Sign in

Abel Jansma

@abelaer.bsky.social
673 followers 350 following 116 posts

Compositionality in natural and artificial intelligence || Fellow @emergenceDIEP, University of Amsterdam

PostsRepliesMedia
Abel Jansma @abelaer.bsky.social · 14/09/2026
link to paper: arxiv.org/abs/2609.11948
arxiv.org
Climate-model factor separation with Shapley values and efficient sampling
Climate models often feature nonlinear responses to changes in model parameters or boundary conditions. Factor separation asks how much of the resulting change should be assigned to altered factors an...
000
Abel Jansma @abelaer.bsky.social · 14/09/2026
I'm not a climate scientist, and only found this paper after discussing attribution methods with AI chatbots. I think it's important to stay modest and talk to real experts, so huge thanks to @climatesamwell.bsky.social NASA GISS director Gavin Schmidt for discussing these ideas!
100
Abel Jansma @abelaer.bsky.social · 14/09/2026
In fact, there are more clever sampling tricks. Here I use a basic trick called reversal-paired sampling, which made climate model attribution *several orders of magnitude* faster!
100
Abel Jansma @abelaer.bsky.social · 14/09/2026
ML researchers have developed efficient sampling methods to approximate Shapley values, but these now also make it possible to efficiently estimate the effect of, say, CO2 and CH4 on climate predictions:
100
Abel Jansma @abelaer.bsky.social · 14/09/2026
The proof of the 2021 conjecture can be directly imported from the game theory literature---a very satisfying case of interdisciplinary convergence. The conjecture is now a theorem, which has immediate practical advantages.
100
Abel Jansma @abelaer.bsky.social · 14/09/2026
I immediately recognised the mathematics: it turns out that Möbius inversions and Shapley values have been *independently discovered* in the climate modelling literature! So cool, I love both these ideas, and previously connected them here: arxiv.org/abs/2510.05786
t.co
https://arxiv.org/abs/2510.05786
100
Abel Jansma @abelaer.bsky.social · 14/09/2026
The conjecture was stated in a collaborative paper between universities in the UK & Canada, and NASA's institute for space studies.
100
Abel Jansma @abelaer.bsky.social · 14/09/2026
🚨 New paper! arxiv.org/abs/2609.11948 In 2021, a group of climate scientists realised that different attribution methods in climate modelling gave the same answer, and conjectured that this holds in general. It turns you can prove this with Game Theory! 🧵
120
Abel Jansma @abelaer.bsky.social · 22/06/2026
This is now accepted to two ICML workshops: Compositional Learning, and Mech Interp. Link here: t.co/l0AJB5Hi9Z
t.co
https://openreview.net/pdf?id=sMUVXkAksS
000
Abel Jansma @abelaer.bsky.social · 22/06/2026
Finally, it turns out that these synergy vectors encode not just idiomatic meaning, but also the abstract notion of "idiomaticity". This means you can use them as steering vectors to let models interpret your input literally, or figuratively!
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
Pythia-1B checkpoints show that it takes a while for the model to learn how to encode the semantic synergy.
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
This works for sentence embedding models, and for next-token predictors, with the latter having the strongest semantics in intermediate layers.
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
In fact, these extracted synergy vectors are often *better* at expressing idiomatic meaning than the *full idiom embedding*!
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
The algebra of the Möbius inversion is crucial here---none of the other residuals work!
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
If a phrase is compositional, this residual should vanish. If meaning appears only at the level of the whole phrase, q should point toward that extra meaning. This works! Idioms like “kick the bucket” have synergy aligned with “to die”, unlike literals like “kick the ball”.
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
Together with Patrick Forré, I recently generalised Möbius transforms to vector-valued functions, so we can now take a derivative of the encoder! For 3 words, for example, you get ∂T(abc)=T(abc)-T(ab)-T(bc)+T(b)
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
How do LLMs represent phrases like “kick the bucket”, where the meaning is not in the words but ‘higher-order’? Idea: exploit the part-whole relationship of language to take a discrete derivative of the embedding function T.
100
Abel Jansma @abelaer.bsky.social · 22/06/2026
New paper alert 🚨 “A Compositional Calculus for Semantic Synergy in Language Model Embeddings” A training-free way to extract synergy from LLM embedding vectors. 🧵
100
Abel Jansma @abelaer.bsky.social · 18/03/2026
Gave a talk at the Softmax Alignment meeting—pretty nice lecture room.
011
Abel Jansma @abelaer.bsky.social · 18/03/2026
I asked somebody in the Berkeley AI safety scene if they meditate. “No,” they said, “I’m not really bliss bottlenecked at the moment”.
020
Abel Jansma @abelaer.bsky.social · 10/03/2026
Every year, our apricot is the first to blossom, which is easy to remember because “apricot” shares a linguistic root with “precocious”.
020
Abel Jansma @abelaer.bsky.social · 04/03/2026
I hope that when I'm 88 I'm also still able to embrace new technologies with such joy and curiosity: www-cs-faculty.stanford.edu/~knuth/paper...
www-cs-faculty.stanford.edu
010
Abel Jansma @abelaer.bsky.social · 23/02/2026
I think people just disagree on what representing something with a graph means. You seem to mean something very specific, along the lines of eq. 4. But if I want to represent the interactions of an up-to n-th order Ising model, a hypergraph can do that, and a pairwise graph can't--you need 2^n dof.
210
Abel Jansma @abelaer.bsky.social · 16/12/2025
ok let's make a list of all ℵ₀ genders then. Now construct a new flag where the color of the first band differs from that of the 1st gender's flag, the color of the second band differs from that of the 2nd gender, etc. This defines a new gender that was not in your list of ℵ₀ genders. □
130
Abel Jansma @abelaer.bsky.social · 02/12/2025
I found this surprising and somewhat worrying: openAI is using an AI model to align their model spec. #NeurIPS2025
010
Abel Jansma @abelaer.bsky.social · 22/10/2025
🔗 Erik’s post: theintrinsicperspective.com/p/i-figured-... 🔗 The paper: arxiv.org/abs/2510.02649
theintrinsicperspective.com
I Figured Out How to Engineer Emergence
What's come of my return to science
010
Abel Jansma @abelaer.bsky.social · 22/10/2025
A few months ago I started discussing causal emergence with Erik Hoel. This led to a really fun collaboration, and a new approach to “engineer emergence”. Erik just published an overview of the ideas, goals, and dreams:
120
Abel Jansma @abelaer.bsky.social · 14/10/2025
I know you like showing pictures of lenses but this seems a little excessive
010
Reposted by Abel Jansma
Rebecca R Helm @rebeccarhelm.bsky.social · 09/10/2025
Many sea stars begin life as young fairy-like creature (called a brachiolaria) that float through the open ocean. Eventually, a small star forms within them (here in yellow). The fairy-like brachiolaria sinks under the star’s weight, and the star pops out! 🎥@the_story_of_a_biologist (on Insta)
7509179
Abel Jansma @abelaer.bsky.social · 08/10/2025
There's so much more in the paper, largely thanks to Patrick who really pushed this to the next level. We're already working on applications: SVs are often used for XAI, but now we can do this for vector-valued functions--the kind implemented by transformers... Stay tuned!
110
Abel Jansma @abelaer.bsky.social · 08/10/2025
To summarise: ⬆️Möbius inversions construct higher-order structure. ⬇️Shapley values project this down again, in the 'right' way. We derive generalisations of both, to directed acyclic multigraphs, and group-valued functions.
120
Abel Jansma @abelaer.bsky.social · 08/10/2025
This shows how intimately related Shapley values and Möbius inversions are: we derive an expression that expresses Shapley values *purely in terms of the incidence algebra*!
120
Abel Jansma @abelaer.bsky.social · 08/10/2025
Doing so required also generalising the Möbius inversion theorem to this setting (prev. only defined for ring-valued functions). We show that it's a natural theorem in the *path algebra* of the graph:
110
Abel Jansma @abelaer.bsky.social · 08/10/2025
But we go further. Classical Shapley values only work for real-valued functions on power sets of players (or lattices). We generalise them even beyond posets to ✅vector/group-valued fns ✅weighted directed acyclic multigraphs , and prove uniqueness!
120
Abel Jansma @abelaer.bsky.social · 08/10/2025
That’s exactly what we do. We reinterpret Shapley values as projection operators: a recursive re-attribution of higher-order synergy to lower-order parts. This turns Shapley values into a general projection framework for hierarchical structure, valid far beyond game theory.
140
Abel Jansma @abelaer.bsky.social · 08/10/2025
Möbius inversions are a way to derive higher-order interactions ion a system's mereology. I wrote a blog post about this here 👉https://abeljansma.nl/2025/01/28/mereoPhysics.html If Shapley values are truly general, we should be able to express them for any Möbius inversion/higher-order structure.
120
Abel Jansma @abelaer.bsky.social · 08/10/2025
But Shapley values (SVs) aren’t just about fairness. They're really a projection operator: the right way to push higher-order structure back down to lower levels. So… can we do this more generally? 🤔 Enter Möbius inversions...
110
Abel Jansma @abelaer.bsky.social · 08/10/2025
If a group of people earn a payoff together, how should it be fairly distributed? Shapley values are weighted sums of sub-coalition "synergies", and provably the fairest possible distribution. It earned Shapley the Nobel Prize. 🧮
110
Abel Jansma @abelaer.bsky.social · 08/10/2025
🚨New paper: arxiv.org/abs/2510.05786 *Shapley values beyond game theory* We show that Shapley values aren’t just about dividing payoffs--they are the right way to project down any higher-order structure. We generalise them, and Möbius inversions, in important ways: 🧵
arxiv.org
Möbius transforms and Shapley values for vector-valued functions on weighted directed acyclic multigraphs
We generalize the concept of Möbius inversion and Shapley values to directed acyclic multigraphs and weighted versions thereof. We further allow value functions (games) and thus their Möbius transform...
1103
Abel Jansma @abelaer.bsky.social · 23/09/2025
This is how Rota originally introduced the incidence algebra. Everyone since has (correctly) required the ring to be commutative. Did people in the '60s just refer to commutative rings as associative rings?
000
Abel Jansma @abelaer.bsky.social · 23/09/2025
Just to balance out the discourse: the #NeurIPS2025 review process for this paper went great. Fair reviews, mostly responsive reviewers, and a thoughtful AC that caught a possible conflict of interest. Definitely improved the paper.
010
Abel Jansma @abelaer.bsky.social · 22/09/2025
The #NeurIPS2025 version is now online: arxiv.org/pdf/2501.11447 It includes a new analysis to show that LLM semantics can be decomposed: the negativity of "horribly bad" is redundantly encoded in the two words, whereas "not bad" has synergistic semantics (i.e. negation):
030
Abel Jansma @abelaer.bsky.social · 19/09/2025
The "partial causality decomposition" was just accepted for a spotlight at #NeurIPS2025! The final version includes a decomposition of LLM semantics---the Arxiv version should be updated soon. Stay tuned!
010
Abel Jansma @abelaer.bsky.social · 12/08/2025
Hi! sorry I'm not on often on this site. They're now online: www.youtube.com/@dutchinstit...
youtube.com
Dutch Institute for Emergent Phenomena
100
Abel Jansma @abelaer.bsky.social · 05/08/2025
While I'm flattered, it's a bit weird that google's AI defers to me when you search for this:
000
Abel Jansma @abelaer.bsky.social · 14/05/2025
Yes!
100
Abel Jansma @abelaer.bsky.social · 06/05/2025
Next week we're organising a workshop on the role of analogies in (artificial) intelligence, with: Melanie Mitchell (@melaniemitchell.bsky.social), Martha Lewis, Jules Hedges (‪@julesh.mathstodon.xyz.ap.brid.gy‬), and Han van der Maas. Register here: www.d-iep.org/workshopanal...
d-iep.org
WORKSHOPANALOGIES | DIEP
251
Abel Jansma @abelaer.bsky.social · 08/04/2025
Don't take my word for it--take Reviewer 2's: "I found the paper extremely interesting and deep" A gentle introduction is available at abeljansma.nl/2025/01/28/m...
abeljansma.nl
Complex Systems and Quantitative Mereology
010
Abel Jansma @abelaer.bsky.social · 08/04/2025
My new approach to higher-order interactions in complex systems is now published in Physical Review Research: journals.aps.org/prresearch/a...
journals.aps.org
Mereological approach to higher-order structure in complex systems: From macro to micro with M\"obius
Relating macroscopic observables to microscopic interactions is a central challenge in the study of complex systems. While current approaches often focus on pairwise interactions, a complete understan...
121
Abel Jansma @abelaer.bsky.social · 07/02/2025
The link doesn’t seem to work for me
100