Social Cognition
Reverse Correlation Task
Also called: noise-based reverse correlation, two-image forced choice reverse correlation, classification image task, 2IFC reverse correlation
Picture what a trait looks like to each person: participants choose the more trustworthy, dominant or happy of two noisy versions of one average face, and the chosen noise averages into a classification image with its informational value.
8-10 minUntimed
Free to start: 10 free participants per account, then $0.99 per participant for academic research, or any yearly plan. Preview any test free after you sign up.
- Papers, 2021 to 2025
- 39
- Evidence grade
- C
- Registered clinical studies
- 0
What it measures
Shows what a trait looks like to each person. On each trial participants see the same average face twice, covered with a random noise pattern and its opposite, and pick the one that looks more trustworthy, dominant or happy. The chosen patterns average into a classification image; the main score is its informational value.
Constructs
- Social perception of faces
- Mental representations of traits
- Face evaluation: trustworthiness and dominance
- Reading facial expressions
Research uses
- Mental images of trustworthy, dominant or happy faces in social and personality research
- Comparing face impressions between groups of perceivers, at the group level
- Changes in face impressions after learning or social experience
How the paradigm works
The reverse correlation task shows what a trait looks like to a person, without asking them to describe it. On each trial the participant sees the same face twice, side by side, each covered with random visual noise: one with a noise pattern added and the other with the same pattern subtracted. The participant picks the face that looks more trustworthy (or more dominant, or happier, as a setting), by first impression and with no time limit. Because the noise is random, the chosen faces differ from the others only in the features the participant used, so the average of the chosen noise patterns, the classification image, pictures those features (Mangini and Biederman, 2004; Dotsch and Todorov, 2012). In Dotsch and Todorov's study, images of trustworthy faces wore a slight smile and images of dominant faces looked more masculine.
More about the paradigm
The face is an average of the neutral photographs of 96 people of the Face Research Lab London Set (DeBruine and Jones, 2017), with women and men, and each of the set's four groups, weighted equally, so no real person's face is shown. The noise is the sinusoid noise of the literature: patches of sinusoids at five scales, six orientations and two phases, 4,092 random contrasts per pattern. Every session uses the same 300 patterns (400 or 500 as a setting), each in its own order, so the images of different people can be averaged into group images and compared. Four practice trials come first, and a short rest follows every 100 trials.
The main score is the informational value (infoVal) of the classification image (Brinkman et al., 2019): how far the image's length lies above the lengths of images made by random choices on the same patterns, in robust standard deviations, with the corrections of Schmitz, Rougier and Yzerbyt (2020). Above 1.96 the image holds more signal than random choices would give, and Brinkman and colleagues advise interpreting only such images. The classification image is stored with each result as the 4,092 values of its noise, drawn on the base face on the researcher's result page and downloadable there; it is research output and is never shown to the participant. The task also reports the share of choices of the left face and the median choice time. Presses within 100 ms of the faces are counted as anticipations and not taken as answers. Scores support research and do not diagnose.
Scores
- Informational Value (infoVal)z
- How far the classification image's length lies above the lengths of images from random choices on the same patterns, in robust standard deviations (Brinkman et al., 2019). 1.96 or more means more signal than random choices would give. Null under 100 trials.
- Under 1.96 is flagged for review
- Classification Imagescore
- The 4,092 noise values of the mean chosen pattern, keyed by parameter number from 0: +1 for the face with a pattern added, -1 for the face with it subtracted. Drawn on the base face on the result page; averaged across people for a group image.
- Research output, never shown to the participant
- Left Face Chosen%
- Percent of choices of the face on the left. The face with the pattern added is on the left on half the trials.
- Quality check
- Median Choice Timems
- Median time from the faces appearing to the choice, over choices at or after 100 ms.
- Informational; the answer has no time limit
- Trials Answeredtrials
- Choices in the scored run.
- Fewer than planned is flagged for review
- Planned Trialstrials
- Trials the session planned: 300, or 400 or 500 as a setting.
- Informational
- Anticipationscount
- Presses within 100 ms of the faces appearing, not taken as answers.
- Quality check
Use and evidence
39 papers from 2021 to 2025 name this task in their title or abstract, up 34% compared with 2016 to 2020. That ranks it 273 of 395 test families in our census. See the counts for every test.
Show the numbers
| Year | Papers |
|---|---|
| 2025 | 7 |
| 2024 | 12 |
| 2023 | 5 |
| 2022 | 6 |
| 2021 | 9 |
| 2020 | 4 |
| 2019 | 9 |
| 2018 | 6 |
| 2017 | 6 |
| 2016 | 4 |
| 2015 | 4 |
| 2014 | 6 |
| 2013 | 4 |
| 2012 | 1 |
| 2011 | 2 |
| 2010 | 5 |
| 2009 | 0 |
| 2008 | 0 |
| 2007 | 0 |
| 2006 | 1 |
| 2005 | 1 |
| 2004 | 1 |
| 2003 | 0 |
| 2002 | 0 |
| 2001 | 0 |
| 2000 | 0 |
Use grade D
30 to 99 papers from 2021 to 2025.
Evidence grade C
At least 2 test-retest papers, 30 reliability or validity papers or 3 registered clinical studies.
- Papers on test-retest reliability
- 2
- Papers on reliability or validity
- 6
- Registered clinical studies using it as an outcome
- 0
Note on the counts: a few counted papers use reverse correlation in neurophysiology.
How the grades work
Use: A is 1,000 or more papers from 2021 to 2025, B 300 to 999, C 100 to 299, D 30 to 99 and E fewer than 30. Evidence: A means at least 25 papers report its test-retest reliability and at least 25 registered clinical studies use it as an outcome; B, at least 8 test-retest papers, and either 5 registered clinical studies or 150 reliability or validity papers; C, at least 2 test-retest papers, 30 reliability or validity papers or 3 registered clinical studies; D, little published measurement evidence found. The counts measure use. They include every paper that names the task in its title or abstract. Grades describe the published research on this task, not ConductCognition's version of it.
Source: ConductCognition test census 0.14.0, counted 2 November 2026 from OpenAlex and ClinicalTrials.gov.
Task design
The trial sequence as the test runs in ConductCognition.
- Face
- One face, the average of the neutral photographs of 96 people of the Face Research Lab London Set, with women and men, and each of the set's four groups, weighted equally: 256 by 256 px in gray, shown twice side by side.
- Noise
- Patches of sinusoids of 2 cycles at 5 scales (1, 2, 4, 8 and 16 patches to a side), 6 orientations and 2 phases: 4,092 random contrasts from -1 to 1 per pattern. One face has the pattern added and the other has it subtracted; the added face is on the left on half the trials.
- Question
- Which face looks more trustworthy (standard), more dominant or happier, answered by click, tap, or F and J, with no time limit. Presses within 100 ms of the faces are counted and not taken as answers. A 300 ms blank follows each choice.
- Trials
- 300 (standard; 400 or 500), the first patterns of one fixed set of 500, in an order drawn for each session. A rest after every 100.
- Practice
- 4 pairs of other patterns, each followed by a note of the face chosen for 1,500 ms.
- Classification image
- The mean of the chosen patterns, +1 for a face with the pattern added and -1 for a face with it subtracted, stored as 4,092 values and drawn on the base face on the researcher's result page. Never shown to the participant.
- Informational value
- The image's length against the lengths of 10,000 images from random choices on the same patterns: (length - median) / (1.4826 x median absolute deviation). 1.96 or more marks an image with signal.
The method follows Mangini and Biederman (2004) and Dotsch and Todorov (2012) with a face and noise of our own: the face is an average our code made, so no real person's face is shown, and the noise is drawn in the browser from each pattern's number. The informational value uses the Euclidean norm and the 1.4826 constant, as Schmitz, Rougier and Yzerbyt (2020) correct the published code. No option asks about a social group, and no screen labels a face.
Quality checks
Each session is marked valid, needs review or invalid, with the reason, before you see its scores.
Scored rows: reverse_correlation_test from the scored run, one choice between two noisy faces per trial with no time limit. No face is the right choice, so accuracy is not judged against chance. The same side on at least 90% of choices invalidates. Choices under 200 ms on at least half the trials invalidate. A run short of its planned trials and an informational value under 1.96 are flagged for review.
Why these rules
There is no correct option to compare with chance, so accuracy is not judged against chance. The face with the pattern added is on the left on half the trials, drawn at random, so choosing one side nearly every time leaves a classification image of random noise; such a session is invalid. Brinkman et al. (2019) advise against interpreting an image whose informational value is under 1.96. An answer under 200 ms is faster than a simple visual reaction, so it cannot reflect the item. Following response-time effort screening for untimed tests (Wise and Kong 2005), the session is invalid only when at least half the answers are that fast.
No percentile. Scores and quality flags only.
Running this test
Settings you can change
| Setting | Standard | Allowed |
|---|---|---|
| Trait | Trustworthy (standard) | Trustworthy (standard), Dominant, Happy |
| Number of trials | 300 (standard) | 300 (standard), 400, 500 |
Yearly plans can change settings within these limits. Free and pay-as-you-go studies use the standard settings.
Need a setting this test does not have?
Devices and languages
- Untimed test: runs on desktops, laptops and tablets. Participant devices
- Languages: English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese.
- Available on: Every plan, including the free plan.
- No percentile. Scores and quality flags only. How norms work
- Recruiting online? See the setup guides for Prolific, Sona Systems, CloudResearch Connect and Qualtrics.
Cite and describe
Participants completed the Reverse Correlation Task in a web browser (ConductCognition task version 1.0.0, scoring version 1.0.0; Mangini & Biederman, 2004; Dotsch & Todorov, 2012; Brinkman et al., 2019), with the standard settings (trait: Trustworthy; number of trials: 300). The main score was informational value (infoVal; how far the classification image's length lies above the lengths of images from random choices on the same patterns, in robust standard deviations (Brinkman et al., 2019). 1.96 or more means more signal than random choices would give).
ConductCognition versions: task version 1.0.0, scoring version 1.0.0. Every export records the versions each session used, for a methods section.
- Mangini, M. C., & Biederman, I. (2004). Making the ineffable explicit: Estimating the information employed for face classifications. Cognitive Science, 28(2), 209-226. https://doi.org/10.1207/s15516709cog2802_4
- Dotsch, R., & Todorov, A. (2012). Reverse correlating social face perception. Social Psychological and Personality Science, 3(5), 562-571. https://doi.org/10.1177/1948550611430272Two noisy versions of one base face, the noise and its opposite; trustworthiness and dominance.
- Brinkman, L., Todorov, A., & Dotsch, R. (2017). Visualising mental representations: A primer on noise-based reverse correlation in social psychology. European Review of Social Psychology, 28(1), 333-361. https://doi.org/10.1080/10463283.2017.1381469The sinusoid noise of 4,092 parameters, base faces, 300 to 1,000 trials, and drawing classification images.
- Brinkman, L., Goffin, S., van de Schoot, R., van Haren, N. E. M., Dotsch, R., & Aarts, H. (2019). Quantifying the informational value of classification images. Behavior Research Methods, 51(5), 2059-2073. https://doi.org/10.3758/s13428-019-01232-2The informational value and its 1.96 threshold.
- Schmitz, M., Rougier, M., & Yzerbyt, V. (2020). Comment on "Quantifying the informational value of classification images": A miscomputation of the infoVal metric. Behavior Research Methods, 52(3), 1383-1386. https://doi.org/10.3758/s13428-019-01295-1The Euclidean norm and the 1.4826 constant.
- Oosterhof, N. N., & Todorov, A. (2008). The functional basis of face evaluation. Proceedings of the National Academy of Sciences, 105(32), 11087-11092. https://doi.org/10.1073/pnas.0805664105Trustworthiness and dominance as the two main dimensions of face evaluation.
- DeBruine, L., & Jones, B. (2017). Face Research Lab London Set [Data set]. figshare. https://doi.org/10.6084/m9.figshare.5047666.v5
Source papers
- Brinkman L, Goffin S, van de Schoot R, van Haren NEM, Dotsch R, Aarts H (2019). Quantifying the informational value of classification images. Behavior Research Methods, 51(5):2059-2073.Paradigm source, Adults
- DeBruine L, Jones B (2017). Face Research Lab London Set. figshare. Dataset. doi:10.6084/m9.figshare.5047666.v5Paradigm source, Adults
Questions
How long does the Reverse Correlation Task take?
About 8 to 10 minutes, as listed in the test library.
What device does a participant need?
Untimed test: runs on desktops, laptops and tablets. The browser must be Chrome, Edge, Firefox or Safari.
What scores does it produce?
Informational Value (infoVal), Classification Image, Left Face Chosen, Median Choice Time, Trials Answered, Planned Trials and Anticipations. Each session is also marked valid, needs review or invalid, with the reason.
How widely is the Reverse Correlation Task used in research?
39 papers from 2021 to 2025 name it in their title or abstract, up 34% compared with 2016 to 2020. That ranks it 273 of 395 test families in the ConductCognition test census (use grade D). Source: ConductCognition test census 0.14.0, counted 2 November 2026 from OpenAlex and ClinicalTrials.gov.
How much reliability evidence does the Reverse Correlation Task have?
In the published research, 2 papers report its test-retest reliability and 0 registered clinical studies use it as an outcome (evidence grade C in the ConductCognition test census).
Does it report a percentile?
No. Scores and quality flags only.
Which settings can a researcher change?
Trait (Trustworthy (standard), Dominant, Happy) and number of trials (300 (standard), 400, 500). Settings can be changed on yearly plans; free and pay-as-you-go studies use the standard settings.
Which languages can participants use?
Participants can take this test in English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese. English is on every plan, and the other languages come with yearly plans. Have a native speaker read the machine-translated text before launch.
Is this a clinical test?
No. ConductCognition is for research use only. It runs, scores and exports the test; clinical interpretation is outside the platform.
Every test: the full test list.
Run this test in your next study
Start free with 10 free participants on any test, then use participant credits or a yearly plan.