drug effects). Model comparison showed good convergence for all three RL models (R < 1.2) and that behaviour was best captured by the full four-parameter model (separate reward and punishment learning rates, reinforcement sensitivity, stimulus stickiness). Model-derived parameter effects across the full 80 trials were: a substantial increase in reward learning rate under LSD (mean LSD 0.87 v. placebo 0.28), with the posterior 99.9% HDI for the difference excluding zero. Punishment learning rate was also elevated (LSD mean 0.48 v. placebo 0.39) with the 99% HDI excluding zero. Net, LSD increased reward learning more than punishment learning (difference excluded zero at high credibility). Phase-specific modelling indicated that during acquisition the reward learning rate was markedly higher under LSD (mean 0.72 v. placebo 0.17; 99% HDI excluding zero) while punishment learning during acquisition was not significantly elevated. During reversal both reward (LSD mean 0.96 v. placebo 0.77) and punishment learning rates (LSD mean 0.42 v. placebo 0.31) were elevated under LSD (drug differences credible at the 90% HDI), and during reversal there was no credible difference in LSD's effect on reward versus punishment rates. Regarding exploratory/exploitative parameters, modelling the entire task showed stimulus stickiness was decreased by LSD (LSD mean 0.23 v. placebo 0.43; drug difference credible at the 90% HDI), indicating increased value-free exploration. Reinforcement sensitivity across the full task was not credibly changed (LSD mean 4.70 v. placebo 5.57; 95% HDI included zero). Phase-wise, during acquisition LSD reduced both stimulus stickiness (0.09 v. 0.46) and reinforcement sensitivity (4.92 v. 6.54), implying increased exploration by both value-free and value-based metrics. During reversal, stimulus stickiness remained lower under LSD (0.36 v. 0.58) but reinforcement sensitivity was increased under LSD (3.64 v. 2.47), suggesting greater trial-by-trial value-driven exploitation in the reversal phase. Linking model parameters to raw measures, a higher reward learning rate during acquisition under LSD predicted more perseverative errors (β = 26.94, p = 0.02), an effect not present under placebo. Stimulus stickiness during reversal did not significantly correlate with perseveration in either condition. Additional exploratory analyses were reported in supplementary materials.

The authors interpret their findings as showing that LSD enhances the rate at which human participants update value representations following prediction errors, particularly for rewards, and promotes exploratory behaviour. They emphasise that increased learning rates are compatible with theoretical accounts that psychedelics relax prior beliefs and thereby increase sensitivity to new information, a mechanism that could plausibly support revision of entrenched maladaptive associations in psychiatric disorders. Two distinct forms of increased exploration were reported: a value-free reduction in stimulus stickiness (greater tendency to switch irrespective of outcome), present across acquisition and reversal, and phase-dependent changes in value-based exploration (reinforcement sensitivity was reduced during acquisition but increased during reversal). The reward learning rate was elevated in both acquisition and reversal, whereas punishment learning rate increases were most evident in the reversal phase. The authors note that under LSD better initial acquisition predicted greater perseverative responding upon reversal, suggesting that associations formed during LSD may become more strongly expressed and harder to update if both learning and later testing occur under the drug. The discussion recognises uncertainty about the precise neurochemical mechanisms, noting candidate roles for 5-HT2A-mediated serotonergic effects and for dopaminergic influences (including interactions between the two systems). The authors cite converging animal and genetic findings that link serotonergic and dopaminergic manipulations to changes in learning rates and stimulus stickiness, but stress that LSD acts at multiple receptor types and that the present design cannot dissociate these. They also highlight timing as a critical factor: when acquisition occurs before drug administration, LSD (or 5-HT2A modulation) can improve subsequent reversal learning, whereas when acquisition and reversal are both conducted under the drug, newly formed priors may be reinforced. Limitations acknowledged by the authors include the inability to determine receptor-specific causes because LSD binds many targets beyond 5-HT2A, the absence of concurrent measurements of subjective effects or plasma LSD levels at the time of task administration, and an inability to reproduce the learning–perseveration relationship in simulated data despite successful parameter recovery analyses. The authors conclude that the principal finding is that LSD enhances feedback-driven belief updating (most strongly for reward) and increases exploration, with potential relevance for understanding how LSD might be harnessed therapeutically to revise harmful associations.