In Search of Lost Time: Temporal Competence in Large Language Models
Authors: Ryan Liu, David Lee, Icey Siyi Ai, Marwa Abdulhai, Rachit Dubey, Thomas L. Griffiths
Organizations: Department of Computer Science, Princeton University · Department of Operations Research and Financial Engineering, Princeton University · Department of Industrial Engineering and Operations Research, Columbia University · Department of EECS, University of California, Berkeley · Department of Communication, University of California, Los Angeles · Department of Psychology, Princeton University
Today's large language models (LLMs) increasingly interact with time: AI agents schedule tasks, coordinate with people and other AI systems, and even decide how long to think. Yet LLMs only ever encounter time secondhand---in descriptions rather than experience. Does this hinder their reasoning and decisions that involve time? Our paper traces this question from a model's knowledge to its actions through four successive evaluations. We find that models know the rules of time, which can be read, but struggle to predict durations, which must be lived; while they are highly accurate (92.0-97.5%) on a collection of questions built from the theorems of various time-related literatures, they underestimate their own completion times by up to 99.5% and fail to identify which tasks will take them longer to run. And while they are better at estimating durations of human tasks, their duration judgments change based on how an event is worded, magnifying similar human biases by up to 1.3×. Moving from estimation to action, we find that some LLMs disregard time costs completely when answering user queries, ignoring a 1,200-fold increase in wait time entirely. Together, our experiments reveal new challenges posed to LLMs as they navigate a dimension they cannot directly perceive, and highlight how their blind spots can even be simple cognitive abilities that humans take for granted.
Figures & tables
Information Input
Mental Processes
Work Output
Interacting w/ Others
Summarize sources
Math problems
Brick wall
Social media posts
Fact-check claims
Logic questions
Brick steps
Reply to comments
Table 1: Vignette domains grouped by O*NET Generalized Work Activities branch.
Axiom / Theorem
Claude Opus 5
000 GPT-5.6 Sol
Gemini 3.7 Flash
Learning
100.0 †
100.0 †
100.0 †
Decay
100.0 †
100.0 †
100.0 †
Screening
0 99.2 (98.1–100.0)
0 90.0 (82.0–96.1)
0 95.2 (88.3–99.8)
Loading
0 99.3 (97.4–100.0)
0 92.5 (86.9–97.3)
0 72.3 (50.2–91.5)
Recovery
100.0 †
0 99.4 (98.3–100.0)
100.0 †
Interruption count
0 74.1 (59.3–87.4)
0 97.9 (95.6–99.6)
100.0 †
Table 2: Axiom and theorem accuracy across vignettes (%), averaged over all domains and responses.
Figure 1: LLM estimates of human task-completion times , shown on log scale. White circles show median point estimates across 30 repetitions. Dark, medium, and light intervals show median 50%, 80%, and 95% uncertainty bounds. Red markers show HCAST’s geometric-mean human completion-times; crosses indicate references outside the model’s 95% interval.
Figure 2: Model estimates of their own agent completion times. Markers and intervals follow Figure 1 . Red markers show geometric-mean completion times across successful METR runs.
Figure 3: LLMs’ duration judgments react to wording differently from humans. They are more sensitive to verb features and less sensitive to context features. (a) Human vs model effects per sentence feature; the diagonal is human-level sensitivity. (b) Sensitivity gain by feature group, with 95% intervals. (c) The ratio of light-verb over simple-verb duration estimates for Wittenberg and Levy (2017) stimuli, one example per event class.
Figure 4: Sensitivity to time cost varies across models. Points indicate mean rates that the model chooses the tool across 30 tasks. Lines are least-squares fits in log10t~ . Gray diamonds and dashed lines represent the base rates of tool use with no stated delay; this baseline is excluded from the fits.
Figure 5: Across settings and models, sensitivity to time cost increases when full completion durations are provided. Points show mean tool use across 30 tasks. Lines are least-squares fits in log10t~ . Hollow points and dashed lines mark durations estimated by the model (Section 6.1); filled points and solid lines mark true durations provided in the prompt (Section 6.2).
Appendix figures & tables35 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Order
Scored
Accuracy
Unanimous
Δ accuracy (95% CI)
Claude Opus 5
Fixed
351
65.5
72.9
+2.2 ( −1.9 , +6.5 )
Shuffled
353
67.7
74.6
GPT-5.6 Sol
Fixed
384
93.5
98.4
+0.3 ( +0.0 , +1.3 )
Shuffled
384
93.8
100.0
Gemini 3.7 Flash
Fixed
384
94.5
96.9
−0.8 ( −2.3 , +0.0 )
Shuffled
384
93.8
100.0
Appendix
Table 3: Accuracy and consistency (%) under fixed and shuffled option order. Unanimous is the share of items answered with the same canonical label on all six shuffled reruns. Δ is shuffled minus fixed accuracy, reported with a 95% paired cluster bootstrap interval.
Figure 6: Scored rows by the option a model chose (left) and by the letter that option appeared under (right). Under shuffling the label counts barely move, but the slot counts flatten, indicative of a model responding to option content rather than option position.
Model
Gold in (a)
Gold in (b)
Gold in (c)
Gold in (d)
Claude Opus 5
62.0 (47.6, 75.9)
69.9 (55.7, 82.7)
67.6 (54.0, 79.6)
71.0 (58.0, 82.4)
GPT-5.6 Sol
94.1 (86.4, 100)
91.1 (81.0, 98.8)
93.1 (85.3, 99.1)
96.2 (90.7, 100)
Gemini 3.7 Flash
94.1 (86.4, 100)
91.1 (81.0, 98.8)
93.1 (85.3, 99.1)
96.2 (90.8, 100)
Appendix
Table 4: Shuffled-arm pooled accuracy (percent) by the display slot the gold was permuted into, with 95% cluster-bootstrap intervals. Row counts per cell range from 73 to 116.
Domain
Axiom/theorem
Queries per cell
BS
BW
LQ
MP
CR
FC
SP
SS
Total
learning
60
0
0
0
0
0
0
0
0
0/480
decay
60
0
0
0
0
0
0
0
0
0/480
screening
120
0
0
0
0
107
120
0
108
335/960
loading
60
0
0
0
0
0
60
0
0
60/480
recovery
60
0
0
0
0
0
0
0
6
6/480
Appendix
Table 5: Claude Opus 5 content-classifier refusals in the main run, by axiom or theorem and domain. Queries per cell is the total number of queries made in each domain per axiom/theorem row. Cells are refused responses out of runs in that row and domain. Domains: BS = brick steps, BW = brick wall, LQ = logic questions, MP = math problems, CR = customer replies, FC = fact-check claims, SP = social posts, SS = summarize sources.
Model
50% interval
80% interval
95% interval
Claude Opus 5
43.75% (7/16)
56.25% (9/16)
93.75% (15/16)
GPT-5.6 Sol
43.75% (7/16)
75.00% (12/16)
87.50% (14/16)
Gemini 3.7 Flash
18.75% (3/16)
31.25% (5/16)
50.00% (8/16)
All model–task pairs
35.42% (17/48)
54.17% (26/48)
77.08% (37/48)
Appendix
Table 6: Reference containment for human completion-time estimates. Each cell reports the proportion and number of tasks for which the HCAST human geometric-mean reference falls within the model’s median uncertainty interval across 30 repetitions.
Estimator and target
50% interval
80% interval
95% interval
Claude Opus 4.6 → self
40.00% (6/15)
66.67% (10/15)
66.67% (10/15)
GPT-5.2 → self
0.00% (0/16)
0.00% (0/16)
0.00% (0/16)
Gemini 3.1 Pro Preview → Gemini 3 Pro
0.00% (0/16)
0.00% (0/16)
0.00% (0/16)
All evaluable model–task pairs
12.77% (6/47)
21.28% (10/47)
21.28% (10/47)
Appendix
Table 7: Reference containment for agent-time estimates. Each cell reports the proportion and number of evaluable tasks for which the geometric mean duration across successful METR runs falls within the model’s median uncertainty interval across 30 repetitions. The Gemini condition is proxy self-estimation: Gemini 3.1 Pro Preview estimates Gemini 3 Pro.
Estimator
Original
Additional
Factor (95% CI)
Adj. p
Claude Opus 4.6
8.93
7.01
0.785 [0.577, 1.067]
.113
GPT-5.2
18.19
11.22
0.617 [0.507, 0.751]
<.001
Gemini proxy
8.55
6.31
0.738 [0.567, 0.960]
.053
Appendix
Table 8: Effect of additional task information on self-estimates. Forecasts are geometric means, in minutes, of task-level median predictions. Factor is the additional-information forecast divided by the original forecast. Confidence intervals are unadjusted 95% intervals from paired t -tests on log forecasts; p -values use Holm correction across the three estimators.
Model
NM
Bias factor
Abs. error factor
Within 2×
Within 4×
ρ
Claude Opus 5
497
0.94
1.96
52.5%
95.6%
0.61
GPT-5.6 Sol
500
0.70
1.61
73.4%
96.8%
0.86
Gemini 3.7 Flash
498
0.50
2.04
46.2%
99.4%
0.82
Appendix
Table 10: WildChat response-time estimation results. A bias factor below one indicates underestimation; the absolute error factor summarizes error magnitude irrespective of direction. Within- 2× and within- 4× report the proportions of valid requests in the corresponding multiplicative ranges.
Sensitivity gain
Judge
ρ
Bias
All features
Predicate
Context
One annotator (reference)
0.49
–
1
1
1
Claude Opus 5
0.688 [0.67, 0.71]
+0.49
1.35 [1.22, 1.47]
1.62 [1.50, 1.76]
0.89 [0.69, 1.11]
GPT-5.6 Sol
0.619 [0.60, 0.64]
+0.01
1.27 [1.13, 1.39]
1.57 [1.40, 1.72]
0.76 [0.52, 0.99]
Gemini 3.7 Flash
0.632 [0.61, 0.65]
+0.50
1.30 [1.15, 1.42]
1.66 [1.49, 1.83]
0.69 [0.47, 0.92]
Qwen2.5-32B
0.430
-0.91
0.67 [0.30, 1.03]
1.13 [0.75, 1.51]
0.37 [-0.17, 0.90]
Appendix
Table 11: Duration judgments on 5000 UDS-T events. ρ : Spearman correlation with the mean human label. Bias: mean model label minus mean human label, in scale steps. Gain: sensitivity to sentence features relative to people ( g ; defined in Section 5), over all features and over the predicate and context groups. Brackets give 95% bootstrap intervals. Qwen2.5-32B was run on the 1000-event subset only.
Model
Calls
Tokens (in / out)
Cost
Claude Opus 5
14315
7.13M / 0.21M
$40.99
GPT-5.6 Sol
12202
2.87M / 0.69M
$25.17
Gemini 3.7 Flash
12211
2.01M / 1.72M
$7.95
Qwen2.5-32B
1042
–
local
Total
$74.11
Appendix
Table 12: API usage for all duration runs: the full sample, four additional runs and three prompt variants on the 1000-event subset, and pilot calls.
Figure 7: Per-model version of Figure 3 a: human effect ( x ) against model effect ( y ) for each of the ten features, one panel per model, with ±1 SE. Filled markers are predicate features, open markers context features. Qwen2.5-32B is estimated on the 1000-event subset.
Feature
Humans (5k)
Claude Opus 5
GPT-5.6 Sol
Gemini 3.7
Qwen 2.5
Humans (1k)
Non-verbal pred.
+1.38 (0.09)
+2.76 (0.12)
+2.69 (0.14)
+3.00 (0.13)
+1.77 (0.35)
+1.16 (0.18)
Simple present
+1.14 (0.08)
+2.03 (0.11)
+2.02 (0.13)
+2.44 (0.12)
+0.99 (0.35)
+1.36 (0.18)
Perfect
+1.27 (0.14)
+1.56 (0.19)
+1.28 (0.23)
+1.03 (0.22)
+1.23 (0.50)
+0.72 (0.28)
Progressive
+0.92 (0.12)
+1.19 (0.16)
+1.33 (0.18)
+1.29 (0.16)
+0.54 (0.35)
+0.62 (0.25)
Imperative
-0.20 (0.14)
-0.37 (0.16)
-0.35 (0.18)
-0.28 (0.18)
-0.02 (0.53)
-0.03 (0.28)
Plural-noun subj.
+1.15 (0.11)
+1.25 (0.14)
+1.00 (0.18)
+1.06 (0.17)
+0.20 (0.42)
+0.89 (0.24)
Appendix
Table 13: Effect of each feature on the duration label, in scale steps (standard error in parentheses). Upper block: predicate features; lower block: context features.
Model
Prompt
ρ
Gain (all)
Gain (predicate)
Gain (context)
Claude Opus 5
base (mean of 5)
0.676
1.16
1.82
0.72
reworded
0.651
1.09
1.85
0.57
reversed
0.679
1.14
1.87
0.64
numeric
0.653
1.12
1.85
0.62
GPT-5.6 Sol
base (mean of 5)
0.591
1.05
1.73
0.58
reworded
0.608
0.97
1.64
0.52
Appendix
Table 14: Agreement and sensitivity gain under each prompt, on the 1000-event subset. Base values average five runs; each variant is a single run.
Figure 8: (a) Correlation between each model’s labels and the mean human label, with the single-annotator reference. (b) Sensitivity gain by feature group with 95% intervals; small gray markers give the three prompt variants. Both panels use the 1000-event subset, where the repeated runs and prompt variants were collected.
Same category
Mean difference (steps)
SD of ρ
SD of gain
Two annotators
29.2%
1.82
–
–
Claude Opus 5, two runs
92.2%
0.15
0.004
0.058
GPT-5.6 Sol, two runs
80.4%
0.45
0.007
0.080
Gemini 3.7 Flash, two runs
89.4%
0.20
0.008
0.035
Appendix
Table 15: Agreement between two annotators on the same event, and between two runs of the same model, on the 1000-event subset. The last two columns give the standard deviation of ρ and of the pooled sensitivity gain across five runs.
inst.
sec.
min.
hours
days
weeks
months
years
dec.
cent.
forever
Humans
6.0
8.6
27.7
11.5
11.1
6.5
10.2
12.5
2.6
0.7
2.6
Claude Opus 5
2.4
6.6
27.2
11.5
8.8
4.8
8.1
26.3
2.0
0.4
2.0
GPT-5.6 Sol
9.8
13.1
21.0
10.3
7.7
2.9
8.5
22.7
1.0
0.4
2.6
Gemini 3.7 Flash
5.1
10.5
19.1
11.3
10.0
4.2
11.7
20.6
2.7
0.9
4.0
Qwen2.5-32B †
27.2
5.1
34.5
4.6
7.7
1.0
3.2
9.7
0.4
0.5
6.1
Appendix
Table 16: Share of labels (%) in each duration category. Humans and the API models: 5000 events. Qwen2.5-32B † : 1000-event subset.
Figure 9: Per-model version of Figure 3 c, including Qwen2.5-32B: light-verb estimate divided by simple-verb estimate, as a geometric mean over the events of each class (log scale). Error bars: SD across five runs.
Event
Class
Humans
Claude Opus 5
GPT-5.6 Sol
Gemini 3.7 Flash
Qwen2.5-32B
embracing
punctive
0.50
0.80
1.00
0.83
1.67
hugging
punctive
1.21
1.00
1.00
1.00
1.00
cuddling
punctive
0.10
0.07
0.03
0.03
0.30
kicking
punctive
0.67
1.00
0.67
1.00
1.00
shaking
punctive
0.41
0.67
0.50
1.00
0.17
poking
punctive
0.40
1.00
1.00
1.00
1.00
Appendix
Table 17: Light-verb / simple-verb duration ratio for each Wittenberg & Levy event.
Model
Reasoning
N
Tool
Acc.
Δp
ρ
pp [95% CI]
OR [95% CI]
Var.
Cubic equations
Claude Opus 5
Default
2,100
0.0%
100.0%
3
–
–
–
0/30
None
1,912
72.5%
42.4%
10
0.79
−7.1
[ −10.8 , −3.4 ]
0.26
[0.14, 0.48]
18/30
GPT-5.6 Sol
Default
2,100
13.4%
99.7%
7
0.86
+0.1
[ −1.0 , +1.2 ]
1.02
[0.88, 1.18]
16/30
None
2,100
53.0%
12.7%
19
0.91
+0.5
[ −1.0 , +2.1 ]
1.08
[0.85, 1.37]
13/30
Gemini 3.7 Flash
Default
2,068
26.1%
97.4%
18
0.81
−3.1
[ −5.2 , −0.9 ]
0.56
[0.42, 0.74]
13/30
Appendix
Table 18: Results summary. Tool use, accuracy, perceived benefit, and within-task sensitivity to t~ for each condition. Quantities are defined in Appendix D.4.1 .
Model
Reasoning
N
Tool
Acc.
Δp
ρ
Cubic equations
Claude Opus 5
Default
300
0.0%
100.0%
4
–
None
267
81.4%
40.2%
13
0.59
GPT-5.6 Sol
Default
300
15.3%
99.7%
7
0.61
None
300
61.0%
10.3%
18
0.89
Gemini 3.7 Flash
Default
295
32.5%
98.6%
18
0.80
Appendix
Table 19: No-delay baseline summary. Same quantities as Table 18 , for trials in which no time cost is stated.
Figure 10: Accuracy and calculator use by equation difficulty. Each point averages the 10 tasks at that level, pooled over all seven t~ ; bars are 95% bootstrap intervals over tasks. Gray diamonds beside each series mark the no-delay baseline.
Model
Reasoning
No t~
1 s
5 s
10 s
20 s
30 s
3 min
20 min
Cubic equations
Claude Opus 5
Default
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
None
81.4
77.3
76.9
71.1
79.6
73.8
67.0
50.2
GPT-5.6 Sol
Default
15.3
13.3
15.0
12.7
12.7
13.0
12.7
14.7
None
61.0
51.7
55.7
51.3
50.3
53.7
54.3
53.7
Gemini 3.7 Flash
Default
32.5
32.7
30.1
26.6
22.5
24.0
25.1
22.0
Appendix
Table 20: Tool use by stated time cost. Each entry is the unweighted mean of the 30 task-level rates at that t~ , matching Figure 4 . The first column is the no-delay baseline.
Calculator use (%)
Accuracy (%)
Model
Reasoning
Easy
Medium
Hard
Easy
Medium
Hard
Claude Opus 5
Default
0.0
0.0
0.0
100.0
100.0
100.0
None
39.8
78.8
99.0
47.7
55.3
24.3
GPT-5.6 Sol
Default
0.3
6.6
33.4
100.0
99.6
99.6
None
26.3
43.0
89.6
18.0
11.1
8.9
Gemini 3.7 Flash
Default
0.0
12.6
65.6
100.0
98.5
93.8
Appendix
Table 21: Calculator use and accuracy by equation difficulty. Cubic tasks only, ten per level. Each entry is the unweighted mean of the ten task-level rates, pooled over all seven t~ .
Median pno-tool (%)
Median Δp (pp)
Model
Reasoning
Easy
Med.
Hard
Easy
Med.
Hard
Cubic equations
Claude Opus 5
Default
95
91
85
2
4
8
None
85
75
35
5
10
35
GPT-5.6 Sol
Default
97
95
65
3
4
33
None
93
85
35
6
13
63
Appendix
Table 22: Elicited confidence estimates. Medians over retained trials. pno-tool and ptool are the model’s stated probabilities of answering correctly without and with the tool; Δp is their difference. Cubic entries are broken out by difficulty.
Figure 11: Tool use increases with the perceived accuracy benefit from using the tool, demonstrated by the positive relationship between the two axes. Points are task– t~ cells, with median reported benefit Δp on the horizontal axis and tool-use rate on the vertical axis. Color denotes reasoning condition; marker and line style denote t~ . Curves are descriptive logistic fits, each drawn over the Δp range observed at that t~ . For simplicity, only the shortest and longest delay values ( t~ ) are shown.
Domain
Model
Matched tasks
Median difference (s)
Cubics
Claude Opus 5
0
–
GPT-5.6 Sol
16
+6.69
Gemini 3.7 Flash
13
+2.77
Search
Claude Opus 5
4
+2.06
GPT-5.6 Sol
6
+2.46
Gemini 3.7 Flash
14
+1.29
Appendix
Table 23: Model-call durations. All results use default reasoning. For each task and stated time cost where both tool and no-tool trials occur, we subtract the mean no-tool duration from the mean tool duration. Then, the differences are averaged equally across matched delays within each task, and the median across matched tasks is reported. Duration is the sum of recorded model-call latencies in the task-answering conversation, excluding confidence estimation and tool execution. Positive values indicate longer model-call durations with tool use.
Model
Reasoning
Estimated (s)
Recorded (s)
Ratio ( × )
Cubic equations
Claude Opus 5
Default
100.0
8.2
12.0 [ 7.7 – 17.3 ]
None
120.0
2.6
43.3 [ 15.3 – 102.6 ]
GPT-5.6 Sol
Default
25.0
6.1
4.8 [ 2.7 – 7.9 ]
None
25.0
2.0
12.2 [ 4.7 – 21.6 ]
Gemini 3.7 Flash
Default
18.0
4.1
4.1 [ 2.3 – 6.7 ]
Appendix
Table 24: Estimated and recorded response time without the tool. Estimated is the median of the model’s estimates of its total response time without the tool; Recorded is the median over questions of the model’s recorded time answering directly in Section 6.1 (Appendix D.5.1 ). Ratio is the median over questions of estimated divided by recorded time, with the 10th–90th percentile range; values above one indicate overestimation.
Model
Reasoning
Estimated (s)
Recorded (s)
Ratio ( × )
Cubic equations
Claude Opus 5
Default
85.0
12.8
6.2
None
79.0
8.1
9.4
GPT-5.6 Sol
Default
15.0
13.8
1.1
None
7.0
5.9
1.1
Gemini 3.7 Flash
Default
10.0
7.2
1.5
Appendix
Table 25: Estimated and recorded response time with the tool. Estimated is the median of the model’s with-tool estimate minus the stated time cost t~ , i.e., its own time on the tool route; Recorded is the median over questions of the recorded model-call time on that route, excluding the tool’s running time. Ratio is the median over questions of estimated divided by recorded time.
Believed faster, by t~
Actual
Model
Reasoning
1 s
5 s
10 s
20 s
30 s
3 min
20 min
at 1 s
Cubic equations
Claude Opus 5
Default
89
67
61
45
31
2
0
0
None
99
92
71
57
25
0
0
0
GPT-5.6 Sol
Default
97
83
66
40
23
0
0
3
None
97
85
78
45
30
0
0
0
Appendix
Table 26: Tool route estimated to be faster, by stated time cost. Share of estimates (%) in which the model’s estimated total time with the tool is below its estimated time without it, by t~ . The last column gives the share of questions (%) on which the recorded tool route plus 1 s is in fact faster than answering directly.
Source
Median time (s)
Model
Reasoning
Sec. 6.1
Forced
No tool
Tool
Cubic equations
Claude Opus 5
Default
30/60
30/60
8.0
12.8
None
48/60
12/60
3.0
8.1
GPT-5.6 Sol
Default
46/60
14/60
6.0
13.8
None
43/60
17/60
2.0
5.9
Appendix
Table 27: Completion times provided in Section 6.2. Sec. 6.1 and Forced count the 60 question–action times (30 questions, two actions) taken from Section 6.1 trials and from forced runs. No tool and Tool are medians over questions of the provided times, the latter before adding t~ .
Model
Reasoning
N
Tool
ρ
pp [95% CI]
OR [95% CI]
Var.
Cubic equations
Claude Opus 5
Default
2,098
0.0 %
–
–
–
0/30
None
2,039
46.4 %
0.83
−11.5 [ −17.1 , −6.0 ]
0.15 [ 0.06 , 0.38 ]
13/30
GPT-5.6 Sol
Default
2,100
2.8 %
0.75
−1.9 [ −3.8 , 0.0 ]
0.37 [ 0.20 , 0.70 ]
9/30
None
2,100
9.1 %
0.78
−5.7 [ −8.9 , −2.4 ]
0.32 [ 0.21 , 0.47 ]
13/30
Gemini 3.7 Flash
Default
2,058
11.3 %
0.79
−6.7 [ −10.9 , −2.5 ]
0.15 [ 0.07 , 0.34 ]
11/30
Appendix
Table 28: Results summary when durations are provided (Section 6.2). Quantities as in Table 18 , defined in Appendix D.4.1 .
Model
Reasoning
1 s
5 s
10 s
20 s
30 s
3 min
20 min
Cubic equations
Claude Opus 5
Default
0.0
0.0
0.0
0.0
0.0
0.0
0.0
None
50.7
56.9
53.8
53.6
54.3
33.2
20.3
GPT-5.6 Sol
Default
6.7
2.3
2.3
4.0
3.3
0.7
0.0
None
17.0
12.7
10.7
13.7
7.7
0.7
1.3
Gemini 3.7 Flash
Default
21.2
20.3
13.0
11.3
11.9
3.0
2.0
Appendix
Table 29: Tool use by stated time cost when durations are provided. Each entry is the unweighted mean of the 30 task-level rates at that t~ , matching the filled points in Figure 5 ; Table 20 gives the same quantities for Section 6.1.
Figure 12: Perceived accuracy benefit and tool choice when durations are provided. Same layout and fits as Figure 11 , for Section 6.2. Points are task– t~ cells, with each task’s median reported benefit Δp from Section 6.1 on the horizontal axis and its tool-use rate on the vertical axis. Color denotes reasoning condition; marker and line style denote the t~ . Each curve is drawn only over the Δp range observed at that t~ .
Figure 13: Cubic tool use by difficulty when durations are estimated and provided. Points show mean tool use across the 10 cubics of each difficulty. Hollow points with dashed lines show durations estimated by the model (Section 6.1); filled points with solid lines show durations provided in the prompt (Section 6.2).
Model
Reasoning
Easy
Medium
Hard
Claude Opus 5
Default
–
–
–
None
−10.3→
−7.0
−12.3→
−18.5
−1.9→
−9.2
GPT-5.6 Sol
Default
0.0→
0.0
+0.3→
−0.1
+0.1→
−5.5
None
+2.6→
0.0
−1.6→
−4.3
+0.7→
−12.7
Gemini 3.7 Flash
Default
–
−3.4→
−3.0
−6.3→
−18.1
None
−1.0→
−1.2
−8.3→
−11.3
−3.2→
−16.3
Appendix
Table 30: Change in time sensitivity by equation difficulty. Slope of tool use per 10× increase in t~ (percentage points), from least-squares fits in log10t~ to the mean tool use of the ten cubics at each level, as estimated → provided (Sections 6.1 and 6.2). Dashes mark no calculator use in either section.
Large language models cannot estimate how long their own tasks take. We investigate this limitation through four experiments across 68 tasks and four model families. Pre-task estimates overshoot actual duration by 4--7× (p<0.001), with models predicting human-scale minutes for tasks completing in seconds. Relative ordering fares no better: on task pairs designed to expose heuristic reliance, models score at or below chance (GPT-5: 18% on counter-intuitive pairs, p=0.033), systematically failing when complexity labels mislead. Post-hoc recall is disconnected from reality -- estimates diverge from actuals by an order of magnitude in either direction. These failures persist in multi-step agentic settings, with errors of 5--10×. The models possess propositional knowledge about duration from training but lack experiential grounding in their own inference time, with practical implications for agent scheduling, planning and time-critical scenarios.
Human perception of time is subjective. Temporal illusions show that the brain relies on context and relational cues when judging duration. Prior studies established these effects with visual and auditory stimuli. We ask whether written narratives alone can cause these illusions, and whether LLMs respond as human participants do. We introduce a benchmark of 6,684 paired scenarios spanning five illusions, each judged from the character's and the reader's perspective. In our human study (60 participants), participants' judgments matched the predicted directions for all five illusions in at least one perspective (7 of 10 conditions). Across 14 LLMs, models reproduce every effect humans show in the same direction, but on average more strongly. Most effects persisted even after cue words were removed and after controlling for 15 lexical and semantic features. These results show that models and humans agree in direction but not in strength, which cautions against treating LLM judgments as substitutes for human responses for subjective temporal reasoning.
Temporal reasoning is widely assumed to decompose into two modes: temporal arithmetic (date and time calculations), and temporal semantics (understanding temporal and causal relations). This dichotomy shapes how benchmarks are designed, how models are trained, and how progress is measured, yet its empirical validity has never been directly tested. We provide two lines of evidence that challenge it. First, we test whether tasks labelled as arithmetic or semantic by the benchmarks support transferable gains within their purported mode. Finetuning on these subsets yields predominantly dataset-specific gains (+6.8 to +23.1 percentage points), while transfer to a different dataset in the same mode ranges from −3.2 to +1.1 percentage points. Across the two benchmark families used in the transfer analysis, this pattern provides little evidence of coherent, transferable temporal abilities. Second, using cognitive-demand-based model evaluation, we examine which capabilities actually drive performance on these benchmarks. We find that quantitative and logical reasoning, the assumed proxies for temporal arithmetic and temporal semantics, are not the primary drivers; knowledge is. Together, these findings show that current temporal reasoning benchmarks do not cleanly realise the arithmetic-semantics distinction and instead largely reward models for remembering how time typically works.
Auss Abbood, Fernando Martínez-Plumed, José Hernández-Orallo +1
Language Technology Lab, University of Cambridge, UK · Valencian Research Institute for Artificial Intelligence (VRAIN) Universitat Politècnica de València, Spain · Leverhulme Centre for the Future of Intelligence, University of Cambridge, UK