https://jkjin.com/Recent content on Welcome to Jikai Jin's websiteHugoen-usSun, 02 Aug 2026 11:21:23 -0400https://jkjin.com/publication/llm_causal_representation_learning/Wed, 08 Jul 2026 00:00:00 +0000https://jkjin.com/publication/llm_causal_representation_learning/We propose a causal representation learning framework to understand the hierarchical structure of language model capabilities, revealing causal relationships between general problem-solving, instruction-following, and mathematical reasoning abilities.https://jkjin.com/gallery/salt_lake_city/Wed, 27 May 2026 00:00:00 +0000https://jkjin.com/gallery/salt_lake_city/https://jkjin.com/gallery/farm_night/Fri, 24 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/farm_night/https://jkjin.com/gallery/rwanda/Fri, 24 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/rwanda/https://jkjin.com/gallery/descanso_garden/Fri, 24 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/descanso_garden/https://jkjin.com/gallery/boston/Fri, 24 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/boston/https://jkjin.com/gallery/datong/Fri, 24 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/datong/https://jkjin.com/gallery/new_york/Fri, 24 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/new_york/https://jkjin.com/gallery/stanford_spring/Thu, 23 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/stanford_spring/https://jkjin.com/gallery/seattle/Mon, 20 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/seattle/https://jkjin.com/gallery/oceanside/Wed, 15 Apr 2026 00:00:00 +0000https://jkjin.com/gallery/oceanside/https://jkjin.com/publication/prescriptive-scaling/Tue, 17 Feb 2026 00:00:00 +0000https://jkjin.com/publication/prescriptive-scaling/Prescriptive scaling laws: estimate stable capability boundaries vs. pretraining compute, track temporal shifts, and release the Proteus 2k evaluation dataset.https://jkjin.com/post/prescriptive-scaling/Sat, 07 Feb 2026 00:00:00 +0000https://jkjin.com/post/prescriptive-scaling/Given a pre-training compute budget, we can predict the best downstream performance strong post-training can reliably attain, and detect when that boundary shifts.https://jkjin.com/publication/latent-bandit/Wed, 04 Feb 2026 00:00:00 +0000https://jkjin.com/publication/latent-bandit/We propose adaptive, state-model-free bandit algorithms for latent-state (confounded, non-stationary) environments.https://jkjin.com/publication/abstention/Wed, 28 Jan 2026 00:00:00 +0000https://jkjin.com/publication/abstention/We study policy learning with the ability to abstain, and demonstrate its usefulness in ensuring safety.https://jkjin.com/publication/general-functional/Fri, 05 Dec 2025 00:00:00 +0000https://jkjin.com/publication/general-functional/Sharp minimax theory for structure-agnostic estimation of general linear functionals, including optimality of first-order debiasing and doubly robust estimators.https://jkjin.com/publication/hard_to_be_normal/Thu, 03 Jul 2025 00:00:00 +0000https://jkjin.com/publication/hard_to_be_normal/We consider structure-agnostic causal inference that estimates treatment effect estimation using black-box ML estimates of nuisance functions, and show that the celebrated DML is optimal when the treatment noise is Gaussian. When the noise is non-Gaussian, we propose ACE, a novel class of higher-order structure-agnostic estimators.https://jkjin.com/publication/llm_inequality/Mon, 09 Jun 2025 00:00:00 +0000https://jkjin.com/publication/llm_inequality/We build a novel benchmark for evaluating how well can state-of-the-art LLMs solve inequality proving problems. Our evaluation studies provide insights on their advanced math reasoning capabilities and highlight their common reasoning flaws.https://jkjin.com/publication/optimality-of-debiasing-for-ate-estimation/Fri, 02 May 2025 00:00:00 +0000https://jkjin.com/publication/optimality-of-debiasing-for-ate-estimation/We show that first-order debiasing of black-box ML estimators is optimal for estimating average treatment effect.https://jkjin.com/publication/grokking/Sat, 30 Nov 2024 00:00:00 +0000https://jkjin.com/publication/grokking/We investigate the “grokking” phenomenon in deep learning on some simple setups, and show that it is caused by a dichotomy of the implicit biases between the early phase and late phase during training.https://jkjin.com/publication/causal-representation-learning/Wed, 25 Sep 2024 00:00:00 +0000https://jkjin.com/publication/causal-representation-learning/We study the best-achievable identification guarantees and provable identification algorithms for causal representation learning when hard interventions are not available.https://jkjin.com/publication/operlearning/Sat, 30 Sep 2023 00:00:00 +0000https://jkjin.com/publication/operlearning/We consider the problem of learning a linear operator between Sobolev RKHSs from noisy data. Different from its finite-dimensional counterpart where regularized least squares is optimal, we prove that estimators with a certain multilevel structure is necessary (and sufficient) to achieve optimality.https://jkjin.com/publication/matrixsensing/Fri, 27 Jan 2023 00:00:00 +0000https://jkjin.com/publication/matrixsensing/We prove that GD applied to the matrix sensing problem has intriguing properties – with small initialization and early stopping, it follows an incremental/greedy low-rank learning procedure. This form of simplicity bias allows GD to recover the ground-truth, despite over-parameterization and non-convexity.https://jkjin.com/publication/robustness/Fri, 27 May 2022 00:00:00 +0000https://jkjin.com/publication/robustness/We provide theoretical evidence that the hardness of robust generalization may stem from the expressive power of deep neural networks. Even when standard generalization is easy, robust generalization provably requires the size of DNNs to be exponentially large.https://jkjin.com/publication/riemann_ahpe/Thu, 10 Feb 2022 00:00:00 +0000https://jkjin.com/publication/riemann_ahpe/We provide an improved analysis of the convergence rates of clipping algorithms, theoretically justifying their superior performance in deep learning.https://jkjin.com/publication/dro/Tue, 05 Oct 2021 00:00:00 +0000https://jkjin.com/publication/dro/We proposed the first non-asymptotic analysis of algorithms for DRO with non-convex losses. Our algorithm incorporates momentum and adaptive step size, and has superior empirical performance.https://jkjin.com/publication/example/Mon, 05 Oct 2020 00:00:00 +0000https://jkjin.com/publication/example/We provide an improved analysis of the convergence rates of clipping algorithms, theoretically justifying their superior performance in deep learning.https://jkjin.com/awards/Fri, 01 Dec 2017 00:00:00 +0000https://jkjin.com/awards/List of awards and honors.https://jkjin.com/news/Fri, 01 Dec 2017 00:00:00 +0000https://jkjin.com/news/List of news.