Sign in

Akshay K. Jagadish

@akjagadish.bsky.social
176 followers 495 following 42 posts

Interested in natural and artificial minds. akjagadish.github.io

PostsRepliesMedia
Reposted by Akshay K. Jagadish
Brenden Lake @brendenlake.bsky.social · 26/08/2026
Excited about new work from my CCN keynote (38m): youtu.be/v3J-vJfxhOE?... We often choose between Bayesian and neural net models of behavior. But what if there's a spectrum, and combining both makes better predictions? Introducing BBT w @akjagadish.bsky.social & Guangyuan arxiv.org/abs/2608.22154
0356
Reposted by Akshay K. Jagadish
Eric Schulz @ericschulz.bsky.social · 27/06/2026
Very excited about this. One of the first steps to truly automate the discovery process in the cognitive sciences.
071
Reposted by Akshay K. Jagadish
Griffiths Computational Cognitive Science Lab @cocoscilab.bsky.social · 27/06/2026
New preprint explores how using AI to discover psychological theories and propose experiments to test them can create an automated cognitive scientist
092
Reposted by Akshay K. Jagadish
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
1/ 🚨 New preprint: "Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist" Introducing AutoCog 🤖 — a fully autonomous AI system that runs the entire scientific discovery cycle in cognitive science to surface novel theories of human behavior 🧵
1185
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
Co-let w/ @younesstrittmatter.bsky.social Kudos to the all-start team of @georgekafetzis.bsky.social @norijacoby.bsky.social, @ericschulz.bsky.social, @nathanieldaw.bsky.social, @suyoghc.bsky.social, and @cocoscilab.bsky.social
083
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
18/ If you’re interested in automating parts of the cognitive science workflow, you should also check out this great concurrent work by @benpry.bsky.social and colleagues bsky.app/profile/benp...
020
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
17/ 📄 Paper: arxiv.org/abs/2606.26448 📋 Preregistration: osf.io/f7kes
arxiv.org
Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist
Across the sciences, autonomous systems are increasingly being used in closed-loop discovery, proposing new theories and designing and running experiments to test them. This approach is yet to be appl...
220
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
16/ 🔭 The beginning of a new era in cognitive science? - A way to accelerate the search for unified theories of cognition (Newell 1990) - A testbed to study the trajectory of discovery itself (Simon 1977)
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
15/ 👩‍🔬 Where do human scientists fit? Role shifts from execution to specification — asking which theories generalize across hundreds of experiments, integrate from computation to implementation, and unify domains. The loop does the search; every run is a machine-readable trace others can resume.
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
14/ ⚠️ Not a magic bullet yet. - Theories advance based on generalization — simplicity, novelty, and unification aren't scored directly. - Can't yet verify the code faithfully realizes the verbal theory. - LLMs still lean toward training-data mechanisms; creativity needs dedicated optimization.
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
13/ 🔬 We preregistered the predictions (osf.io/f7kes) and tested them on new participants. All three held: it out-predicted WADD, Tallying & TTB on discriminating stimuli; equal advantages mattered more in the low-value region; and the effect weakened when shifted up. An echo of prospect theory.
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
12/ 🆕 A cardinal-cue variant: choices show diminishing sensitivity to feature value. People seem to pass cue values through a concave, diminishing-returns transform before weighting by validity — so a given advantage counts for more at low values than high (0→1 > 4→5).
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
11/ 🪢 Classic MAD, binary cues: subjective weights follow a power function of validity. One subject-specific exponent subsumes three heuristics — →0 Tallying, =1 WADD, large TTB. Parsimonious too (cf. Rmus et al. 2026), surfaced with no parameter-fitting and no explicit pressure for parsimony.
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
10/ 🧑‍🔬 What happens with real humans? Surfaced theories keep improving over cycles — beating the seeds on all ten datasets, with the best theory's error dropping ~an order of magnitude. They also generalized to a held-out decision study, again outperforming the seeds.
110
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
9/ ✅ Can it recover known ground-truth theories? After five cycles, yes — canonical heuristics (TTB, Tallying, WADD) and non-canonical ones (Take-the-Worst, Single-cue). Fed pure noise, it returns a random-guessing model. Discoveries are driven by the DATA, not the LLM's priors.
110
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
8/ 🔁 A case study: seeded with Tallying and Take-the-Best, with behavior simulated from Weighted Additive (WADD). AutoCog proposes experiments where the incumbents disagree + a metric to capture it, analyzes the gap, arbitrates, and synthesizes a new theory — here, the ground truth: WADD.
Illustrative case-study of AUTOCOG on synthetic WADD-generated responses A single discovery cycle (center)
where AUTOCOG is tasked to recover WADD while being Seeded with TTB and Tallying (top-right). Experimental design(right): the system constructs an experiment where the option pairs (cue matrices for Option A and Option B over five
validity-ordered cues) are chosen to make the seed theories disagree. Behavioral response (bottom): on each designed trial, the
deterministic choices predicted by TTB, Tallying, and WADD. Metric (right): the discriminating statistic on which TTB and
Tallying make opposing choices. Analysis (left): on the TTB-match rate over disagreement trials, the simulated participants fall
between the TTB prediction and the Tallying prediction, nearer Tallying. Theory Revision (top): the Arbitrator, observing that
the data fall squarely between TTB and Tallying, concludes that participants integrate both the number of supporting cues and
their validities and issues a new-theory verdict; the Theory Reviser proposes WADD replacing TTB.
130
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
7/ 🎰 The testbed: multi-attribute decision-making (MAD) — a domain with a rich history and many competing theories. In AutoCog's experiments, people choose between two products described by "expert ratings" of varying reliability.
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
6/ 🔁 How it works LLM agents in each cycle: ① design an experiment where they disagree (adversarial) ② collect behavior: people online, a foundation model, or a cognitive model ③ an arbiter scores both & flags the loser ④ an agent writes the successor
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
5/ 🧩 So what's the bottleneck? Everything's in place except the generative step — theory-building: the creative leap from the accumulated failures of old models to a better one. Unlike theorem-proving, a theory of mind can't be compiled, solved, or synthesized. AutoCog automates exactly this.
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
4/ ⏱️ Why is now the right time for psychology? Experiments run online and are analyzed in code. Three pieces matured at once: large-scale online studies, optimal experimental design, and cognitive models synthesized as code. A closed loop combines all three. tinyurl.com/2fuxprvx
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
3/ 🔥 AI for scientific discovery is everywhere. Across the sciences, self-driving labs that design experiments, collect data, analyze, and iterate are becoming the norm — see recent work from @sakanaai.bsky.social, @futurehousesf.bsky.social @ucberkeleyofficial.bsky.social tinyurl.com/ukctwxep
nature.com
An autonomous laboratory for the accelerated synthesis of inorganic materials - Nature
An autonomous laboratory, the A-Lab, is presented that combines computations, literature data, machine learning and active learning, which discovered and synthesized 41 novel compounds from a set of 5...
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
🧵 TL;DR - recovers known and nuisance decision-making strategies - beats its seed theories on all 10 datasets; generalizes to held-out studies - surfaces 2 new regularities, with 1 preregistered - new future for cognitive scientists: from execution to specification? arxiv.org/abs/2606.26448
arxiv.org
Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist
Across the sciences, autonomous systems are increasingly being used in closed-loop discovery, proposing new theories and designing and running experiments to test them. This approach is yet to be appl...
120
Akshay K. Jagadish @akjagadish.bsky.social · 26/06/2026
1/ 🚨 New preprint: "Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist" Introducing AutoCog 🤖 — a fully autonomous AI system that runs the entire scientific discovery cycle in cognitive science to surface novel theories of human behavior 🧵
1185
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
Joint work in collaboration w/ @milenamr7.bsky.social @kristinwitte.bsky.social @marvinmathony.bsky.social @marcelbinz.bsky.social and @ericschulz.bsky.social
021
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
If you like this space, please keep an eye out! 📄 Prepring: arxiv.org/abs/2603.20988
arxiv.org
Can we automatize scientific discovery in the cognitive sciences?
The cognitive sciences aim to understand intelligence by formalizing underlying operations as computational models. Traditionally, this follows a cycle of discovery where researchers develop paradigms...
011
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
10/ 👩‍🔬 Role of human scientists We define the question, the design space, the model family & the decisive tests theories must pass — the loop amplifies our ability to search, falsify & learn. Surfaced theories are then validated in-vivo with real humans. Human + machine intuition.
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
9/ ⚠️ The risks The central risk isn't technical, it's epistemic: persuasive-looking results that are scientifically hollow. Grammars can be too narrow, synthetic data can drift, program search is rugged, "interestingness" can be gamed. Hack any of these & the theory is moot.
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
8/ 📚 A warning, via Borges The Library of Babel holds every possible book, mostly gibberish. An automated science of the mind risks the same: every possible task, dataset & model — most of it worthless. The goal isn't more books. It's the rare volumes that matter.
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
7/ 🔁 Closing the loop The hardest step. Pure info-gain criteria are myopic — they reward discriminability even when it's scientifically dull. 💡 An AI-agent critic optimizing for interestingness: novelty, parsimony, invariance & generalization (Zhang et al. 2023) w/ @jeffclune.com
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
6/ 🦎 Synthesizing models Replace hand-crafted model classes with LLM-based program synthesis. 💡 Our GeCCo pipeline writes cognitive models as Python programs & refines them on predictive performance — turning modeling into high-throughput search of algorithmic space.
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
5/ 🦙 Generating data Simulate behavior with foundation models of cognition like Centaur, producing de novo data often indistinguishable from humans. 💡 Condition on demographics & psychometrics — e.g. a 30yo with high OCD traits on a planning task (@marcelbinz.bsky.social et al. 2025).
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
4/ 🧪 Proposing experiments Use an LLM as an intelligent experiment sampler — proposing paradigms positioned to elicit a target behavior, then refining on feedback. 💡 A generative grammar over MDPs is one start: it spans bandits to multi-step planning (@mariaeckstein.bsky.social et al. 2025).
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
3/ 🧠 An in-silico science of the mind We envision all four stages implemented fully in silico: ① Propose experiments ② Generate data ③ Synthesize models ④ Close the loop As AI grows more capable, each step is becoming feasible. We're at a genuine inflection point.
An automated cycle of scientific discovery in the cognitive sciences. An experimentalist proposes experiments, a cognitive foundation model generates behavioral data. A modeller proposes and tests different computational models. Finally, a critic judges the “interestingness” of the results and parsed that signal back to the experimentalist for further optimization.
130
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
2/ 💪 Motivation The discovery process in cognitive science follows a 4-stage loop: propose a paradigm → collect data → handcraft & fit models → review & iterate. It works, but the space explored is bottlenecked by researcher intuition, capacity & training — and it's slow.
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
1/ 🚨 New opinion piece: "Can we automatize scientific discovery in the cognitive sciences?" We lay out a vision for a fully automated, in-silico science of the mind, where modern AI systems run every stage of the scientific discovery cycle in cognitive science 🧵 #AutomatedDiscovery #AI4Science
1186
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
Led by the inimitable Liyi Zhang and in collaboration w/ @brendenlake.bsky.social and @cocoscilab.bsky.social
010
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
Check out the preprint and code below 👇 Paper: arxiv.org/pdf/2606.09856 Code: github.com/zhang-liyi/l...
arxiv.org
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
PPT substantially improves both accuracy and calibration in LLM inductive reasoning.
111
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
Introducing program-based posterior training (PPT): 1. Use an LLM to generate open-world scenarios 2. Translate each scenario into a probabilistic programs 3. Run probabilistic inference algorithms to get posterior distributions 4. Fine-tune an LLM on the constructed soft labels
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
Here, we leveraged two of their biggest strengths, in conjunction with post-training, to remedy this: 1. Generating real-world scenarios 2. Program synthesis
arxiv.org
110
Akshay K. Jagadish @akjagadish.bsky.social · 22/06/2026
LLMs are great at lot of things, inductive reasoning is **still** not one of them.
110
Reposted by Akshay K. Jagadish
Griffiths Computational Cognitive Science Lab @cocoscilab.bsky.social · 18/12/2025
Excited to announce a new book telling the story of mathematical approaches to studying the mind, from the origins of cognitive science to modern AI! The Laws of Thought will be published in February and is available for pre-order now.
217041
Reposted by Akshay K. Jagadish
Milena Rmus @milenamr7.bsky.social · 26/05/2025
Preprint update, co-led with @akjagadish.bsky.social, with @marvinmathony.bsky.social, Tobias Ludwig and @ericschulz.bsky.social!
arxiv.org
Generating Computational Cognitive Models using Large Language Models
Computational cognitive models, which formalize theories of cognition, enable researchers to quantify cognitive processes and arbitrate between competing theories by fitting models to behavioral data....
0167
Akshay K. Jagadish @akjagadish.bsky.social · 11/04/2025
Please apply to join us at the Helmholtz Munich. It’s a great lab to do science! ☺️
040
Akshay K. Jagadish @akjagadish.bsky.social · 17/03/2025
Great to see our work being covered by @nytimes.com! I still cannot believe it 😱
070
Akshay K. Jagadish @akjagadish.bsky.social · 11/03/2025
Congrats on the new book, Chris! Looking forward to reading it soon ☺️
010
Reposted by Akshay K. Jagadish
Mirko Thalmann @mirkothm.bsky.social · 03/02/2025
Ever wondered why only some memories 🧠 come easily? Our latest work (osf.io/preprints/ps...) led by S. Haridi, with @ericschulz.bsky.social, shows that targeted memory retrieval speeds up with precise semantic and temporal retrieval cues. Hence, crafting cues can give you instant access to memories⚡
0137
Reposted by Akshay K. Jagadish
Ziv Ben-Zion, PhD @zivbenzion.bsky.social · 08/03/2025
🚀 Just 5 days since publication, our paper is already the #1 trending article on @natureportfolio.nature.com npj Digital Medicine! 🥳 🤖 Check out how we induced anxiety in Chat-GPT using traumatic narratives - then calmed it down with mindfulness & meditation: www.nature.com/articles/s41... 📝
154
Akshay K. Jagadish @akjagadish.bsky.social · 08/03/2025
🚨 Fresh off the press! We showed that one can extract an anxious response from an LLM by including traumatic narratives into the prompt, and interestingly, we can bring the response back to baseline by using mindfulness prompts 😌 It was a fun collaboration, thanks for bringing us on board!! :)
031
Reposted by Akshay K. Jagadish
Milena Rmus @milenamr7.bsky.social · 26/02/2025
About a month late posting this, but here's a new project with @ericschulz.bsky.social, @akjagadish.bsky.social, @marvinmathony.bsky.social and Tobias Ludwig We are using LLMs to propose cognitive models in learning and decision making data. Presenting this work at RLDM! arxiv.org/abs/2502.00879
arxiv.org
Towards Automation of Cognitive Modeling using Large Language Models
Computational cognitive models, which formalize theories of cognition, enable researchers to quantify cognitive processes and arbitrate between competing theories by fitting models to behavioral data....
0218