Dhanya Sridhar

philippebrouillard@gmail.com

Biography

Dhanya Sridhar is an assistant professor in the Department of Computer Science and Operations Research (DIRO) at Université de Montréal, a core academic member of Mila – Quebec Artificial Intelligence Institute, and a Canada CIFAR AI Chair.

She was a postdoctoral researcher at Columbia University and received her doctorate from the University of California, Santa Cruz.

In brief, Sridhar’s research focuses on combining causality and machine learning in service of AI systems that are robust to distribution shifts, adapt to new tasks efficiently and discover new knowledge alongside us.

Current Students

Philippe Brouillard

PhD - Université de Montréal

Co-supervisor :

Alexandre Drouin

Independent visiting researcher - University of Maryland College Park

navita.goyal@mila.quebec

sophia.gunluk@mila.quebec

Sophia Gunluk

PhD - Université de Montréal

Shruti Joshi

PhD - Université de Montréal

joshi.shruti@mila.quebec

Tejas Kasetty

Master's Research - Université de Montréal

Principal supervisor :

Guillaume Lajoie

tejaskasetty@gmail.com

cristian-dragos.manta@mila.quebec

Cristian Dragos Manta

PhD - Université de Montréal

Principal supervisor :

Yoshua Bengio

Tom Marty

PhD - Université de Montréal

tom.marty@mila.quebec

mizu.nishikawa-toomey@mila.quebec

Mizu Nishikawa-Toomey

PhD - Université de Montréal

Principal supervisor :

Laurent Charlin

Theo Saulus

PhD - Université de Montréal

Principal supervisor :

Simon Lacoste-Julien

theo.saulus@mila.quebec

Publications

High-resolution dissection of concept acquisition in different families of protein language models

Shawn Whitfield

Tom Marty

Robert M. Vernon

Christopher J. Langmead

Quentin Fournier

Protein language models have been increasingly successful on tasks ranging from fitness prediction to functional design, yet what biological… (see more) knowledge they acquire and where it is encoded within their internal representations remain underexplored. Through a high-resolution layer-by-layer interpretability analysis of 8 models from the ESM2 and AMPLIFY families on 22 concepts from human proteome annotations, we found that these models encode concepts of increasing levels of complexity along their depth: basic physicochemical properties and linear motifs are best captured by early-layer embeddings, secondary structure from subsequent layers, and domain-level semantics from middle layers. Principal component projections of these embeddings showed that they separate biologically meaningful protein groupings, and molecular-biology-inspired interventions demonstrated that pLM embeddings can discriminate phosphomimic-active from inactive mutants. Perhaps surprisingly, we observed that pretraining data and compute had a greater impact on the linear emergence of biological concepts than scaling up parameters. By revealing where biological knowledge is captured in pLMs and which choices shape its emergence, our work offers insights to develop more robust, biologically grounded protein language models.

2026-07-22

bioRxiv (accepted)

Unsupervised Causal Abstractions Discovery

Theo Saulus

Simon Lacoste-Julien

Causal abstractions formalize when a high-level structural causal model (SCM) captures the interventional behavior of a lower-level SCM. Exi… (see more)sting applications of this notion largely follow a hypothesis-testing paradigm: an expert proposes a candidate high-level model and then evaluates if the low-level system implements it. We study the complementary problem of learning a high-level model directly from low-level measurements. Our contributions leverage hypotheses from low-rank causal discovery, and can be summarized as follows: (1) we show that observations generated by a low-rank graph induce latents that form a causal abstraction, (2) we provide identifiability results about these latents, and (3) we propose a practical objective to learn this high-level SCM.

2026-06-16

arXiv (preprint)

A Compression Perspective on Simplicity Bias

Mizu Nishikawa-Toomey

Sarthak Mittal

Guillaume Lajoie

Deep neural networks exhibit a simplicity bias, a well-documented tendency to favor simple functions over complex ones. In this work, we cas… (see more)t new light on this phenomenon through the lens of the Minimum Description Length principle, formalizing supervised learning as a problem of optimal two-part lossless compression. Our theory explains how simplicity bias governs feature selection in neural networks through a fundamental trade-off between model complexity (the cost of describing the hypothesis) and predictive power (the cost of describing the data). Our framework predicts that as the amount of available training data increases, learners transition through qualitatively different features -- from simple spurious shortcuts to complex features -- only when the reduction in data encoding cost justifies the increased model complexity. Consequently, we identify distinct data regimes where increasing data promotes robustness by ruling out trivial shortcuts, and conversely, regimes where limiting data can act as a form of complexity-based regularization, preventing the learning of unreliable complex environmental cues. We validate our theory on a semi-synthetic benchmark showing that the feature selection of neural networks follows the same trajectory of solutions as optimal two-part compressors.

2026-03-25

arXiv (preprint)

Learning a Spatial Partitioning and its Causal Relations from Temporal Data

Yaniv Gurwicz

Peer Nowack

Jakob Runge

David Rolnick

Scientific research often seeks to understand the causal structure underlying high-level variables in a system. For example, climate scienti… (see more)sts study how phenomena, such as El Niño, affect other climate processes at remote locations across the globe. However, scientists typically collect low-level measurements, such as geographically distributed temperature readings. From these, one needs to learn both a mapping to causally-relevant latent variables, such as a high-level representation of the El Niño phenomenon and other processes, as well as the causal model over them. The challenge is that this task, called causal representation learning, is highly underdetermined from observational data alone, requiring other constraints during learning to resolve the indeterminacies. In this work, we consider the task of partitioning observed variables into disentangled factors, such as extracting regions from geographically gridded measurement data in climate research or capturing brain regions from neural activity data. We demonstrate the identifiability of the resulting model and propose a differentiable method, Causal Discovery with Single-parent Decoding (CDSD), that simultaneously learns, from temporal data, the underlying latents and a causal graph over them. We assess the validity of our theoretical results using simulated data and showcase the practical validity of our method in an application to real-world data from the climate science field.

2026-03-09

Conference on Causal Learning and Reasoning (poster)

Who Guards the Guardians? The Challenges of Evaluating Identifiability of Learned Representations

Wieland Brendel

Identifiability in representation learning is commonly evaluated using standard metrics (e.g., MCC, DCI, R^2) on synthetic benchmarks with k… (see more)nown ground-truth factors. These metrics are assumed to reflect recovery up to the equivalence class guaranteed by identifiability theory. We show that this assumption holds only under specific structural conditions: each metric implicitly encodes assumptions about both the data-generating process (DGP) and the encoder. When these assumptions are violated, metrics become misspecified and can produce systematic false positives and false negatives. Such failures occur both within classical identifiability regimes and in post-hoc settings where identifiability is most needed. We introduce a taxonomy separating DGP assumptions from encoder geometry, use it to characterise the validity domains of existing metrics, and release an evaluation suite for reproducible stress testing and comparison.

2026-02-26

arXiv (unknown)

Causality is Key for Interpretability Claims to Generalise

Shruti Joshi

Aaron Mueller

David Klindt

Wieland Brendel

Patrik Reizinger

Interpretability research on large language models (LLMs) has yielded important insights into model behaviour, yet recurring pitfalls persis… (see more)t: findings that do not generalise, and causal interpretations that outrun the evidence. Our position is that causal inference specifies what constitutes a valid mapping from model activations to invariant high-level structures, the data or assumptions needed to achieve it, and the inferences it can support. Specifically, Pearl's causal hierarchy clarifies what an interpretability study can justify. Observations establish associations between model behaviour and internal components. Interventions (e.g., ablations or activation patching) support claims how these edits affect a behavioural metric (e.g., average change in token probabilities) over a set of prompts. However, counterfactual claims -- i.e., asking what the model output would have been for the same prompt under an unobserved intervention -- remain largely unverifiable without controlled supervision. We show how causal representation learning (CRL) operationalises this hierarchy, specifying which variables are recoverable from activations and under what assumptions. Together, these motivate a diagnostic framework that helps practitioners select methods and evaluations matching claims to evidence such that findings generalise.

2026-02-17

arXiv (preprint)

The Role of Causal Features in Strategic Classification for Robustness and Alignment

Sophia Gunluk

Antonio Gois

Nir Rosenfeld

Nidhi Hegde

Simon Lacoste-Julien

In strategic classification, an institution (e.g., a bank) anticipates adaptation from users who change their features to increase utility i… (see more)n a classification task (e.g., loan repayment). Since a key challenge is the distribution shift induced by users, we turn to causal models, which have been shown to bound the worst-case out-of-distribution (OOD) risk, and establish several new results that link causality and strategic classification. First, we show that causal classification leads to optimal classification error after any sufficiently large adaptation, when the noise is bounded in a certain way. Second, when these assumptions do not hold, we show OOD cross-entropy risk of optimal classifiers decomposes into an OOD bias term and a term arising from not using all observable features, allowing us to determine when causal classifiers have an advantage. Finally, we show that causal classifiers can align long-term incentives between institutions and users, contrasting with previous work that highlights social costs of such approaches. We validate our theory empirically on synthetic data, finding that our results predict behavior in practice.

2026-02-02

Artificial Intelligence and Statistics (poster)

Position: Causality is Key for Interpretability Claims to Generalise

Shruti Joshi

Aaron Mueller

David Klindt

Wieland Brendel

Patrik Reizinger

Interpretability research on large language models (LLMs) has produced methods that align model components to high-level concepts, yet their… (see more) use has been accompanied by recurring failures: findings that do not generalise, and causal language that outruns the evidence. Our position is that Pearl’s causal hierarchy formally defines what constitutes a good alignment, what data or assumptions it requires, and what inferences it supports. Specifically, observations of model behaviour support only associational claims; interventions enable cause-effect claims, but not necessarily predictions of model behaviour; counterfactuals, or predictions of behaviour on unseen examples, are often unverifiable in current studies. We show how interpretability research can benefit from causal representation learning (CRL), which provides tools for provably extracting semantic variables and their relationships from activations, and outline practical requirements for generalisable insights: robustness to distribution shifts, sensitivity to assumptions, and compositionality of interventions. Our diagnostic framework helps practitioners select appropriate methods and mitigate failures to ensure that claims match evidence and findings generalise.

2025-12-31

International Conference on Machine Learning (Accept (regular))

Causal Differentiating Concepts: Interpreting LM Behavior via Causal Representation Learning

Navita Goyal

Hal Daumé III

Alexandre Drouin

2025-09-17

NeurIPS.cc/2025/Conference (spotlight)

Leveraging Structure Between Environments: Phylogenetic Regularization Incentivizes Disentangled Representations

Jason Hartford

Recently, learning invariant predictors across varying environments has been shown to improve the generalization of supervised learning meth… (see more)ods. This line of investigation holds great potential for application to biological problem settings, where data is often naturally heterogeneous. Biological samples often originate from different distributions, or environments. However, in biological contexts, the standard "invariant prediction" setting may not completely fit: the optimal predictor may in fact vary across biological environments. There also exists strong domain knowledge about the relationships between environments, such as the evolutionary history of a set of species, or the differentiation process of cell types. Most work on generic invariant predictors have not assumed the existence of structured relationships between environments. However, this prior knowledge about environments themselves has already been shown to improve prediction through a particular form of regularization applied when learning a set of predictors. In this work, we empirically evaluate whether a regularization strategy that exploits environment-based prior information can be used to learn representations that better disentangle causal factors that generate observed data. We find evidence that these methods do in fact improve the disentanglement of latent embeddings. We also show a setting where these methods can leverage phylogenetic information to estimate the number of latent causal features.

2025-07-25

Transactions on Machine Learning Research (accepted)

In-context learning and Occam's razor

A central goal of machine learning is generalization. While the No Free Lunch Theorem states that we cannot obtain theoretical guarantees fo… (see more)r generalization without further assumptions, in practice we observe that simple models which explain the training data generalize best: a principle called Occam's razor. Despite the need for simple models, most current approaches in machine learning only minimize the training error, and at best indirectly promote simplicity through regularization or architecture design. Here, we draw a connection between Occam's razor and in-context learning: an emergent ability of certain sequence models like Transformers to learn at inference time from past observations in a sequence. In particular, we show that the next-token prediction loss used to train in-context learners is directly equivalent to a data compression technique called prequential coding, and that minimizing this loss amounts to jointly minimizing both the training error and the complexity of the model that was implicitly learned from context. Our theory and the empirical experiments we use to support it not only provide a normative account of in-context learning, but also elucidate the shortcomings of current in-context learning methods, suggesting ways in which they can be improved. We make our code available at https://github.com/3rdCore/PrequentialCode.

2025-07-14

International Conference on Machine Learning (Accept (poster))

proceedings.mlr.press

Next-Token Prediction Should be Ambiguity-Sensitive: A Meta-Learning Perspective

The rapid adaptation ability of auto-regressive foundation models is often attributed to the diversity of their pre-training data. This is b… (see more)ecause, from a Bayesian standpoint, minimizing prediction error in such settings requires integrating over all plausible latent hypotheses consistent with observations. While this behavior is desirable in principle, it often proves too ambitious in practice: under high ambiguity, the number of plausible latent alternatives makes Bayes-optimal prediction computationally intractable. Cognitive science has long recognized this limitation, suggesting that under such conditions, heuristics or information-seeking strategies are preferable to exhaustive inference. Translating this insight to next-token prediction, we hypothesize that low- and high-ambiguity predictions pose different computational demands, making ambiguity-agnostic next-token prediction a detrimental inductive bias. To test this, we introduce MetaHMM, a synthetic sequence meta-learning benchmark with rich compositional structure and a tractable Bayesian oracle. We show that Transformers indeed struggle with high-ambiguity predictions across model sizes. Motivated by cognitive theories, we propose a method to convert pre-trained models into Monte Carlo predictors that decouple task inference from token prediction. Preliminary results show substantial gains in ambiguous contexts through improved capacity allocation and test-time scalable inference, though challenges remain.

2025-06-18

ArXiv (preprint)