Source-linked AI summary
Signatures of human-like processing in Transformer forward passes
Jennifer Hu, Michael A. Lepori, Michael Franke
TL;DR
Output-level comparisons have left open whether Transformer computation resembles the real-time processing that produces human responses. This paper analyzes layer-wise dynamics across 20 open-source models and 6 domains, finding competitor-interference signatures and improved prediction of human processing beyond static measures. It therefore presents AI models as potential explicit processing models of human cognition, while noting that larger models are not always more human-like.
Problem
It remains unclear whether the internal layer-wise processing of models resembles the cognitive processing involved in producing human responses.
Source: §1 · p. 2Method
The study analyzes layer-wise Transformer dynamics across 20 open-source models and 6 domains, comparing dynamic measures with static output measures.
Source: §1 · p. 2, §3.2.1 · p. 7Results
Takeaways & Limitations
Mechanistic analyses can treat AI systems as potential explicit models of human processing rather than only as black-box mappings from stimuli to responses.
Source: §4 · pp. 9–10Many model–human measure combinations produce R2 values near 0, so strong alignment is not expected for every combination.
Source: §3.2.3 · p. 8
Abstract
from arXiv · showhide
Modern AI models are increasingly being used as theoretical tools to study human cognition. One dominant approach is to evaluate whether human-derived measures are predicted by a model's output: that is, the end-product of a forward pass. However, recent advances in mechanistic interpretability have begun to reveal the internal processes that give rise to model outputs, raising the question of whether models might use human-like processing strategies. Here, we investigate the relationship between real-time processing in humans and layer-time dynamics of computation in Transformers, testing 20 open-source models in 6 domains. We first explore whether forward passes show mechanistic signatures of competitor interference, taking high-level inspiration from cognitive theories. We find that models indeed appear to initially favor a competing incorrect answer in the cases where we would expect decision conflict in humans. We then systematically test whether forward-pass dynamics predict signatures of processing in humans, above and beyond properties of the model's output probability distribution. We find that dynamic measures improve prediction of human processing measures relative to static final-layer measures. Moreover, across our experiments, larger models do not always show more human-like processing patterns. Our work suggests a new way of using AI models to study human cognition: not just as a black box mapping stimuli to responses, but potentially also as explicit processing models.
1 Introduction
The paper asks whether Transformer forward passes reveal human-like processing dynamics, rather than merely matching human behavior through final outputs. Across two experiments, it examines competitor interference and whether layer-wise dynamics predict human processing measures.
1 Introduction
- The paper extends output-level comparisons by using mechanistic interpretability to study internal processes supporting model outputs.Earlier work linked model layers to fast and slow human processing measures, motivating the paper’s central question.Source: §1 · pp. 1–2
- Experiment 1 tests competitor interference, while Experiment 2 quantitatively evaluates whether processing-difficulty measures predict human processing load above static output measures.The research questions also examine how model size affects similarity between model and human processing.Source: §1 · p. 2, Figure 1 · p. 2
- Dynamic forward-pass measures improve prediction of human processing indicators beyond static measures from final or intermediate layers across 20 models and 6 domains.The study covers text and vision models alongside multiple human behavioral measures.Source: §1 · pp. 2–3
- Models initially favor salient competing answers before correct answers on items expected to induce human-like decision conflict.The study interprets this pattern as delayed decision-making and two-stage processing.Source: §1 · p. 2, Figure 1 · p. 2
- Larger models are not always the most predictive of human processing, and mid-sized models show the strongest signs of two-stage processing.These model-size effects qualify the relationship between scale and human-like processing.Source: §1 · p. 2
- The apparent similarity between model and human processing may emerge from general-purpose objectives such as next-token prediction or image recognition.This is presented as a possible explanation rather than an established mechanism.Source: §1 · p. 3
2 Domains for evaluating cognitive processing patterns
The experiments span text and vision domains designed to test cognitive processing patterns, especially competition between salient incorrect and correct answers. Domains vary in task, modality, human response measure, and availability of competitor conditions.
Domains for evaluating cognitive processing patterns
- The domain set includes capitals, animal categorization, gender bias, syllogisms, and object recognition, with Competitor items designed to induce competition between intuitive and correct answers.Object recognition lacks paired intuitive and correct answers and is therefore used only in Experiment 2.Source: §2 · pp. 3–4
Capitals (recall)
- Capital recall uses 62 items, including 42 Competitor cases where a salient incorrect city is the entity’s most populous city and 20 NoCompetitor cases.Human participants typed answers while the study recorded response time, accuracy, and typing-derived measures.Source: §2 · p. 3
Capitals (recognition)
- Capital recognition presents the correct and intuitive options directly and measures human accuracy and response time for 42 Competitor items.Data came from 41 human participants.Source: §2 · p. 3
Animal categories
- Animal categorization contrasts atypical exemplars that compete with the correct category against typical exemplars, using accuracy, response time, and mousetracking measures.The stimuli include 6 Competitor and 13 NoCompetitor items, with data from 108 participants.Source: §2 · p. 3
Gender bias
- Gender-bias stimuli manipulate contextual cues so that Competitor items conflict with prevalent profession–gender associations, but the dataset has no associated human data.Consequently, this domain is used only in Experiment 1.Source: §2 · pp. 3–4
Capitals (recall)
- One typing-derived dependent variable was recorded for only 30 of the 56 capital-recall participants because of technical difficulties.This constrains interpretation of that particular measure rather than the entire dataset.Source: §2 · p. 3
Syllogisms
- Syllogism items test logical validity, with Competitor cases creating conflict between logical validity and prior beliefs; human accuracy and response time are measured.The set contains 48 Competitor and 144 NoCompetitor stimuli.Source: §2 · p. 4
Object recognition
- Object recognition tests out-of-distribution vision performance across 17 image datasets and 16 ImageNet categories, using human accuracy and response time.Because these stimuli lack paired correct and intuitive answers, the domain is an exploratory Experiment 2 generalization test.Source: §2 · p. 4
3 Investigating cognitive processing strategies in a forward pass
The experiments use layer-wise Transformer dynamics to identify competitor interference and test whether dynamic processing measures predict human processing beyond static outputs. Across 20 models and multiple domains, forward-pass dynamics capture human-processing patterns, but model size has task-dependent effects.
3 Investigating cognitive processing strategies in a forward pass
- The study evaluates 18 autoregressive language models and two vision Transformers, spanning 124M to 405B parameters, with logit-lens readouts checked against tuned-lens results.The evaluation covers six model families for text and two ViT models for vision.Source: §3 · pp. 4–5
3.1 Experiment 1: Competitor interference effects in LMs
- Models generally show stronger competitor-interference signatures for items with salient intuitive answers, although COM is an exception in gender-bias and syllogism domains.Both COM and TTD are generally higher in Competitor than NoCompetitor conditions.Source: §3.1 · p. 5, Figure 2 · p. 6
- Example layer trajectories show distinct strategies: intermediate-sized OLMo-2 exhibits two-stage processing, GPT-2 favors the intuitive answer, and Llama-3.1 405B favors the correct answer.In NoCompetitor items, all three example models consistently favor the correct answer across layers.Source: §3.1 · p. 6
3.2.1 Candidate metrics
- Processing dynamics are derived from layer-wise distributions using static single-layer measures and dynamic aggregates or change-point measures such as AUC and argmax-delta.Dynamic measures summarize influence over layers or the layer where a metric changes most rapidly.Source: §3.2.1 · pp. 6–7
3.2.2 Analyses
- The predictive analyses use item-level R2 and Bayes-factor comparisons between baseline static-output regressions and models augmented with processing measures.Human accuracy and processing-related dependent variables are modeled with participant random intercepts.Source: §3.2.2 · p. 7
- Competitor-interference metrics often correlate strongly with human dependent variables and substantially improve prediction beyond output measures.The analyses compare baseline mixed-effects models using final-layer static predictors with critical models adding a dynamic measure.Source: §3.2.2 · p. 7, §3.2.3 · p. 8
3.2.3 Results
- Dynamic measures improve prediction of human processing measures beyond static final-layer outputs, including for vision and midpoint-layer baselines.Critical models frequently improved over baselines, with log BF > 0.Source: §3.2.2 · p. 7, §3.2.3 · p. 8
- Model size does not uniformly increase human-like predictive performance; its effects vary by task and human dependent variable.Larger models help for some capitals and syllogism measures, hurt performance for animal categorization, and show diminishing returns for capitals recall.Source: §3.2.3 · pp. 8–9
4 General discussion
The study finds high-level alignment between human and machine processing, while showing that forward-pass dynamics can characterize model difficulty and predict human processing beyond static outputs. It presents this as a foundation for treating AI systems as explicit processing models, while leaving finer-grained temporal comparisons for future work.
4 General discussion
- Processing metrics from forward-pass dynamics predicted human accuracy and processing measures beyond static final-layer logits across models, domains, task modalities, and input modalities.Larger models were not always most predictive of human processing, and mid-sized models were most likely to show competitor interference effects.Source: §4 · p. 9
- The approach compares high-level summary statistics of model processing trajectories with human processing rather than matching specific time slices or neural pathways.This supports comparisons on arbitrarily complex inputs whose specific algorithms or brain pathways may be unclear.Source: §4 · p. 9
- Processing metrics could characterize model difficulty and support evaluations with better construct validity or more efficient early exiting.The paper frames these metrics as candidate measures of load within models.Source: §4 · p. 9
- The findings provide a proof-of-concept for high-level alignment between human and machine processing and motivate studying AI systems as explicit models of human processing.Proposed directions include testing architectural or training changes and generating predictions about human difficulty in novel settings.Source: §4 · p. 9
- The work is presented as a first step toward using mechanistic analyses of AI models to understand processing in the human mind.The authors also identify whether a large pretrained model resembles an individual processing stream or an aggregate across a population as an open question.Source: §4 · p. 10
A.1 Human experiments
The human experiments measured accuracy, response time, and task-specific behavioral indicators of processing load across recall, recognition, categorization, syllogism, and object-recognition tasks. Data came from new Prolific samples and released datasets, with task-specific preprocessing and model-comparison measures.
Capitals recall
- Capitals recall used 56 English-speaking US participants who freely typed answers, with keystrokes and timestamps logged for response and typing analyses.Trials with extreme response times or fewer recorded keystrokes than final-answer characters were excluded.Source: §A.1 · pp. 17–18
Capitals recognition
- Capitals recognition used 41 participants who chose between correct and intuitive answers with randomized answer-side placement, predicting only accuracy and response time.Trials with response times more than two standard deviations from the mean were excluded.Source: §A.1 · p. 18
Animal categories
- Animal categorization used 19 exemplars and 108 participants, contrasting atypical competitor items with typical no-competitor items using mouse trajectories.Participants were divided between click and touch groups, which was included as a fixed factor.Source: §A.1 · p. 18
Human behavioral measures
- Human processing load was measured with response time and behavioral trajectories or typing patterns, alongside accuracy.Measures included mouse AUC, MAD, x-axis flips, maximum acceleration time, keypress timing, extra keypresses, and backspaces.Source: §A.1 · p. 18
Syllogisms
- Syllogism data comprised 192 items varying realistic versus nonsense content and whether validity conflicted with prior beliefs, with accuracy and response time analyzed.Because subject identifiers were unavailable, analyses included random intercepts for item variations.Source: §A.1 · pp. 18–19
Object recognition
- Object recognition compared humans and pretrained vision Transformers on 17 out-of-distribution image datasets spanning 16 ImageNet categories and varied image manipulations.Condition variables accounted for the 12 parametric degradation datasets but not the five non-parametric manipulations.Source: §A.1 · p. 19
A.2 Model evaluation
Model evaluation used prompts tailored to six domains and intermediate-layer readouts to compare processing dynamics with human measures. The study included language and vision models, with forced-choice inputs averaged across answer orderings where applicable.
Implementation
- The model inventory and evaluation setup were documented in Table 3, with inference run using nnsight on NVIDIA A100 and H100 GPUs.Inference was conducted during April–May 2025.Source: §A.2 · p. 19, Table 3 · p. 20
Text-domain evaluation
- Capitals recognition averaged metrics across contexts placing the correct and incorrect options in opposite orders.This controlled for answer-order presentation in the forced-choice model evaluation.Source: §A.2 · p. 19
- The evaluation covered domain-specific inputs for gender bias, syllogisms, capitals recall and recognition, and animal categories.Animal-category prompts used “A” or “An” according to the exemplar’s first letter, while syllogism prompts explicitly requested valid or invalid judgments.Source: §A.2 · pp. 19–20
A.3 Annotating responses in capitals recall experiment
The capitals recall responses were labeled by GPT-4o using a fixed five-category scheme that distinguishes correct, intuitive, alternate-city, uncertain, and other answers. The resulting label distribution is reported in Figure 5.
Labeling scheme
- GPT-4o labeled each response as Correct, Intuitive, Alternate city, Not sure, or Other.The categories were intended to distinguish exact or near-correct answers, intuitive competitors, other cities, uncertainty, and miscellaneous responses.Source: §A.3 · pp. 20–21
- The annotation prompt explicitly allowed minor typos and spelling variations when assigning the best-fitting label.Examples included capitalization changes and misspellings.Source: §A.3 · pp. 20–21
Annotation results
- Figure 5 shows the distribution of GPT-4o-assigned labels for free responses in the capitals recall experiment.The figure summarizes the output of the annotation procedure rather than participant response times or keystroke measures.Source: Figure 5 · p. 21, §A.3 · p. 21
Annotation procedure
- For each trial, the prompt substituted the actual entity, correct answer, intuitive answer, and human response into the labeling template.This made the annotation decision conditional on the item-specific alternatives and response.Source: §A.3 · p. 21
B Details of model-derived metrics
The paper defines static and layer-time measures of uncertainty, confidence, relative confidence, and answer boosting, then uses regression models to compare their predictive value.
B.1 Uncertainty
- Uncertainty is measured by next-token entropy, with final-layer entropy as the output measure and summed entropy or largest-decrease layer as processing measures.The entropy depends only on context, not on a particular answer option.Source: §B.1 · p. 22
B.2 Confidence in correct answer
- Confidence in the correct answer uses first-token log probability and reciprocal rank, with final-layer, area-under-curve, and characteristic-layer variants.Reciprocal rank ranges from 1 for the top-ranked token to 1/|V| for the bottom-ranked token.Source: §B.2 · p. 22
B.3 Confidence in correct answer, relative to incorrect answer
- Relative confidence is the correct-minus-incorrect log-probability difference, summarized by final-layer value, signed AUCs, and the largest-increase layer.The positive and negative AUCs separately quantify preference for the correct and incorrect answers and are not redundant.Source: §B.3 · p. 22, Figure 6 · p. 23
B.1 Uncertainty
- Layer-time measures complement final- or midpoint-layer values by aggregating trajectories or identifying layers with the largest changes.Dynamic measures use area under curves or characteristic layer indices; signed measures separately track positive and negative regions.Source: §B.1 · p. 22, §B.2 · p. 22, §B.3 · p. 22
B.4 Boosting of correct answer, relative to incorrect answer
- Boosting is estimated from residual-stream layer changes projected into vocabulary space, capturing whether layers promote the correct answer over the incorrect answer.The projection along ⟨1,1⟩ represents increasing the correct answer’s probability by increasing its log probability rather than decreasing the incorrect answer’s.Source: §B.4 · p. 23
- Boosting dynamics use positive and negative projection areas plus the layer of maximum projection, rather than final-layer output measures.The maximum projection is used because the projection itself already measures change.Source: §B.4 · pp. 23–24
C Details of regression analyses
- Model comparisons use centered and scaled linear or generalized linear mixed-effects models, with distributions chosen for continuous, binary, count, and time-based outcomes.Time-based dependent variables are log-transformed, and predictors with no item-level variation are dropped for the relevant model and task.Source: C Details of regression analyses · p. 24
D.1 Task accuracy
Task accuracy is evaluated from final-layer outputs, while Experiment 1 reports model-level competitor-interference measures across tasks.
D.1 Task accuracy
- Final-layer accuracy compares total answer probability for text tasks and highest category probability for vision tasks against the correct answer.Figure 7 reports accuracy for each model and humans on each task.Source: §D.1 · p. 24
D.2 Experiment 1: Competitor interference effects by model
- Figure 8 shows each model’s competitor-interference measures, COM and TTD.Source: §D.2 · p. 24
D.3 Experiment 2: Comparison of logit lens and tuned lens
The paper compares logit-lens and tuned-lens predictions of human dependent variables and focuses on logit-lens results because their distributions are largely similar.
D.3 Experiment 2: Comparison of logit lens and tuned lens
- Logit-lens and tuned-lens models achieve largely similar distributions of R2 values when predicting human dependent variables.The main text therefore focuses on logit-lens results for models with pretrained tuned lenses.Source: §D.3 · p. 24
D.4 Experiment 2: Results from vision domain
Experiment 2 includes a vision-domain analysis of its results from object-recognition datasets.
D.4 Experiment 2: Results from vision domain
- Figure 10 reports the Experiment 2 results from the object-recognition datasets.Source: §D.4 · p. 24
D.5 Experiment 2: Log Bayes factors with respect to midpoint-layer baseline
Experiment 2 compares dynamic processing measures against static midpoint-layer baselines when predicting human processing measures across text and vision domains.
D.5 Experiment 2: Log Bayes factors with respect to midpoint-layer baseline
- Log Bayes Factors compare critical regression models with static midpoint-layer baselines across text and vision domains.The midpoint-layer baseline provides the reference for assessing whether processing dynamics add predictive value.Source: §D.5 · p. 24, Figure 11 · p. 27
- Model accuracy is reported separately for text-based and vision-based tasks, with task-specific gaps in available human or model data.The accuracy figure notes missing human gender-bias data and missing Llama-3.1 405B syllogism data.Source: Figure 7 · p. 25
- The broader experiments evaluate competitor-interference measures and other processing dynamics against human processing indicators across multiple domains and model representations.Competitor effects are summarized with change-of-mind and time-to-decision measures, while R2 distributions compare logit and tuned lenses.Source: Figure 8 · p. 26, Figure 9 · p. 26
- R2 summarizes how well model processing measures predict groups of human dependent variables in the vision domain.Figure 10 reports model-level R2 and mean R2, alongside final-layer log Bayes Factors.Source: Figure 10 · p. 26