Sarthak Mittal
@sarthmit
Graduate Student at @Mila_Quebec and Visiting Researcher @Meta. Prior Research Intern at @Apple, @MorganStanley, @NVIDIAAI and @YorkUniversity
🚀 New Preprint! 🚀 In-Context Parametric Inference: Point or Distribution Estimators? Thrilled to share our work on inferring probabilistic model parameters explicitly conditioned on data, in collab with @Yoshua_Bengio, @FelineAutomaton & @g_lajoie_! 🔗arxiv.org/abs/2502.11617
Explicit latents or implicit marginalization? at #ICML2025 📌 Tue, 11 am 📍East Exhibition Hall A-B (E-1603) Come check out surprising results on whether explicitly incentivizing learning of correct latents improves generalization over implicitly marginalizing it!
🎉Personal update: I'm thrilled to announce that I'm joining Imperial College London @imperialcollege as an Assistant Professor of Computing @ICComputing starting January 2026. My future lab and I will continue to work on building better Generative Models 🤖, the hardest…
⛵️ Excited to share 𝚂𝙰𝙸𝙻𝙾𝚁: a method for *learning to search* with learned world + reward models to plan in the latent space at test-time. Unlike behavior cloning, 𝚂𝙰𝙸𝙻𝙾𝚁 recovers from mistakes without any additional data, DAgger corrections, or ground truth rewards.
Say ahoy to 𝚂𝙰𝙸𝙻𝙾𝚁⛵: a new paradigm of *learning to search* from demonstrations, enabling test-time reasoning about how to recover from mistakes w/o any additional human feedback! 𝚂𝙰𝙸𝙻𝙾𝚁 ⛵ out-performs Diffusion Policies trained via behavioral cloning on 5-10x data!
Physics says it's fine to be lazy New preprint on minimum-excess-work guidance: arxiv.org/abs/2505.13375 Check out the thread below 👇
New preprint alert 🚨 How can you guide diffusion and flow-based generative models when data is scarce but you have domain knowledge? We introduce Minimum Excess Work, a physics-inspired method for efficiently integrating sparse constraints. Thread below 👇arxiv.org/abs/2505.13375
A great collab with former labmates @AntChen_ & Dongyan! Interesting cognitive limitation in LMs: strong disjunctive bias leads to poor performance on conjunctive causal inference tasks. Mirrors adult human biases—possibly a byproduct of training data priors.
Language model (LM) agents are all the rage now—but they may have cognitive biases when inferring causal relationships! We eval LMs on psych task to find: - LMs struggle with certain simple causal relationships - They show biases similar to human adults (but not children) 🧵⬇️
Happy to share that Compositional Risk Minimization has been accepted at #ICML2025 📌Extensive theoretical analysis along with a practical approach for extrapolating classifiers to novel compositions! 📜 arxiv.org/abs/2410.06303
that’s a wrap for #AI4Mat and #FPIWorkshop at #ICLR2025 🎁 thanks for coming, off to sleep for the next 2 weeks 😴 #FPIWorkshop: @tara_aksa @AlexanderTong7 @bose_joey @sarthmit @k_neklyudov @YuanqiD #AI4Mat: @MiretSantiago Rocío M @anoopnm007 @SteMartiniani @MoosaviSMohamad
Great talk by Grant Rotskoff linking sampling with nonequilibrium physics in the #FPIworkshop. Come by for the poster session (Peridot 202-203) for the next hour. #ICLR25
FPI workshop off to a great start with Emtiyaz Khan talking about Adaptive Bayesian Intelligence! Come check it out in Peridot 202-203 #FPIWorkshop #ICLR25
Come check out the workshop and hear about novel works and contributions from an exciting lineup of speakers and panelists!
FPI workshop off to a great start with Emtiyaz Khan talking about Adaptive Bayesian Intelligence! Come check it out in Peridot 202-203 #FPIWorkshop #ICLR25
GDM Mech Interp Update: We study if SAEs help probes generalise OOD (they don't 😢). Based on this + parallel negative results on real-world tasks, we're de-prioritising SAE work. Our guess is that SAEs aren't useless, but also aren't a game-changer More + new research in 🧵
Can only speak for my ICML reviewing batch, but the hack of putting scary, convoluted and wrong math still works.
If you're interested in how to keep challenging neural networks throughout training, check out our latest preprint! #sample_efficiency #scaling_laws
🚀 New Paper Alert! Can we generate informative synthetic data that truly helps a downstream learner? Introducing Deliberate Practice for Synthetic Data (DP)—a dynamic framework that focuses on where the model struggles most to generate useful synthetic training examples. 🔥…