Is this you? Claim this profile to correct it, add a bio and choose the work people see first.
Claim this profileAcademic lineage
Possible advisorsa guess from early papers, not confirmed
- Sergey LevineSuggested from co-authorship
- Yoshua BengioSuggested from co-authorship
Is this you? Claim this profile to confirm or dismiss it.
Works125 from public data
- Toward Causal Representation Learning1,684
Fundamental concepts of causal inference are reviewed and related to crucial open problems of machine learning, including transfer and generalization, thereby assaying how causality can contribute to modern machine learning research.
- 829
- Inductive biases for deep learning of higher-level cognition491
This work considers a larger list of inductive biases that humans and animals exploit, focusing on those which concern mostly higher-level and sequential conscious processing, to help build AI systems benefiting from humans’ abilities in terms of flexible out-of-distribution and systematic generalization.
- Aligning Text-to-Image Diffusion Models with Reward Backpropagation322
AlignProp is proposed, a method that aligns diffusion models to downstream reward functions using end-to-end backpropagation of the reward gradient through the denoising process, and achieves higher rewards in fewer training steps than alternatives, while being conceptually simpler, making it a straightforward choice for optimizing diffusion models for differentiable reward functions of interest.
- 320
- Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning265
This work leverages Monte Carlo Tree Search (MCTS) to iteratively collect preference data, utilizing its look-ahead ability to break down instance-level rewards into more granular step-level signals, to enhance consistency in intermediate steps.
- 215
- InfoBot: Transfer and Exploration via the Information Bottleneck191
This work proposes to learn about decision states from prior experience by training a goal-conditioned policy with an information bottleneck, and finds that this simple mechanism effectively identifies decision states, even in partially observed settings.
- 167
- SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement160
SWE-Search is proposed, a multi-agent framework that integrates Monte Carlo Tree Search (MCTS) with a self-improvement mechanism to enhance software agents' performance on repository-level software tasks and reveals how performance scales with increased inference-time compute through deeper search, providing a pathway to improve software agents without requiring larger models or additional training data.
- A Theory for Emergence of Complex Skills in Language Models128
A statistical framework that relates cross-entropy loss of LLMs to competence on the basic skills that underlie language tasks and Mathematical analysis showing that the Scaling Laws imply a strong form of inductive bias that allows the pre-trained model to learn very efficiently.
- 116
- Learning to Induce Causal Structure86
This work designs a neural network architecture that learns the mapping from both observational and interventional data to graph structures via supervised training on synthetic graphs and achieves state-of-the-art performance on naturalistic graphs for low sample complexity.
- Skill-Mix: a Flexible and Expandable Family of Evaluations for AI models79
This work develops a methodology for designing and administering a new evaluation to measure ability to combine skills using GPT-4 as well as the open LLaMA-2 70B model, and sketches how the methodology can lead to a Skill-Mix based eco-system of open evaluations for AI capabilities of future models.
- 70
- 70
- 59
- AI-Assisted Generation of Difficult Math Questions48
A design framework that combines the strengths of LLMs with a human-in-the-loop approach to generate a diverse array of challenging math questions and a striking relationship observed between models' performance on the new dataset, suggesting that successfully solving the question in MATH$^2$ requires a nontrivial combination of two distinct math skills.
- 47
- Accelerating Greedy Coordinate Gradient and General Prompt Optimization via Probe Sampling46
This work studies a new algorithm called probe sampling, a mechanism that dynamically determines how similar a smaller draft model's predictions are to the target model's predictions for prompt candidates, which is able to accelerate other prompt optimization techniques and adversarial methods.
- DIBS: Diversity Inducing Information Bottleneck in Model Ensembles45
This paper explicitly optimize a diversity inducing adversarial loss for learning the stochastic latent variables and thereby obtain diversity in the output predictions necessary for modeling multi-modal data.
- 41
- Reasoning Robustness of LLMs to Adversarial Typographical Errors40
An Adversarial Typo Attack algorithm is designed that iteratively samples typos for words that are important to the query and selects the edit that is most likely to succeed in attacking, which shows that LLMs are sensitive to minimal adversarial typographical changes.
- 40
- 36
- Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors34
Results indicate that turning slow derivations into fast procedural hints enables LLMs to remember how to reason, not just what to conclude.
- Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning33
This work introduces Instruct-SkillMix, an automated approach for creating diverse, high quality SFT data for instruction-following, and suggests plausible reasons for why creating open instruction-tuning datasets via naive crowd-sourcing has proved difficult.
- 33
- 31
- 29
- GFlowOut: Dropout with Generative Flow Networks28
GFlowOut leverages the recently proposed probabilistic framework of Generative Flow Networks to learn the posterior distribution over dropout masks and empirically demonstrates that GFlowOut results in predictive distributions that generalize better to out-of-distribution data, and provide uncertainty estimates which lead to better performance in downstream tasks.
- 28
- Can AI Be as Creative as Humans?26
It is proved in theory that AI can be as creative as humans under the condition that it can properly fit the data generated by human creators, thereby reducing the debate on AI's creativity to the question of its ability to fit a sufficient amount of data.
- Discrete Key-Value Bottleneck26
The ability of the discrete key-value bottleneck to minimize the effect of learning under distribution shifts is theoretically investigated and it is shown that it reduces the complexity of the hypothesis class.
- Zero-Shot Object-Centric Representation Learning25
This work studies current object-centric methods through the lens of zero-shot generalization by introducing a benchmark comprising eight different synthetic and real-world datasets, and introduces a novel fine-tuning strategy to adapt pre-trained vision encoders for the task of object discovery.
- 24
- 23
- 20
- Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision19
It is shown that inference compute itself can serve as supervision and by generating parallel rollouts and converting them into reference estimates, models can learn without human labels-critically, even in non-verifiable domains like healthcare guidance where no programmatic checker exists.
- Is Independence all you need? On the Generalization of Representations Learned from Correlated Data19
This work bridges the gap to real-world scenarios by analyzing the behavior of most prominent methods and disentanglement scores on correlated data in a large scale empirical study (including 3900 models).
- 19
- Rethinking Thinking Tokens: LLMs as Improvement Operators18
This work trains an 8B thinking model with Reinforcement Learning to make it consistent with PDR as the inference method, and identifies an interesting inference family Parallel-Distill-Refine (PDR), which performs an interesting inference family Parallel-Distill-Refine, which provides performance superior to long CoT.
- 16
- 16
- 16
- Representation Learning in Deep RL via Discrete Information Bottleneck15
This work introduces a simple, yet effective, bottleneck that can be integrated with any existing self-supervised objective for RL, and finds that compressed representations with RepDIB can lead to strong performance improvements, as the learned bottlenecks help predict only the relevant state while ignoring irrelevant information.
- 15
- 15
- 15
- 15
- Scaling Test-Time Compute for Agentic Coding14
This work proposes a test-time scaling framework for agentic coding based on compact representations of rollout trajectories that converts each rollout into a structured summary that preserves its salient hypotheses, progress, and failure modes while discarding low-signal trace details.
- Generalizing from SIMPLE to HARD Visual Reasoning: Can We Mitigate Modality Imbalance in VLMs?14
A synthetic framework for assessing the ability of VLMs to perform algorithmic visual reasoning, comprising three tasks: Table Readout, Grid Navigation, and Visual Analogy, and shows that explicit image-to-text conversion is important in promoting S2H generalization on images, by transferring reasoning from text.
- 14
- Unlearning via Sparse Representations14
A nearly compute-free zero-shot unlearning technique based on a discrete representational bottleneck that efficiently unlearns the forget set and incurs negligible damage to the model's performance on the rest of the data set is proposed.
- 14
- Bayesian-LoRA: LoRA based Parameter Efficient Fine-Tuning using Optimal Quantization levels and Rank Values trough Differentiable Bayesian Gates13
Bayesian-LoRA is proposed which approaches low-rank adaptation and quantization from a Bayesian perspective by employing a prior distribution on both quantization levels and rank values, and is able to fine-tune a pre-trained model on a specific downstream task, finding the optimal rank values and quantization levels for every low-rank matrix.
- Data for Mathematical Copilots: Better Ways of Presenting Proofs for Machine Learning13
Boosting the capabilities of large language models, or any forthcoming advancements in AI-based mathematical assistants (copilots or ``thought partners''), necessitates a course correction both in the design of mathematical datasets and the evaluation criteria of the models'mathematical ability.
- Discrete Factorial Representations as an Abstraction for Goal Conditioned Reinforcement Learning13
It is shown that applying a discretizing bottleneck can improve performance in goal-conditioned RL setups, and a theorem is proved lower-bounding the expected return on out-of-distribution goals, while still allowing for specifying goals with expressive combinatorial structure.
- 12
- Stateful active facilitator: Coordination and Environmental Heterogeneity in Cooperative Multi-Agent Reinforcement Learning12
A Centralized Training Decentralized Execution learning approach called Stateful Active Facilitator (SAF) is proposed that enables agents to work efficiently in high-coordination and high-heterogeneity environments through a differentiable and shared knowledge source used during training and dynamic selection from a shared pool of policies.
- On the Generalization and Adaption Performance of Causal Models11
This work systematically study the generalization and adaption performance of such modular neural causal models by comparing it to monolithic models and structured models where the set of predictors is not constrained to causal parents.
- 11
- Leveraging Communication Topologies Between Learning Agents in Deep Reinforcement Learning11
This work draws upon results from the networked optimization literatures suggesting that arranging learning agents in communication networks other than fully connected topologies (the implicit way agents are commonly arranged in) can improve learning.
- HypoSpace: A Diagnostic Benchmark for Set-Valued Hypothesis Generation under Underdetermination and Sublinear Coverage Bounds10
HypoSpace is introduced, a diagnostic suite that treats LLMs as samplers of finite hypothesis sets and measures three complementary indicators: Validity (precision of proposals consistent with observations), Uniqueness (non-redundancy among proposals), and Recovery (coverage of the enumerated admissible set).
- Cycle Consistency Driven Object Discovery10
This work introduces a method that explicitly optimizes the constraint that each object in a scene should be associated with a distinct slot, and formalizes this constraint by introducing consistency objectives which are cyclic in nature.
- 10
- 10
- Learning Beyond Pattern Matching? Assaying Mathematical Understanding in LLMs9
This paper assesses the domain knowledge of LLMs through its understanding of different mathematical skills required to solve problems, and finds evidence of domain understanding during in-context learning.
- Spotlight Attention: Robust Object-Centric Learning With a Spatial Locality Prior9
This work incorporates a spatial-locality prior into state-of-the-art object-centric vision models and obtains significant improvements in segmenting objects in both synthetic and real-world datasets.
- 9
- Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models8
This work presents an inference pipeline that attains best-in-class performance on IMO-style math problems at an average inference cost orders of magnitude below competing methods while using only general-purpose off-the-shelf models.
- On the Impossibility of Retrain Equivalence in Machine Unlearning8
These results consistently show that Retrain Equivalence is an ill-posed target for local unlearning algorithms, so long as the target models are trained in stages, so long as the target models are trained in stages.
- Unnatural Languages Are Not Bugs but Features for LLMs8
This work demonstrates that unnatural languages - strings that appear incomprehensible to humans but maintain semantic meanings for LLMs - contain latent features usable by models, and demonstrates that models fine-tuned on unnatural versions of instruction datasets perform on-par with those trained on natural language.
- Masked Generative Priors Improve World Models Sequence Modelling Capabilities8
This work replaces the traditional MLP prior with a Masked Generative Prior (e.g., MaskGIT Prior) and introduces GIT-STORM, a Transformer-based World Models to continuous action environments for the first time, addressing a significant gap in prior research.
- 8
- Learning the Arrow of Time for Problems in Reinforcement Learning8
This work illustrates how a learned arrow of time can capture salient information about the environment, which in turn can be used to measure reachability, detect side-effects and to obtain an intrinsic reward signal.
- Top-K Training of GANs: Improving Generators by Making Critics Less Critical8
A simple modification to the Generative Adversarial Network (GAN) training algorithm is introduced that materially improves results with no increase in computational cost: when updating the generator parameters, it is shown that this `top-k update' procedure is a generally applicable improvement.
- Communication Topologies Between Learning Agents in Deep Reinforcement Learning.8
Results from the networked optimization and collective intelligence literatures suggest that arranging learning agents in less than fully connected topologies (the implicit way agents are commonly arranged in) can improve learning.
- 8
- FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics7
FML-Bench is proposed, a benchmark of 18 fundamental ML research tasks across 10 domains that separates agent strategy from execution infrastructure and defines 12 process-level behavioral metrics and reveals that early convergence and directionally focused exploration are significantly associated with final performance, while solution diversity and compute cost are not.
- 7
- 6
- DiscoGen: Learning to Discover Gene Regulatory Networks6
DiscoGen is introduced, a neural network-based GRN discovery method that can denoise gene expression measurements and handle interventional data and it is demonstrated that the model outperforms SOTA neuralNetwork-based causal discovery methods.
- 6
- Object-Centric Temporal Consistency via Conditional Autoregressive Inductive Biases5
Conditional Autoregressive Slot Attention (CA-SA), a framework that enhances the temporal consistency of extracted object-centric representations in video-centric vision tasks, and outperforms the considered baselines on downstream tasks, such as video prediction and visual question-answering tasks.
- 5
- 5
- 5
- 5
- Physical Reasoning and Object Planning for Household Embodied Agents4
The CommonSense Object Affordance Task (COAT), a novel framework designed to analyze reasoning capabilities in commonsense scenarios, is introduced, which not only advances the understanding of physical commonsense reasoning in language models but also paves the way for future improvements in household agent intelligence.
- 3
- Narrowing the Focus: Learned Optimizers for Pretrained Models3
This work proposes a novel optimizer technique that learns a layer-specific linear combination of update directions provided by a set of base optimizers, effectively adapting its strategy to the specific model and dataset.
- Discrete Compositional Representations as an Abstraction for Goal Conditioned Reinforcement Learning3
- 3
- 3
- 3
- 3
- TC-VAE: Uncovering Out-of-Distribution Data Generative Factors2
The proposed TC-VAE, a generative model optimized using a lower bound of the joint total correlation between the learned latent representations and the input data, is shown to be able to uncover OOD generative factors on different datasets and outperforms on average the related baselines in terms of downstream disentanglement metrics.
- 2
- 2
- 2
- 2
- 2
- 1
- Test-time Adaptation with Slot-Centric Models1
It is shown the proposed GFS-Nets suffices to break the reconstruction-segmentation trade-off, and slow inference greatly improves segmentation in out-of-distribution scenes.
- 1
- 1
- 1
- 1
- 1
- Experience Graphs: The Data Foundation for Self-Improving Agents–
Trellis turns inference-time search from disposable computation into a durable institutional asset: logs made databases reliable; experience graphs may make agents cumulative.
- –
- –
- –
- –
- –
- –
- –
- –
- Leveraging the Third Dimension in Contrastive Learning–
Two distinct approaches to incorporating depth signals into the self-Supervised Learning framework improve the robustness and generalization of the baseline SSL methods, though the first approach (with depth-channel concatenation) is superior.
- Coordinating Policies Among Multiple Agents via an Intelligent Communication Channel–
This paper proposes an alternative approach whereby agents communicate through an intelligent facilitator that learns to sift through and interpret signals provided by all agents to improve the agents’ collective performance.
- –
- Maximum Entropy Models for Fast Adaptation.–
This paper proposes a regularization scheme using a max-entropy prior on the learned features of a neural network; such that the extracted features make minimal assumptions about the training data.
- –
- –
Publication data from OpenAlex; citation counts are the higher of OpenAlex and Semantic Scholar, last synced 2026-10-12. One-sentence summaries under some papers are written by Semantic Scholar’s model. Citation counts may be lower than on Google Scholar, which indexes more sources.
Report an error
Wrong papers, two people merged into one, or a profile that should not be here? Tell us and we will fix or hide it.