Trial-history biases in evidence accumulation can give rise to apparent lapses in decision-making

Perceptual decision-makers often deviate from optimal decision-theory, showing dependencies on recent trial history and making evidence-insensitive “lapse” errors. History biases are typically modeled as additive biases on evidence and thought to be overcome by strong stimuli, whereas lapses are assumed to be evidence-independent errors from nuisance processes (e.g., inattention or motor mistakes). Emerging findings show covariation between history dependence and lapses and overlapping neural substrates, suggesting a link. This work asks whether a single mechanism—normative accumulation under misbeliefs of non-stationarity—can produce both phenomena. The hypothesis is that trial-to-trial updates to the initial state of a drift-diffusion accumulator, driven by recent choices and outcomes, yield history-dependent biases and apparent evidence-insensitive errors when choices are pooled across trials.

Prior studies report sequential biases in 2AFC tasks across species and model them via DDM initial state shifts or logistic additive biases. Lapses have been treated as evidence-independent errors, often attributed to exploration or inattention. Reports show that learning reduces both history dependence and lapse probabilities and that lapses can occur in runs rather than independently. Neural perturbations (secondary motor cortex, striatum) affect both lapse rates and trial-history influences, hinting at shared substrates. Logistic psychometric fits are widely used; however, they break down when history-induced initial-state deviations are large, leading to heavy-tailed curves and apparent lapses. Related theoretical work on non-stationary priors and variable precision also yields heavy-tailed psychometrics, but links between sequential biases and lapse rates have been underexplored. This study integrates these literatures by proposing history-dependent initial states as a unified source of history biases and apparent lapses.

Participants: 152 adult male Long Evans rats performed a fixed-duration auditory evidence accumulation ('Poisson Clicks') task; a subset of 6 rats performed a novel reaction-time (RT) variant. Inclusion criteria ensured asymptotic performance and stable accuracy. Task: Two simultaneous Poisson click trains from left/right speakers; in fixed-duration task, rats maintain fixation during stimulus and choose the side with more clicks; in RT task, stimulus plays during fixation and stops when fixation breaks, and rats report which side had the higher Poisson rate. Model: A drift-diffusion framework with trial-history-dependent initial states (HISt). Across-trial updating uses a sum of exponential filters over choice-outcome pairs h ∈ {Rw, Lw, Rl, Ll}: f^h(n) = β^h f^h(n−1) + η 1_{o_{n−1}=h}; I(n) = Σ_h f^h(n). Within-trial, the accumulator evolves with pulse-based evidence, sensory adaptation (strength φ, timescale τ_φ), sensory noise, accumulator noise, leak/feedback (λ), and bounds ±B, numerically solved via Fokker-Planck methods adapted to discrete pulses. Choices are modeled as the probability that x(T) exceeds bias; true lapses occur on a fraction κ of trials via (i) motor error/exploration (random biased choices, parameter p), (ii) inattention (deterministic choice toward sign(i(n)−p)), or (iii) a hybrid sigmoidal dependence on initial state. Psychometric curves: 4-parameter logistic fits (K0, K1, threshold x0, sensitivity b) to characterize asymptotes (lapse rates), threshold, and slope. History modulation metrics: threshold modulation = x0^{Rw}−x0^{Lw}; lapse modulation = 2(K0^{Rw}−K0^{Lw}) + (K1^{Rw}−K1^{Lw}). Model fitting: Maximum likelihood estimation (Julia, Optim, automatic differentiation) on individual rats, with Bayes Information Criterion (BIC) for model comparison. Joint choice+RT modeling: RTs decomposed into decision time plus non-decision time (NDT). NDTs modeled by inverse Gaussian distributions from a separate drift-diffusion with bound ω_k and drift v_k for each choice k ∈ {L,R}, modulated by trial number (α) and previous trial’s outcome (γ). Likelihood integrates accumulation hitting-time distributions with NDT to fit choices and RTs.

• Rats (n=152) showed high accuracy in the Poisson Clicks task (mean 0.79 ± 0.04) with stable performance across trials and exhibited history dependence, typically win-stay/lose-switch. • History-conditioned psychometrics revealed significant comodulations: threshold and lapse rates both shifted with recent wins (e.g., following right wins, both bias and lapse parameters favored right choices). At the population level, sensitivity was unaffected (p=0.8), but threshold (p=3×10⁻⁴) and lapse rates were significantly modulated (left lapse p=8×10⁻⁴; right lapse p=6×10⁻⁷; n=152, two-sided Mann-Whitney U-test). • Across rats, threshold and lapse rate modulations covaried (Pearson r = −0.35, p = 7.28×10⁻⁵), consistent with a shared mechanism. On average, 17 ± 12% of lapses were modulated by trial history, suggesting many are apparent rather than true lapses. • Model comparison strongly favored the accumulator with HISt over the no-HISt variant in 147/152 rats (lower mean per-trial BIC: HISt 0.91 ± 0.01 vs no-HISt 0.93 ± 0.01; paired t-test p = 9.85×10⁻¹⁸). • The HISt model captured individual variability in history modulations: R² = 0.72 (threshold) with slope ≈ 1.02, and R² = 0.69 (lapse rate) with slope ≈ 0.70, comparing model-predicted vs empirical history modulations. • In the RT task (6 rats, 223,231 trials; average accuracy 0.75 ± 0.02), choices showed similar history-dependent threshold and lapse modulations (p=0.69 sensitivity; p=0.004 threshold; p=0.02 left lapse; p=0.02 right lapse), supporting generalization of the HISt framework. • RT signatures matched HISt predictions: error RTs were shorter than correct RTs; RTs after a win were shorter when the stimulus favored the previously rewarded side; history dependence was strongest for fast RTs and waned or reversed for long RTs. • Joint choice+RT fits with HISt reproduced both psychometrics and RT patterns, capturing substantial variance in history-dependent threshold and lapse modulations. Hybrid true-lapse variants slightly improved lapse-rate correspondence but at the expense of threshold correspondence, reinforcing apparent lapses via HISt as the primary driver.

The findings support a unified cognitive mechanism in which misbeliefs about non-stationarity lead to trial-by-trial updates of the initial state of an evidence accumulator, producing both history biases and apparent lapses. Pooling choices across variable initial states yields heavy-tailed psychometric curves that, when fit with logistic functions, manifest as elevated lapse rates. Comodulations of threshold and lapse parameters emerge from the same accumulation variables and are constrained by within-trial parameters (bounds, sensory noise, integration strategy). The model reconciles reports of co-occurring history dependence and lapses, explains modulation of lapse rates by sensory uncertainty, and accounts for RT signatures diagnostic of initial-state variability (shorter error RTs, faster RTs when prior aligns with current stimulus, and stronger history effects for fast decisions). While true lapses (motor error or inattention) contribute, the major share of history-dependent comodulations in thresholds and lapse rates is captured by apparent lapses from HISt. The results suggest that neural circuits involved in integrating prior information and action selection (e.g., secondary motor cortex, dorsomedial striatum) may jointly influence both suboptimalities through their impact on initial accumulator states.

History-dependent updates to the initial state of an accumulator under non-stationary beliefs provide a common explanation for trial-history biases and a substantial fraction of apparent lapses in perceptual decision-making. In large-scale rat datasets, this mechanism quantitatively accounts for history-conditioned changes in psychometric thresholds and lapse rates and predicts distinctive reaction-time signatures. Joint modeling of choices and RTs demonstrates that decisions previously considered stochastic can be made predictable by tracking initial-state dynamics. Future research should probe animals’ implicit models of non-stationary priors, disentangle contributions of potential drift-rate variability across species, and use targeted neural manipulations to identify circuits mediating history-dependent initial states and their interaction with sensory evidence.

• Species and sex: Data come exclusively from adult male Long Evans rats; female rats were excluded due to housing-related aggression, limiting generalizability across sexes and species. • Model scope: The primary mechanism focuses on history-dependent initial states; alternative processes (e.g., drift-rate variability) may dominate in other datasets (humans/primates) where error RTs are longer than correct RTs. • Task design: Fixed-duration and RT variants may not capture settings explicitly designed to counter history biases; lapses in such tasks may involve different mixtures of true lapse processes. • Psychometric fitting: Logistic approximations can obscure heavy-tailed structure arising from initial-state variability; parameter recovery may be sensitive to stimulus range. • Neural inference: While behavioral signatures suggest shared substrates, causal neural mechanisms remain to be directly established; multiple independent circuits may contribute to distinct components (apparent vs true lapses).