PaperlayerView exact revision on arXiv

Source-linked AI summary

The Machine's Internal Clock: Do LLMs Share Human Temporal Illusions?

Catherine Bao, Vivek Srikumar

arXiv:2608.15394v1cs.CL

TL;DR

The paper asks whether written narratives alone can evoke human temporal illusions, addressing limited evidence beyond visual and auditory stimuli. Using a benchmark of 6,684 narrative pairs across five illusions, it finds that humans show only two illusions, whereas 14 LLMs select literature-predicted scenarios for four, likely reflecting literature retrieval rather than human-like temporal biases.

  • Problem

    Evidence is limited on whether written narratives alone evoke subjective, context-dependent temporal illusions in humans and LLMs.

  • Method

    The study builds a benchmark of 6,684 paired narratives spanning five temporal illusions and evaluates human readers and 14 LLMs from character and reader perspectives.

  • Results

    Humans show only two of five illusions, while LLMs select literature-predicted scenarios for four of five, with roughly 70% of analyzed traces invoking psychological research.

  • Takeaways & Limitations

    Model alignment with published temporal distortions diverges from human behavior and is consistent with literature retrieval rather than human-like temporal perception.

  • Takeaways & Limitations

    Some illusion types retain word-count asymmetries and rely on non-universal assumptions about readers’ experiences.

Abstract

from arXiv · show

Human perception of time is subjective. Well-documented temporal illusions show that the brain relies on context and relational cues for judging duration instead of tracking elapsed time directly. Prior studies established these effects with visual and auditory stimuli. Existing LLM evaluations of temporal perception focus on estimating event durations or multi-step temporal reasoning. In this work, we investigate whether written narratives alone can evoke human temporal illusions, using a new benchmark of 6,684 narrative pairs spanning five illusions. We find that human readers (60 participants) prefer expected scenarios in only two of the five illusions, those where the manipulation is directly visible in text rather than requiring readers to internally simulate duration. We evaluate 14 LLMs on the same benchmark. Surprisingly, we find that models pick the literature-predicted scenario across four of the five illusions, diverging from human behavior. Reasoning traces show that ~70% of responses explicitly evoke psychology research, suggesting that this alignment is consistent with retrieval of published findings rather than human-like temporal biases.

1 Introduction

The paper introduces a narrative benchmark for testing subjective, context-dependent temporal perception and finds that textual narratives preserve only some human temporal illusions. Across the benchmark, humans reliably perceive two of five illusions, whereas 14 LLMs match literature-predicted distortions for four.

  • Motivation: The study targets a gap in AI evaluations by examining subjective, context-dependent temporal perception rather than objective temporal reasoning.Existing time-based evaluations assess mathematical and relative temporal tasks, not perceived expansion or compression of time.
  • Benchmark: The benchmark adapts five psychology-based temporal illusions into 6,684 narrative pairs generated from 111 templates.The original psychology literature primarily studies these illusions with visual and auditory stimuli.
  • Human study: Humans reliably perceive only two of the five temporal illusions through written narratives.The human study tests whether narratives alone preserve temporal illusions.
  • LLM evaluation: 14 LLMs replicate literature-predicted temporal distortions across four of the five illusions.The evaluation compares model judgments with annotator behavior and published temporal distortions.
  • Interpretation: Reasoning traces frequently invoke psychology literature explicitly, suggesting that the models’ alignment reflects those published expectations.The benchmark example likewise shows an LLM thinking trace retrieving literature about the relevant temporal illusion.

2 Background and Related Work

Prior temporal-illusion research used visual or auditory stimuli, while existing LLM benchmarks mainly test factual or reasoning-based temporal judgments. This work instead compares two objectively equal durations using psychological literature as ground truth for subjective temporal distortions.

  • Temporal illusions: Psychological studies examined temporal illusions with visual or auditory stimuli over short durations, motivating five narrative-suitable illusions based on attention, density, novelty, and association.The selected illusions span arousal-based and processing-based phenomena.
  • Temporal illusions: Stimulus density and complexity can lengthen perceived duration, although structured complexity over long periods may compress time.Filled intervals with more discrete components are perceived as longer, whereas organized complexity such as music can make long periods seem shorter.
  • Temporal illusions: Familiarity shortens subjective time through more efficient neural processing, and greater spatial familiarity contracts estimated traversal time.The Familiarity-Duration Effect links familiar items to shorter neural responses and shorter subjective duration.
  • LLM temporal evaluation: Existing temporal benchmarks evaluate factual time knowledge or infer event frequency, duration, and plausibility from temporal context.These benchmarks treat time as historical information or test temporal common sense and causal plausibility.
  • Study motivation: Unlike prior datasets assuming real-world ground truth, this work compares objectively equal durations and uses psychological literature to establish ground truth for relative temporal distortions.The approach targets subjective questions without fixed absolute answers by comparing experiences rather than elapsed-time values.

3 Benchmark Design

The benchmark converts five temporal illusions into matched narrative pairs whose manipulations target mechanisms from psychology literature while controlling narrative length and structure. A template-based generation and filtering process yields 6,684 examples, though template families are not statistically independent and category sizes are imperfectly balanced.

  • Paired narrative design: Each template pairs control and experimental narratives differing in one targeted manipulation, with the experimental scenario predicted by prior literature to feel longer.The benchmark translates psychology experiments into narrative prompts and preserves a paired-condition design.
  • Paired narrative design: Conditions are closely matched in length and narrative structure, with longer and shorter versions included where possible and duration judgments collected from character and reader perspectives.These controls address potential effects of text length on perceived duration.
  • Mechanism-based subtypes: Each illusion is divided into mechanism-based subtypes, distinguishing sustained, high-arousal top-down attention in Emotional Time Dilation from brief, stimulus-driven novelty in the Oddball Effect.The same mechanism-to-subtype approach is applied across the remaining illusions, with matched, length-controlled pairs.
  • Generation and filtering: 6,684 examples remained after three paraphrases per example and removal of pairs differing by more than 7 words.The process began with 2,000 randomly sampled examples per illusion type, producing 10,000 candidate questions, plus 250 manually edited golden cases.
  • Limitations: Template-generated examples are not statistically independent, so the effective sample size is smaller than raw counts, and filtering leaves the dataset imperfectly balanced.The authors state that each category nevertheless retains a sufficiently large sample for analysis.

4 Experimental Setup

The study uses complementary human and LLM evaluations to test narrative temporal perception, replication of known temporal illusions, and perspective sensitivity. Both evaluations use matched scenario-selection tasks, with human participants recruited under prespecified criteria and models assessed under standardized prompting and decoding procedures.

  • Research questions: The evaluation addresses whether humans and language models detect known narrative time-perception effects and whether models distinguish character from reader perspectives.These questions are examined through two complementary evaluations.
  • Human evaluation: Human participants were recruited through Prolific and had to be adults in the United States, primary English speakers, prior Prolific participants, and maintain at least a 90% approval rate.Participation was anonymous, and an optional demographic questionnaire was administered.
  • Human evaluation: Participants responded quickly using their first intuition and evaluated randomized scenarios from either the character’s or reader’s perspective.The questions asked which event felt longer to experience as a character or longer to read and process as a reader.
  • LLM evaluation: Models received the same participant instructions and were evaluated on both a curated golden set and a broader filtered dataset using 2-way and 4-way response formats.The 4-way format additionally allowed same or cannot tell responses.
  • LLM evaluation: Open-weight models used local greedy decoding with a 512-token budget, while proprietary models used batch APIs with default sampling and a 1024-token completion budget.Scenario order was randomized per item using fixed seed 42.

5 Results

LLMs selected the literature-predicted scenario across four of five temporal illusions, with strongest performance on structured effects, while humans showed reliable preferences only for textually detectable manipulations. Model reasoning traces frequently invoked psychology research, suggesting literature retrieval rather than human-like temporal biases.

  • LLM evaluation: 0.85–1.00 pexp: Oddball, Familiarity, and Temporal Order reached high agreement across nearly all models, whereas Emotional Time Dilation was markedly weaker.The result supports more consistent model behavior for structured temporal deviations than emotionally grounded temporal reasoning.
  • LLM evaluation: Character- versus reader-perspective framing reversed model selection for Filled Duration, while other illusion types showed smaller shifts.Character narratives favored the empty interval as longer, whereas retrospective reader framing produced the opposite result.
  • LLM evaluation: Interval Intensity consistently exceeded Emotional Arousal and Stakes-and-Expectations, indicating stronger model responses to observable temporal cues than emotional context.Oddball subtypes likewise performed strongly when temporal deviations were clear and structured.
  • Reasoning traces: ∼70% of Qwen3-32B thinking traces used explicit research-oriented framing, and psychology-related terms appeared across four of five illusion categories.Examples included references to psychological research and mentally stretching or prolonging subjective duration.
  • Model behavior: Smaller open-source models showed stronger positional bias, while greater uncertainty corresponded to higher rates of non-committal responses.GPT models tended to avoid ambiguous decisions, Qwen models expressed uncertainty, and Gemma models usually forced a selection.
  • Human evaluation: Only Oddball and Temporal Order produced statistically significant human preferences for the experimental sentence because their textual disruptions were directly detectable.The remaining illusions required readers to internally simulate duration, and their effects contrasted with the robust LLM effects.

6 Discussion, Limitations and Future Work

The study finds that humans and LLMs diverge in temporal-illusion behavior, with readers showing reliable effects for only oddball and temporal order while models do so for four of five illusions. Limitations include narrative asymmetries, non-universal familiarity assumptions, positional bias, and the absence of internal-representation analysis.

  • Limitations: Narrative translations retain unavoidable word-count asymmetries across illusion types, and Familiarity depends on experiences that vary across readers.A routine task for one reader may be novel for another.
  • Limitations: Substantial positional-bias variation means evaluation results may reflect model-specific artifacts when positional heuristics dominate.UNQOVER-style symmetrization can mitigate moderate bias but is less reliable in these cases.
  • Future Work: The study analyzes outputs and reasoning traces but not internal representations, limiting direct tests of the mechanisms behind model behavior.Future work could probe activation vectors associated with emotion, attention, novelty, and other psychological factors during decision-making.
  • Discussion: Readers show reliable effects for only oddball and temporal order, whereas models show effects for four of the five illusions.Model behavior aligns with the psychology literature rather than with the human participants.
  • Future Work: Understanding how models interpret and represent temporal illusions could inform agent design and contribute to behavior modeling.The evaluation addresses whether models exhibit these illusions, not whether they should.

7 Conclusion … Temporal Order Effects

The paper introduces a narrative benchmark translating five temporal illusions into 6,684 paired scenarios and finds that humans show limited transfer from text, whereas LLMs select literature-predicted scenarios for four illusions. The benchmark’s narrative manipulations operationalize emotional arousal, novelty, filled intervals, familiarity, and temporal order while highlighting limitations of imagined rather than directly experienced states.

  • 7 Conclusion / A.1 Full Benchmark Design: 6,684 paired scenarios from 111 templates instantiate five temporal illusions for evaluating whether written narratives evoke duration biases.The benchmark includes multiple question styles and often contrasts longer and shorter versions.
  • 7 Conclusion: Humans showed no reliable illusion transfer except when the manipulation was directly visible in the text, while models chose literature-predicted scenarios for four of five illusions.Approximately 70% of analyzed reasoning traces explicitly invoked psychological research, even without prompts requesting it.
  • A Appendix: Additional Benchmark Information / A.1 Full Benchmark Design: The benchmark translates classic psychology experiments into narrative prompts designed to compare LLM temporal biases with human-like perception.Its core limitation is that readers must simulate relevant states and durations internally rather than undergo the laboratory manipulation directly.
  • Emotional Time Dilation: Narrative emotional time dilation varies situational stakes, cognitive demand, and sudden emotional events to model arousal–attention effects on perceived duration.Readers imagine these states rather than experience physiological arousal directly, so responses may reflect reasoning about what should feel longer.
  • Oddball Effect: The oddball manipulation places a novel deviation within a predictable sequence, distinguishing bottom-up novelty from other attention-based duration mechanisms.Templates vary whether the oddball is contained within, precedes, or influences the target interval.
  • Filled-Duration Effect: The filled-duration design contrasts intervals containing discrete actions with empty waiting periods, while asking for judgments from external and character perspectives.The expected reader perspective favors longer filled intervals, whereas an unengaging empty interval may feel longer to the character.
  • Familiarity Duration Effect: The familiarity-duration design contrasts novelty and routine through public familiarity, declared expertise, labeled framing, and intrinsic task familiarity.Narrative familiarity is culturally and demographically dependent, unlike laboratory novelty manipulations using controlled stimuli.
  • Temporal Order Effects: Temporal order effects are tested through scrambled workflows, repeated versus distinct events, and coherent narratives interrupted by unexpected events.These manipulations target order-violation overestimation, repetition-based compression, and retrospective expansion after disruption.

A.2 Dataset Creation … B.2 Configuration Details

The benchmark pairs closely matched control and experimental narratives across five illusion types, with manually curated gold cases and filtering intended to preserve clarity and length comparability. Evaluation used character- and reader-perspective prompts, standardized response mapping, and model-specific local or batch configurations.

  • A.2 Dataset Creation: Each template pairs control and experimental conditions differing in one targeted manipulation, while matching length, wording, and narrative structure.The experimental condition represents the scenario predicted by prior psychological literature to feel longer.
  • A.2 Dataset Creation: 20 fillers per category were created from 10 manually written fillers, with GPT-4o expansion followed by hand review; sampling produced 2000 examples for each illusion type and 10000 candidate questions.Character names were sampled from the 100 most common U.S. names over the last one hundred years.
  • A.2 Dataset Creation: 250 golden cases were manually curated, with 50 examples per illusion type, to ensure clarity, consistency, and faithful representation.Generated examples were paraphrased three times with Qwen3-32B, and the closest length-matched variant was selected to reduce word-count differences that could bias duration judgments.
  • A.2 Dataset Creation: Examples within template families are not statistically independent, so the effective sample size is smaller than the raw dataset size.The filtered dataset is not perfectly balanced, although each category retains a sufficiently large sample for analysis and future work may rebalance categories.
  • B.1 Evaluation Instructions and Prompts: Models received two sentences and responded only with “Sentence One” or “Sentence Two,” after which answers were mapped to experimental or control conditions using randomized, fixed presentation order.The same sentence order was held constant across character- and reader-perspective evaluations.
  • B.2 Configuration Details: Open-weight models ran locally with HuggingFace Transformers 5.8.0 and PyTorch 2.13.0 on NVIDIA L40S GPUs using 4bit NF4 precision, while proprietary models used batch APIs.Local models used a 512-token generation budget; proprietary models used a 1024-token completion budget with default sampling settings.

B.3 Evaluation Framework and Dataset Validation

The evaluation combines a hand-edited gold dataset with a scalable automated paraphrased dataset and validates their consistency using Pearson and Spearman correlations. Model-based results are highly consistent across datasets, while reader-based agreement is weaker; 4-way evaluation is preferred for fine-grained analysis.

  • Dataset Construction: The gold dataset contains 50 hand-edited examples per illusion, complemented by a larger automated dataset for scalable evaluation.Examples are rewritten to reflect each illusion’s design and are framed from the character’s perspective.
  • Dataset Consistency: Pearson correlation r measures proportional agreement in debiased effect sizes, while Spearman rank correlation ρ tests rank-based consistency and model rank preservation.These complementary measures compare model- and illusion-level results between the gold and paraphrased datasets.
  • Dataset Consistency: Model rankings are highly consistent across the gold and paraphrased datasets for both 2-way and 4-way evaluations.This consistency is observed across both correlation-based measures and both evaluation types.
  • Dataset Consistency: Reader-based evaluations show weaker agreement between the gold and paraphrased datasets.The paper suggests that modeling temporal perception from a reader’s perspective may be more challenging and less consistent for LLMs, but requires targeted human validation.
  • Evaluation Protocols: The 4-way evaluation more faithfully matches the gold dataset and better detects illusion-specific differences, whereas 2-way evaluation is simpler but less sensitive.Accordingly, 2-way evaluation suits coarse comparisons, while 4-way evaluation is preferable for fine-grained analysis and subtle behavioral differences.

B.4 Effects of Additional Factors: Positional Bias, Complexity, and Non-Committal Answers

Additional factors strongly affect some model behaviors: smaller open-source models show greater positional bias and non-committal responding, while prompt length usually has limited effects except for Filled Duration. Qwen-32B data generation does not systematically differ from Qwen-8B behavior.

  • Data Generation: Qwen-32B’s behavior does not systematically differ from Qwen-8B’s, suggesting Qwen-based option generation does not introduce strong model bias.
  • Positional Bias: Smaller open-source models consistently show the greatest positional bias, whereas larger and proprietary models exhibit substantially smaller gaps.A paired McNemar test confirms substantial position bias for many mainly open-source models.
  • Positional Bias: The study applies Li et al. (2020)’s debiasing strategy, which symmetrizes results across option orderings.
  • Non-Committal Answers: GPT-20B and GPT-OSS show strong non-committal tendencies, reaching up to 85% of answers in one category, while Qwen stays around ∼10% overall.Qwen primarily selects can’t tell, whereas Gemma models almost always commit to an available option.
  • Non-Committal Answers: Models more often give non-committal answers from the character’s perspective, and smaller models are more likely to avoid committing.Models also tend to default to only one of the two non-committal options.
  • Prompt Complexity: Prompt length has limited and inconsistent effects across most illusions, generally preserving the same overall pattern of illusion strength.Filled Duration is the exception: longer narratives more often produce experimental-option selections because additional text provides more duration-filling context.

B.5 Additional LLM Analysis Visualizations

The additional visualizations examine model agreement with literature across illusion types, evaluation settings, perspectives, and sub-categories. They also summarize common words in models’ thinking tokens.

  • B.5 Additional LLM Analysis Visualizations: Word-frequency visualizations summarize the most common words in models’ thinking tokens for each illusion category.The analysis includes visualizations for the general categories and for 2-way evaluation.
  • B.5 Additional LLM Analysis Visualizations: Under 4-way evaluation, model agreement with the literature is shown as the decisive-only proportion pexp, with 95% cluster-robust confidence intervals and Benjamini–Hochberg significance markers.Cells marked × indicate non-significance at a false discovery rate of q < 0.05.
  • B.5 Additional LLM Analysis Visualizations: Under 2-way evaluation, model performance is shown as the positional-bias-corrected proportion pdebiased for character and reader perspectives, with confidence intervals and significance markers.Cells marked × indicate non-significance under Benjamini–Hochberg correction at q < 0.05.
  • B.5 Additional LLM Analysis Visualizations: Additional comparisons show perspective differences in agreement with the literature across illusion sub-categories.The visualizations identify specific sub-categories with differences in agreement.

B.6 Additional Human Analysis Visualizations

This section presents additional visualizations of the human evaluation, covering reader–character perspectives, participant demographics, and position bias across evaluation responses.

  • Figure 22 compares reader and character perspectives under 2-way evaluation across illusion types.
  • Figure 23 shows the demographic distribution of participants in the human evaluation.
  • Figure 24 visualizes position bias in the human evaluation.
Loading 2608.15394v1…