I’ll be at ICML in Vienna next week! Looking forward to presenting our work on compositional generalization in Transformers. 📑Arxiv: .💻Code: 🧵A thread summarizing our results.

RT @randall_balestr: Learning by input-space reconstruction is often inefficient and hard to get right (compared to joint-embedding). While….

Join millions who have switched to Grok.

RT @EkdeepL: New paper–accepted as *spotlight* at #ICLR2025! 🧵👇. We show a competition dynamic between several algorithms splits a toy mode….

RT @corefpark: I will be presenting our work on:.- Reproducing many in-context learning phenomena.- Identifying a phase diagram of ICL.- Ex….

RT @EkdeepL: Paper alert—accepted as a NeurIPS *Spotlight*!🧵👇. We build on our past work relating emergence to task compositionality and an….

Drop by poster #700 between 11:30 and 1:00!.

RT @bemoniri: This paper is accepted to #ICML2024!.

RT @Hidenori8Tanaka: Q: Can Transformers generalize by composing functions? If so, how?. A: Yes, they achieve combinatorial generalization!….

And finally, a massive shoutout to my amazing collaborators @EkdeepL @KhonaMikail Robert dick and @Hidenori8Tanaka. This was work done during my internship at @NttResearch Harvard and am super grateful to Hidenori for hosting me last summer.

The results hint at why scratchpad or chain-of-thought are really powerful ideas for LLMs, and suggest that compositional generalization is a useful lens to understand the success of these methods.

Check out our paper 📰 for more results on: (1) how the choice of functions changes ability to compostionally generalize; (2) failure of LSTMs on this task (3) training dynamics of compositional generalization.

Our experiments suggest a particular mechanistic hypothesis for this task: the attention layers select the function to apply and the functions are executed in the MLP layers. We observe this consistently across Transformers of different sizes!!

We also sprinkle in spurious correlations in the training data (in-order compositions) and find that it systematically results in a failure to generalize to out-of-order compositions.

In contrast, Transformers that directly generate the final output of the function composition do not generalize. Generalizing to unseen compositions is an OOD task but the step-by-step format format breaks this task down into multiple sub-tasks which are "in-distribution".

Transformers that generate intermediate steps of the composition, can be trained on as few as 100 function compositions but surprisingly generalize to 4 million unseen compositions — a combinatorial explosion!!! 🤯.

We train Autoregressive Transformers on a synthetic data to study how details of the data generating process affect compositional generalization. The most striking differences occur between Transformers that generate and omit intermediate steps of the composition.

Compositional structure of language is complex. We instead consider a simple synthetic setup where data is generated by a linear chain of compositions, i.e., given task tokens F1, F2, F3 and input token X, the goal is to generate F3(F2(F1(X))).

Natural language has rich compositional structure. This motivates the question: what do Transformers learn when trained on a compositional data generating process? 🤔.

RT @docmilanfar: Such an important lesson-. even the very best, the most successful, may barely win more than half the contested points….

RT @EshwarERA: A tweet thread about our recent paper on (Pareto) optimal learning algorithms for repeated games; i.e. how to learn to play….