Matan Mazor
LightNews LightNews
banner
matanmazor.bsky.social
Matan Mazor
@matanmazor.bsky.social
3.5K followers 250 following 550 posts
Post-doctoral research fellow in cognitive neuroscience (Oxford), interested in complex systems and in simple systems who believe they are complex systems
Posts Replies Media Videos
matanmazor.bsky.social
This is truly amazing: some LLMs can detect when concepts are injected into their internal activations.

transformer-circuits.pub/2025/introsp...
November 17, 2025 at 12:14 PM
matanmazor.bsky.social
The human need for rationalizations has shaped consciousness science since its inception. Descartes saw his view that animals were unconscious automatons “not so much cruel to beasts but respectful to human beings… whom it absolves from any suspicion of crime whenever they kill or eat animals” (Descartes, 1999, cited in Kaldas, 2015). More recently, a new wave of biological naturalism, arguing that consciousness is inherently a property of living beings (Aru et al., 2023; Findlay et al., 2024; Seth, 2025) coincided with concerns about the potentially devastating moral implications of conscious AI (Long et al., 2024). As with Fleming and Michel’s fish, or Descartes’ beasts, there may be good epistemic reasons to favour a theory in which AI cannot be conscious. Yet a full understanding of consciousness science requires taking seriously the force of rationalizations in theory formation and selection.
October 10, 2025 at 7:52 PM
matanmazor.bsky.social
מביך בשביל האוניברסיטה. טוב לדעת שלפחות הרקטור, מונא מארון, תומכת במכתב (מתוך נאום שלה בטקס בוגרי תואר בחינוך)
July 30, 2025 at 1:03 PM
matanmazor.bsky.social
See the full 6 pages for perceptual and meta-perceptual learning effects, prospective confidence, and one very honest closing paragraph:
Finally, while Exp. 2 confirmed that confidence in absence is affected by stimulus size in a way that parallels the effect of size on prospective confidence, it also produced a series of disappointing null findings, with no significant effect of size on confidence in presence and no correlations between size effects on post-decisional confidence and on prospective confidence or perceptual sensitivity. Disappointed by these null findings, we reminded ourselves that Exp. 2 had a smaller sample size and only 8 trials per block (half the number of trials in Exp. 1). As such, we are not confident that these correlations are truly absent — we cannot be certain that we would have detected them if they were present.
July 23, 2025 at 3:39 AM
matanmazor.bsky.social
Despite not being fitted to confidence ratings, model parameters correlated with confidence across subjects. Specifically, a "true visibility" parameter, derived from decisions and RTs, predicted confidence in presence, and a "believed visibility" parameter predicted confidence in absence.
July 23, 2025 at 3:34 AM
matanmazor.bsky.social
We also fitted our ideal-observer model from psycnet.apa.org/fulltext/202..., in which agents observe noisy sensor activations and decide between PRESENT, ABSENT, or WAIT, as a function of their impatience and, crucially, their beliefs about the expected visibility of stimuli.
A) A schematic description of the ideal observer model. B) Comparison of human and simulated data for Exp. 1, for response time, error rate, and confidence. Lines and points represent human data; semi-transparent rectangles represent data simulated from model parameters, which were fitted to reaction time and accuracy (but not confidence) data of individual participants. Rectangles are centred at their mean value, and their height is twice the standard error.
July 23, 2025 at 3:30 AM
matanmazor.bsky.social
The beautiful thing is that confidence in presence tracked true visibility (higher for smaller stimuli), but confidence in absence tracked believed visibility (higher for bigger stimuli). Only the effect of size on confidence in presence correlated with its effect on objective sensitivity.
A) visual sensitivity (measured as d’) as a function of stimulus size. B) mean confidence in correct responses as a function of target presence and size. Error bars represent the standard error; ***: p<.001. C) correlations between the effect of size on perceptual sensitivity and on confidence, separately for target-present and target-absent decisions.
July 23, 2025 at 3:29 AM
matanmazor.bsky.social
Maya and I used a meta-perceptual illusion to obtain a clear behavioural dissociation. In our task, targets were more visible when the display was small, but participants systematically believed that bigger stimuli made targets easier to see.
Experiment 1. A) Stimuli. Left: example frames from target-present (blue) and target-absent (red) trials. Stimuli refreshed at 15 frames per second until a response was recorded. Right: on different trials, stimuli appeared big or small. C) trial structure. Participants performed two 16-trial detection blocks in which the target was the letter S and two blocks in which the target was the letter A. Stimulus presentation was preceded by a frame, and followed by an analog confidence rating scale. The order of the two letters was randomised between participants.
July 23, 2025 at 3:27 AM
matanmazor.bsky.social
Confidence in absence as confidence in counterfactual visibility: a CogSci proceedings paper with star MSc student Maya Schipper, is now out on PsyArXiv:

osf.io/preprints/ps...

🧵👇
July 23, 2025 at 3:23 AM
matanmazor.bsky.social
my laptop:
April 9, 2025 at 7:45 PM
matanmazor.bsky.social
The paper also features fun things like reverse correlation, backwards induction, TAYLOR SWIFT, a new non-parametric sign-consistency test @itayyaron.bsky.social, and some evidence that the Bayesian p(correct) is not a bad approximation of subjective confidence in this task.
Some stuff from the paper: reverse correlation, sign consistency, model predictions for subjective confidence, equations, and taylor swift
March 21, 2025 at 4:59 PM
matanmazor.bsky.social
Third, we demonstrate this principle with partial occlusion. The more we occluded the display, people missed more targets. But occlusion also made *some* people more likely to make “present” responses on target-absent trials: a reflection of their belief that they may have been missing the target.
A figure detailing the experimental paradigm, in which participants made decisions about the presence or absence of letters under different levels of stimulus occlusion.
March 21, 2025 at 4:59 PM
matanmazor.bsky.social
Second, we show that a model parameter that controls the rate of perceptual evidence affects decisions when a target is present, but it is a parameter that controls *meta-perceptual beliefs about the rate of perceptual evidence* that controls decisions when a target is absent.
A table of correlations between model parameters and occlusion effects on accuracy and reaction times, showing that correlations are significant between the parameter that controls the true visibility of stimuli and behaviour on target-present trials, and between the parameter that controls the believed visibility of stimuli and behaviour on target-absent trials.
March 21, 2025 at 4:59 PM
matanmazor.bsky.social
First, we propose a new model of perceptual detection as a series of momentary decisions between 3 actions: decide present / absent, or accumulate more evidence. The perceptual evidence that goes into this process is noisy, but the decisions are optimal for maximising discounted long-term reward.
An ideal observer model of visual detection
March 21, 2025 at 4:59 PM
matanmazor.bsky.social
Our "I would have seen it if it were there" paper — a collaboration with @ranimo.bsky.social and @clarepress.bsky.social — is now out in Psych. Review.

There’s a lot in this paper, but here are what I see as the 3 main takeaways:

psycnet.apa.org/fulltext/202...
The title of the paper, "Beliefs about Perception Shape Perceptual Inference: An Ideal Observer Model of Detection", with the author names, and Fig. 1: a schematic of an inverse optics account of vision.
March 21, 2025 at 4:48 PM
matanmazor.bsky.social
Kahneman took his science seriously.

www.themarker.com/wallstreet/2...
March 15, 2025 at 2:34 PM
matanmazor.bsky.social
Revising an ethics application, and making sure not to make the participants that I am experimenting upon feel like they are being 'experimented upon'
March 14, 2025 at 10:43 AM
matanmazor.bsky.social
From June Huh's Fields Medal interview
While browsing through arXiv postings, rambling between shelves of libraries, or sitting in the dining room of an institute, I met heroes and heroines. The last decade of my life was like living a chapter from ancient mythology. There were hundreds of characters, each one of them with a unique set of abilities. It was, and still is, a privilege to have connected with their minds. On clear days, I can see that I am a small and simple part of a big and complex ancient structure, like a mushroom sprouted from a giant underground fungal network. It feels good.
November 20, 2024 at 8:32 AM
matanmazor.bsky.social
George Ainslie on clarity in scientific writing
November 3, 2024 at 1:41 PM
matanmazor.bsky.social
10/11 subject-level model-fitting revealed a beautiful dissociation: while occlusion effects on accuracy and reaction times in target-present trials were a function of visibility, the same effects in target-absent trials were driven by beliefs about visibility.
June 25, 2024 at 1:37 PM
matanmazor.bsky.social
@ItayYaron @Liad_Mudrik @n_faivre 10/11 subject-level model-fitting revealed a beautiful dissociation: while occlusion effects on accuracy and reaction times in target-present trials were a function of visibility, the same effects in target-absent trials were driven by beliefs about...
June 25, 2024 at 1:36 PM
matanmazor.bsky.social
9/11 (to show that this was true variation and not just noise we used the sign-consistency R package that superstar Itay Yaron developed, together with @liadmudrik.bsky.social , @nfaivre.bsky.social and myself: github.com/mufcItay/sig...). Read more here: osf.io/preprints/psyarxiv/ebg8w
June 25, 2024 at 1:35 PM
matanmazor.bsky.social
9/11 (to show that this was true variation and not just noise we used the sign-consistency R package that superstar @ItayYaron developed, together with @Liad_Mudrik, @n_faivre and myself: https://github.com/mufcItay/signcon). Read more here: <a href="http://osf.io/preprints/psyarxiv/ebg8w" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">http://osf.io/preprints/psyarxiv/ebg8w
June 25, 2024 at 1:34 PM
matanmazor.bsky.social
7/11 so if decisions in target-absent trials are sensitive to occlusion, this is due to something other than visibility itself. According to our model, this something is beliefs about visibility.
June 25, 2024 at 1:29 PM
matanmazor.bsky.social
7/11 so if decisions in target-absent trials are sensitive to occlusion, this is due to something other than visibility itself. According to our model, this something is beliefs about visibility.
June 25, 2024 at 1:28 PM