@tadegquillien.bsky.social
Cognitive scientist at the University of Edinburgh. Causality, computation, evolution.
Lab: https://quillienlab.github.io/
Lab: https://quillienlab.github.io/
A new study by Zach Horne et al. combines history of science and psychology experiments to document the appeal of 'intrinsic' explanations: scientists and laypeople are drawn to explanations that appeal to an object's inherent properties.
www.pnas.org/doi/10.1073/...
www.pnas.org/doi/10.1073/...
September 22, 2025 at 4:36 PM
A new study by Zach Horne et al. combines history of science and psychology experiments to document the appeal of 'intrinsic' explanations: scientists and laypeople are drawn to explanations that appeal to an object's inherent properties.
www.pnas.org/doi/10.1073/...
www.pnas.org/doi/10.1073/...
The problem is that this strategy would also allow us to 'infer' that people have intuitive preferences for nuclear explosions, or sending people to concentration camps:
September 18, 2025 at 5:04 PM
The problem is that this strategy would also allow us to 'infer' that people have intuitive preferences for nuclear explosions, or sending people to concentration camps:
People tend to attribute more intentionality to agents that did something bad. It is tempting to use this effect as a way to covertly measure people's implicit attitudes, as Will Gervais and colleagues do in a recent paper on attitudes toward atheism:
www.pnas.org/doi/10.1073/...
www.pnas.org/doi/10.1073/...
September 18, 2025 at 5:03 PM
People tend to attribute more intentionality to agents that did something bad. It is tempting to use this effect as a way to covertly measure people's implicit attitudes, as Will Gervais and colleagues do in a recent paper on attitudes toward atheism:
www.pnas.org/doi/10.1073/...
www.pnas.org/doi/10.1073/...
We also find that the optimal policy exhibits the same 'outgroup homogeneity bias' as people: it tends to represent outgroup members are more similar to each other than they are. Again, this is especially the case when cognitive resources are very limited.
August 2, 2025 at 5:57 PM
We also find that the optimal policy exhibits the same 'outgroup homogeneity bias' as people: it tends to represent outgroup members are more similar to each other than they are. Again, this is especially the case when cognitive resources are very limited.
Our key result: under limited resources, the optimal policy preferentially encodes information about group membership (blue) and tends to discard individuating information (teal).
This tendency only reverses if group membership has very low task relevance.
This tendency only reverses if group membership has very low task relevance.
August 2, 2025 at 5:57 PM
Our key result: under limited resources, the optimal policy preferentially encodes information about group membership (blue) and tends to discard individuating information (teal).
This tendency only reverses if group membership has very low task relevance.
This tendency only reverses if group membership has very low task relevance.
We study agents who have to predict the behavior of other agents.
Agents have limited cognitive resources: they can only extract so much information from the environment, and have to prioritize which information to encode (group membership or individuating info).
Agents have limited cognitive resources: they can only extract so much information from the environment, and have to prioritize which information to encode (group membership or individuating info).
August 2, 2025 at 5:56 PM
We study agents who have to predict the behavior of other agents.
Agents have limited cognitive resources: they can only extract so much information from the environment, and have to prioritize which information to encode (group membership or individuating info).
Agents have limited cognitive resources: they can only extract so much information from the environment, and have to prioritize which information to encode (group membership or individuating info).
Comparison to alternative models shows the data can't be explained by a direct effect of temporal order on causal judgments: instead the effect is mediated by counterfactual simulation.
July 28, 2025 at 5:23 PM
Comparison to alternative models shows the data can't be explained by a direct effect of temporal order on causal judgments: instead the effect is mediated by counterfactual simulation.
The cool thing is that we are able to reverse-engineer how people simulate counterfactuals by fitting model parameters.
As expected, recent events are less 'stable', i.e. more likely to be replaced by counterfactual alternatives:
As expected, recent events are less 'stable', i.e. more likely to be replaced by counterfactual alternatives:
July 28, 2025 at 5:21 PM
The cool thing is that we are able to reverse-engineer how people simulate counterfactuals by fitting model parameters.
As expected, recent events are less 'stable', i.e. more likely to be replaced by counterfactual alternatives:
As expected, recent events are less 'stable', i.e. more likely to be replaced by counterfactual alternatives:
When implemented via the Counterfactual Effect Size model, these assumptions are enough to explain the novel effects found by Thanawala & Erb (human data in grey, model in orange):
July 28, 2025 at 5:19 PM
When implemented via the Counterfactual Effect Size model, these assumptions are enough to explain the novel effects found by Thanawala & Erb (human data in grey, model in orange):
Counterfactual models predict that normality should influence causal judgments in a different way depending on causal structure.
A fascinating paper by Ozdemir and Walker finds some hints of this pattern in 5- to 7-year old children.
static1.squarespace.com/static/5615d...
A fascinating paper by Ozdemir and Walker finds some hints of this pattern in 5- to 7-year old children.
static1.squarespace.com/static/5615d...
July 22, 2025 at 4:14 PM
Counterfactual models predict that normality should influence causal judgments in a different way depending on causal structure.
A fascinating paper by Ozdemir and Walker finds some hints of this pattern in 5- to 7-year old children.
static1.squarespace.com/static/5615d...
A fascinating paper by Ozdemir and Walker finds some hints of this pattern in 5- to 7-year old children.
static1.squarespace.com/static/5615d...
New paper from Qi, Vul and Powell introduces a clever way to measure a participant's Welfare-Tradeoff Ratio in a single trial: journals.plos.org/plosone/arti...
July 9, 2025 at 4:08 PM
New paper from Qi, Vul and Powell introduces a clever way to measure a participant's Welfare-Tradeoff Ratio in a single trial: journals.plos.org/plosone/arti...
Really cool paper by Pajot et al. showing that a Language of Thought for arithmetic predicts the frequencies of numbers in natural language.
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
July 6, 2025 at 4:30 PM
Really cool paper by Pajot et al. showing that a Language of Thought for arithmetic predicts the frequencies of numbers in natural language.
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
As shown by Tversky and Kahneman, people also often judge that Pr(A&B) > Pr(A).
The conjunction fallacy makes sense given the logic of guessing: sometimes "A&B" conveys a better picture of your subjective distribution than "A".
The conjunction fallacy makes sense given the logic of guessing: sometimes "A&B" conveys a better picture of your subjective distribution than "A".
June 30, 2025 at 4:23 PM
As shown by Tversky and Kahneman, people also often judge that Pr(A&B) > Pr(A).
The conjunction fallacy makes sense given the logic of guessing: sometimes "A&B" conveys a better picture of your subjective distribution than "A".
The conjunction fallacy makes sense given the logic of guessing: sometimes "A&B" conveys a better picture of your subjective distribution than "A".
We tested this idea by asking people to rate the quality of interval guesses, relative to the correct answer.
People tend to say that a guess is good if it encodes a distribution (as inferred by the model) that assigns high probability to the correct answer.
People tend to say that a guess is good if it encodes a distribution (as inferred by the model) that assigns high probability to the correct answer.
June 30, 2025 at 4:19 PM
We tested this idea by asking people to rate the quality of interval guesses, relative to the correct answer.
People tend to say that a guess is good if it encodes a distribution (as inferred by the model) that assigns high probability to the correct answer.
People tend to say that a guess is good if it encodes a distribution (as inferred by the model) that assigns high probability to the correct answer.
Sometimes people guess a quantity by giving an interval (e.g. "I think there are between 165 and 185 countries in the UN").
We argue that these intervals implicitly encode a subjective probability distribution: the interval midpoint and width encode the mean and variance.
We argue that these intervals implicitly encode a subjective probability distribution: the interval midpoint and width encode the mean and variance.
June 30, 2025 at 4:19 PM
Sometimes people guess a quantity by giving an interval (e.g. "I think there are between 165 and 185 countries in the UN").
We argue that these intervals implicitly encode a subjective probability distribution: the interval midpoint and width encode the mean and variance.
We argue that these intervals implicitly encode a subjective probability distribution: the interval midpoint and width encode the mean and variance.
We compared people's answers with a Bayesian decoder calibrated with the production data from our previous study. This normative benchmark fits human inferences well, without fitting any additional parameter.
June 30, 2025 at 4:18 PM
We compared people's answers with a Bayesian decoder calibrated with the production data from our previous study. This normative benchmark fits human inferences well, without fitting any additional parameter.
But do guesses really communicate distributional information?
To find out, we ran an 'inverted' version of the previous study. Participants were given a guess made by someone else, and had to infer which box the speaker was looking at.
To find out, we ran an 'inverted' version of the previous study. Participants were given a guess made by someone else, and had to infer which box the speaker was looking at.
June 30, 2025 at 4:17 PM
But do guesses really communicate distributional information?
To find out, we ran an 'inverted' version of the previous study. Participants were given a guess made by someone else, and had to infer which box the speaker was looking at.
To find out, we ran an 'inverted' version of the previous study. Participants were given a guess made by someone else, and had to infer which box the speaker was looking at.
Our information-theoretic model closely tracks the proportion of participants who generate a given guess.
(Sometimes participants looking at the same box make different guesses. This typically happens when the model thinks these are equally good guesses!)
(Sometimes participants looking at the same box make different guesses. This typically happens when the model thinks these are equally good guesses!)
June 30, 2025 at 4:17 PM
Our information-theoretic model closely tracks the proportion of participants who generate a given guess.
(Sometimes participants looking at the same box make different guesses. This typically happens when the model thinks these are equally good guesses!)
(Sometimes participants looking at the same box make different guesses. This typically happens when the model thinks these are equally good guesses!)
In a second study, we asked participants to compose their own guesses:
June 30, 2025 at 4:16 PM
In a second study, we asked participants to compose their own guesses:
In a first study we asked participants to rate the quality of different guesses one could make in different contexts.
Their ratings are well-predicted by an information-theoretic measure of how faithfully the guess conveys the shape of the true distribution.
Their ratings are well-predicted by an information-theoretic measure of how faithfully the guess conveys the shape of the true distribution.
June 30, 2025 at 4:16 PM
In a first study we asked participants to rate the quality of different guesses one could make in different contexts.
Their ratings are well-predicted by an information-theoretic measure of how faithfully the guess conveys the shape of the true distribution.
Their ratings are well-predicted by an information-theoretic measure of how faithfully the guess conveys the shape of the true distribution.
In this context, we predict that people make guesses that would allow a listener to approximately re-construct the speaker's subjective probability distribution.
June 30, 2025 at 4:15 PM
In this context, we predict that people make guesses that would allow a listener to approximately re-construct the speaker's subjective probability distribution.
To test our proposal systematically, we ask participants to make guesses about what color will come out in a random draw from an urn like this one:
June 30, 2025 at 4:14 PM
To test our proposal systematically, we ask participants to make guesses about what color will come out in a random draw from an urn like this one:
Our point is that this kind of mistake actually makes sense if people are implicitly solving a different problem than the one given by the experimenter.
Namely, their guess functions to encode their subjective probability distribution over what Danielle might be studying:
Namely, their guess functions to encode their subjective probability distribution over what Danielle might be studying:
June 30, 2025 at 4:13 PM
Our point is that this kind of mistake actually makes sense if people are implicitly solving a different problem than the one given by the experimenter.
Namely, their guess functions to encode their subjective probability distribution over what Danielle might be studying:
Namely, their guess functions to encode their subjective probability distribution over what Danielle might be studying:
I wrote a short blog post about our experimental work on the logic of guesses:
xphi.net/2025/06/03/e...
xphi.net/2025/06/03/e...
June 11, 2025 at 3:25 PM
I wrote a short blog post about our experimental work on the logic of guesses:
xphi.net/2025/06/03/e...
xphi.net/2025/06/03/e...
Our #CogSci2025 paper led by Madeleine Horner (with Adam Moore) explores people's lay conception of empathy.
Main finding: people have a robust expectation that agents who feel empathy are more likely to help, especially when helping is costly.
quillienlab.github.io/Horner%20et%...
Main finding: people have a robust expectation that agents who feel empathy are more likely to help, especially when helping is costly.
quillienlab.github.io/Horner%20et%...
April 29, 2025 at 3:51 PM
Our #CogSci2025 paper led by Madeleine Horner (with Adam Moore) explores people's lay conception of empathy.
Main finding: people have a robust expectation that agents who feel empathy are more likely to help, especially when helping is costly.
quillienlab.github.io/Horner%20et%...
Main finding: people have a robust expectation that agents who feel empathy are more likely to help, especially when helping is costly.
quillienlab.github.io/Horner%20et%...