Organizations: F-initiatives, Paris, France · Université Sorbonne Paris Nord, villetaneuse, France · Northwestern University, Chicago, IL, USA · IULM university, Milan, Italy
Human visual search is serial: the fovea must land on a candidate to confirm it, and those landings form a scanpath. Whether multimodal large language models (MLLMs), given the same foveated input, search as humans do bears on their use as models of human vision and on attention-alignment scores. We compare three general-purpose MLLMs with human eye-movement scanpaths on goal-directed search (COCO-Search18), driving each model fixation by fixation through an identical, human-matched foveated view and assessing it along three axes: the decision of target presence, the efficiency of reaching the target, and the gaze process itself. The axes dissociate. On the decision and on target acquisition the models match or exceed humans, detecting present targets near ceiling and reaching them on the first saccade more often than people do. The gaze process is not human. Under the human-matched condition, all three share one signature: low-entropy, large-amplitude, self-consistent scanpaths that agree with themselves far more closely than two humans agree with each other. That is consistent with a single-pass, non-serial architecture rather than a limit of acuity. Matched retinal input reproduces where humans look but not how the looking unfolds in time, and no degradation regime recovers human-like search at human-like success. The gap sits on a process axis that answer-alignment and saliency metrics do not measure. Because they miss it, such metrics cannot certify human-like vision, and zero-shot models suit outcome and spatial questions but not temporal, process-level ones.
Figures & tables
Figure 1: The dissociation on individual trials. Rows are (scene, target) target-present trials ordered by human search difficulty; columns are one human observer and Qwen3.5-35B-A3B (seed 0 ) under four conditions. Green box, target; square, first fixation; numbered circles, fixation order; filled final marker, target found. Where the human accumulates fixations across the scene before confirming the target, the model under sharp and geisler–perry , visually indistinguishable conditions, issues one large saccade to the target and stops: the same outcome reached by a different process, and the divergence is in the order and extent of fixations rather than in their location. Row 6 is a human miss the model does not make. Under gist- k=32 and crop the model does not search longer but terminates in false absence. Rows are illustrative, selected as trials spanning the range of human search difficulty; quantitative claims on the finding axis rest on all 141 target-present trials.
Figure 2: The three-axis dissociation at the human-matched condition. Decision ( d′ , present/absent): the models match or exceed the human reference. Finding (first-saccade TFP@1): the models exceed the human rate (0.49) by a wide margin. Gaze (cross-seed self-consistency): all three lie above the human ↔ human ceiling (0.53) and apart from the human (bar height, cross-seed ScanMatch; annotation, gaze-entropy Cliff’s δ ), with Gemma-4-E4B nearest. The correct outcome is produced by a shared, non-human process.
Figure 3: The search loop. Each episode begins at a forced central fixation; the renderer applies the active foveation condition at the current gaze point, the resulting glimpse is appended to the context, and the model returns a single directive line. A look directive supplies normalised coordinates that are mapped to display pixels and become the next gaze point, closing the loop (red); found or absent terminates the episode (green). No search policy is imposed and the glimpse cap never forces termination: every episode ends on the model’s own decision.
Figure 4: The foveation bracket applied to one scene at a fixed gaze point (red + ). Only geisler–perry is human-matched, and at this geometry it is mild, nearly indistinguishable from sharp ; the gaussian gist ladder and crop are synthetic anchors.
TFP@1
TFP-end
NFix-TP
NFix-TA
TA decl.-abs
dens. CC
condition
Q
G
Gm
Q
G
Gm
Q
G
Gm
Q
G
Gm
Q
G
Gm
Q
G
Gm
human
0.49
0.93
3
5
—
—
sharp
0.97
0.97
0.80
0.97
0.98
0.93
2
2
3
1
4
6
0.94
0.91
0.96
0.55
0.54
0.49
GP
0.97
0.97
0.80
0.98
0.98
0.93
2
2
3
1
4
5
0.93
0.91
0.96
0.58
0.63
0.50
k8
0.93
0.84
0.64
0.95
0.89
0.86
3
2
3
2
4
6
0.84
0.77
0.81
0.65
0.45
0.26
k16
0.78
0.71
0.42
0.91
0.80
0.75
3
3
4
4
4
6
0.78
0.77
0.68
0.70
0.38
0.45
Table 1: Outcome by foveation condition (existence-passed trials); each entry is Qwen/GLM/Gemma (Q/G/Gm) against the human reference (first row; per-model n differs slightly). Finding and outcome measures match or exceed the human values; the dynamic gaze divergence is quantified in Table 2 , though target-absent search length (NFix-TA) also differs from humans for Qwen. TFP@ 1 /TFP-end: target fixated on the first / by the final saccade (TP); NFix-TP/-TA: median fixations on target-present/absent trials; TA decl.-abs: target-absent declared-absent rate; dens. CC: correlation of model and human fixation-density maps.
cond.
model
entropy δ
sacc. amp δ
refix. δ
center bias δ
agent ↔ agent
human ↔ human ceiling
—
—
—
—
0.53
GP
Qwen3.5-35B-A3B
− .67
+ .50
+ .04
− .18
0.84
GLM-4.6V-Flash
− .65
+ .61
− .13
− .31
0.91
Gemma-4-E4B
− .27
+ .23
+ .13
+ .06
0.71
k16
Qwen3.5-35B-A3B
− .56
− .03
+ .49
+ .01
0.78
GLM-4.6V-Flash
− .53
+ .20
+ .30
− .05
0.80
Table 2: Gaze axis at the human-matched condition and at two intermediate degradations, re-analysed from the per-condition tables of the submission (Tables S3–S4). Cliff’s δ against the human distribution ( ∣δ∣>0.33 in bold; sign is agent − human) and cross-seed self-consistency. At GP the deviation is shared: low entropy, large amplitudes, near-human center bias. At k16–k24 the amplitude effect vanishes, refixation rises, and Gemma-4-E4B’s entropy reverses sign, so the GP signature is a property of the legible regime rather than a fixed offset; self-consistency above the ceiling is what persists for the two reasoning-tuned models.
Figure 5: ( 5(a) ) First-saccade targeting (solid) and eventual success (dashed) as functions of the degradation factor k for the three models, with human reference levels; the two quantities fall together, so no k yields human-like search at human-like success. ( 5(b) ) Principal-component analysis of the per-(model, condition) five-statistic gaze signature (PC1, 45%: scanpath length and gaze entropy; PC2, 30%: saccade amplitude versus refixation). Under the legible conditions each model sits in a region offset from the human reference ( ⋆ ); points approach the human only under severe degradation, with Gemma-4-E4B nearest.
model
TP exist.
TA exist.
yes-bias
d′ (GP)
criterion
human
—
—
—
2.91
0.03
Qwen3.5-35B-A3B
0.99
0.81
0.19
3.84
-0.41
GLM-4.6V-Flash
0.99
0.81
0.19
4.23
-0.24
Gemma-4-E4B
0.97
0.80
0.20
3.14
0.38
Table S1 : Detection and decision measures. One-shot existence accuracy and yes-bias (full sharp scene), the in-harness present/absent sensitivity d′ and criterion c at the human-matched condition (Eq. ( S2 )).
Figure S1 : Signal-detection sensitivity ( d′ ) and criterion ( c ) for the in-harness present/absent decision, per model against the human reference.
TFP@1
TFP-end
NFix-TP
NFix-TA
TA decl.-abs
dens. CC
condition
Q
G
Gm
Q
G
Gm
Q
G
Gm
Q
G
Gm
Q
G
Gm
Q
G
Gm
human
0.49
0.93
3
5
—
—
sharp
0.97
0.97
0.80
0.97
0.98
0.93
2
2
3
1
4
6
0.94
0.91
0.96
0.55
0.54
0.49
GP
0.97
0.97
0.80
0.98
0.98
0.93
2
2
3
1
4
5
0.93
0.91
0.96
0.58
0.63
0.50
k8
0.93
0.84
0.64
0.95
0.89
0.86
3
2
3
2
4
6
0.84
0.77
0.81
0.65
0.45
0.26
k16
0.78
0.71
0.42
0.91
0.80
0.75
3
3
4
4
4
6
0.78
0.77
0.68
0.70
0.38
0.45
Table S2 : Outcome by foveation condition; each entry is Qwen / GLM / Gemma against the human reference (first row). TFP@1 and TFP-end follow Eq. ( S3 ); NumFix entries are per-group medians; the last two columns give target-absent declared-absent rate and fixation-density CC with the human map.
Figure S2 : Cumulative target-fixation probability by saccade (Eq. ( S3 )) under sharp (solid) and the human-matched GP (dotted) for the three models, against humans (dashed). The models reach the target on the first saccade; human probability accrues over several fixations.
model
condition
gaze entropy
saccade amp
refixation
scanpath len
center bias
turn angle
coverage
Q
sharp
− .68
+ .52
+ .03
− .27
− .20
+ .03
− .75
GP
− .67
+ .50
+ .04
− .26
− .18
+ .05
− .74
k8
− .69
+ .20
+ .36
− .27
− .05
+ .06
− .65
k16
− .56
− .03
+ .49
− .21
+ .01
+ .12
− .46
k24
− .33
− .08
+ .52
− .07
+ .02
+ .14
− .18
k32
− .15
− .07
+ .50
+ .05
+ .00
+ .24
− .03
Table S3 : Intrinsic gaze signature: Cliff’s δ against the human distribution (Eq. ( S5 ); ∣δ∣>0.33 in bold; sign is agent − human) for every statistic and condition, for all three models. Human medians: gaze entropy 1.58 bits, saccade amplitude 8.57\textdegree , refixation 0.00 , scanpath length 19.4\textdegree , center bias 10.0\textdegree .
Figure S3 : Per-scanpath gaze-statistic distributions at the human-matched condition, human against the three models. The reasoning-tuned models concentrate gaze (low entropy) and make large saccades; Gemma-4-E4B lies closest to the human distribution on the entropy, amplitude and scanpath-length panels (Qwen3.5-35B-A3B is closest on refixation).
Figure S4 : Distribution of ( 4(a) ) saccade directions and ( 4(b) ) turning angles at the human-matched condition. The models’ directional and meander structure departs from the human reference, consistent with the entropy and amplitude effects.
Qwen3.5-35B-A3B
GLM-4.6V-Flash
Gemma-4-E4B
condition
AH
AA
AH
AA
AH
AA
human ceiling
0.53
sharp
0.500
0.839
0.484
0.903
0.473
0.712
GP
0.505
0.843
0.485
0.913
0.469
0.713
k8
0.527
0.814
0.491
0.868
0.462
0.697
k16
0.521
0.776
0.476
0.800
0.405
0.594
Table S4 : Scanpath similarity (ScanMatch, Sec. S3.5 ) by condition: agent ↔ human (AH) and agent ↔ agent (AA, cross-seed determinism) for each model. The human ↔ human agreement ceiling is 0.53.
component
variance
dominant loadings
PC1
45%
+0.62 gaze entropy, +0.62 scanpath len, +0.41 center bias
PC2
30%
+0.69 saccade amp, −0.63 refixation, −0.35 center bias
Table S5 : Principal components of the standardised five-statistic gaze signature across all groups, with the loadings that render the axes interpretable.
Figure S5 : Each statistic under sharp (abscissa) versus the human-matched GP condition (ordinate), per model. Points lie on the identity line: the human-matched foveation changes behaviour negligibly.
sharp
geisler–perry
model
CC
NSS
KL
CC
NSS
KL
Qwen3.5-35B-A3B
0.55
0.57
1.39
0.58
0.60
1.30
GLM-4.6V-Flash
0.54
0.54
1.19
0.63
0.63
0.74
Gemma-4-E4B
0.49
0.48
0.47
0.50
0.48
0.46
Table S6 : Fixation-density agreement with the human map (Sec. S3.6 ) under sharp and the human-matched condition: linear correlation CC, normalised scanpath saliency NSS, and Kullback–Leibler divergence KL. CC and NSS (higher is closer) are highest for GLM-4.6V-Flash; KL (lower is closer) is lowest for Gemma-4-E4B. All three models correlate positively with the human map.
Figure S6 : Spatial agreement with humans (fixation-density CC, abscissa) against temporal divergence (absolute gaze-entropy effect, ordinate). The models match where humans look while diverging in how the looking unfolds.
Figure S7 : First-saccade targeting (solid) and eventual success (dashed) as functions of the synthetic degradation factor k , for the three models, with human reference levels. The two curves fall together; no k yields human-like search at human-like success.
Figure S8 : As degradation increases, the refixation effect (solid) rises (the models revisit locations) and then the declared-absent rate (dashed) rebounds at the most severe degradation as the models default to an absent response.
stratum (n)
human TFP@1
sharp TFP@1 (Q / G / Gm)
near-small (44)
0.54
0.96 / 0.97 / 0.76
near-large (26)
0.65
0.97 / 0.98 / 0.85
far-small (26)
0.37
1.00 / 0.98 / 0.64
far-large (45)
0.39
0.96 / 0.96 / 0.89
Table S7 : First-saccade targeting by eccentricity × target-size stratum (sharp), human against the three models.
Figure S9 : First-saccade targeting per difficulty stratum; the models exceed humans in every eccentricity × size cell, including the hardest.
model
max ∣Δ TFP@1 ∣ over ±0.5/1/1.5\textdegree
Qwen3.5-35B-A3B
0.063
GLM-4.6V-Flash
0.076
Gemma-4-E4B
0.095
Table S8 : Maximum absolute change in TFP@1 across box tolerances of 0.5 , 1 and 1.5 degrees, per model.
model
metric
sharp Δ [95% CI]
GP Δ [95% CI]
Qwen3.5-35B-A3B
gaze entropy
−0.62 [ −0.69,−0.56 ]
−0.61 [ −0.67,−0.54 ]
saccade amp.
+4.57 [ +3.19,+5.95 ]
+4.53 [ +3.15,+5.91 ]
NumFix
−1.53 [ −1.94,−1.13 ]
−1.49 [ −1.89,−1.08 ]
TFP@1
+0.47 [ +0.40,+0.54 ]
+0.48 [ +0.41,+0.55 ]
GLM-4.6V-Flash
gaze entropy
−0.60 [ −0.67,−0.52 ]
−0.60 [ −0.67,−0.52 ]
saccade amp.
+5.58 [ +4.84,+6.32 ]
+5.62 [ +4.88,+6.35 ]
Table S9 : Linear mixed-effects fixed effects against the human reference (Eq. ( S6 ); Δ [95% CI]) under sharp and the human-matched condition, with crossed scene and rater random intercepts. Bold indicates a confidence interval that excludes zero.
Figure S10 : Cross-seed self-consistency as a function of sampling temperature (anchor model, Qwen, core conditions), against the human ↔ human ceiling 0.53.
Visual search is a fundamental cognitive ability. This study investigates whether Multimodal Large Language Models (MLLMs) exhibit human-like difficulty signatures in visual search tasks. We compared search performance of humans (n = 1,250) and MLLMs using identical 2D and 3D stimuli across different set sizes. Both groups showed efficient performance in feature searches, most clearly when the target had a unique color, but performance degradation in conjunction searches as set sizes increased. Additionally, we found strong correlations between human and MLLM error rates (ρ=0.82), which suggests that MLLMs are sensitive to similar objective complexities, such as stimulus heterogeneity. However, differences were found as well: whereas humans invested extra search time to respond accurately on target-absent trials, MLLMs exhibited extreme present/absent response biases in complex searches. We conclude that MLLMs replicate high-level human performance signatures, yet their underlying computations differ significantly.
Renchi Zhang, Joost C. F. de Winter, Dimitra Dodou +2
Faculty of Mechanical Engineering, Delft University of Technology, Delft, The Netherlands
Human observers prioritize visual information according to task goals. Most computational models of naturalistic viewing are gaze-trained for free viewing, leaving open whether goal-directed attention can emerge in systems without gaze supervision. We tested two off-the-shelf vision-language models (VLMs), Qwen3-VL-32B-Thinking and Gemma-4-26B-A4B-it, on 4,887 naturalistic scenes under visual-search and free-viewing instructions. Model predictions were compared with human fixations on the same images under corresponding tasks. Both models aligned more closely with human fixations under matching goals than under mismatched goals. This crossover persisted in target-absent scenes, where alignment could not be explained by simple visual grounding, and appeared in decoder-layer readouts. Furthermore, model-thinking traces were grounded in target semantics during search and in visual prominence during free viewing. These findings show that general-purpose VLMs can generate human-aligned, goal-directed spatial priorities without gaze-specific training, informing theories of goal-directed attention and offering scalable tools for predicting where people look across tasks.
Han Zhang
University of Michigan Transportation Research Institute, Ann Arbor, MI, USA
Visual search has been one of the most productive paradigms in the study of visual attention: the way reaction time scales with the number of items distinguishes parallel, "pop-out" search from serial, attention-demanding search. I ask whether vision-language models (VLMs) exhibit the same behavioral signatures. I adapt four classic paradigms: feature versus conjunction search, spatial-configuration (T-vs-L) search, enumeration, and the tilted/vertical search asymmetry; and present them to current frontier and mid-tier models. Because a single model call has no reaction time, I use the number of reasoning ("thinking") tokens a model spends per trial as a within-model analog of search effort, and I compare against a large public human benchmark (Wolfe et al., 2010). The models reproduce several human signatures: feature search costs flat effort while conjunction effort climbs with set size; frontier models hold accuracy where mid-tier models collapse to chance; and a resolution control shows the conjunction cost is genuine search rather than difficulty resolving small shapes. They also diverge from humans in informative ways. The target-present effort slope exceeds the target-absent slope, reversing the human ordering; enumeration remains accurate where humans would lose count; and a reasoning model with adaptive deliberation declines to deliberate on detection tasks altogether, so that a single search expresses itself as an effort gradient in one model and as an accuracy cliff in another. I argue that psychophysical paradigms, applied behaviorally, are a sharp and inexpensive probe of machine visual cognition, and that the points of divergence are as informative as the points of agreement.