AI-generated summaries

Today's ML research,
without the noise.

Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.

53 Papers today
8h Update frequency
7 Days of history
CascadeFormer: Depth-Tapered Transformers Motivated by Gradient Fan-in Asymmetry
Huzama Ahmad, Cao Viet Hai Nam, Se-Young Yun
NLP Large Language Models Efficient ML
  • Introduction of Gradient Fan-in Asymmetry (GFA) as a structural explanation for layer redundancy in deep transformers.
  • Development of CascadeFormer, which tapers model width with depth to improve efficiency without sacrificing performance.
  • CascadeFlow Pruning (CFP) leverages training gradients for effective layer pruning, outperforming standard methods.
  • Empirical validation of GFA through correlational and interventional studies on models up to 1.2B parameters.
Read more
Kolmogorov Arnold networks (KAN) for aerodynamic prediction: a comparison with MLPs and GNNs
Miguel Jaraiz, Fermin Gutierrez, Pablo Yeste, Miguel Sánchez-Domínguez, Eusebio Valero, Gonzalo Rubio, Lucas Lacasa
Theory Efficient ML Graph Learning
  • KANs adapt activation functions, offering a new approach to neural network architecture.
  • In aerodynamic prediction tasks, KANs show comparable performance to MLPs but are outperformed by GNNs.
  • KANs have faster training times due to lower complexity but suffer from training instabilities.
  • Hyperparameter optimization is crucial for improving KAN performance.
Read more
RSPC: A Benchmark for Modeling Stress and Psychiatric Conditions in Digitally Mediated Relationships using Psychiatrist Annotations
Parmitha Vangapandu, Sai Ganesh Mokkapati, Sathwik Narkedimilli, MSVPJ Sathvik, Timothy Liu, Simon See, Johannes C. Eichstaedt
NLP Large Language Models
  • Introduction of the RSPC dataset linking psychiatric conditions with relational stressors.
  • Benchmarking of transformer models and LLMs for mental health tasks.
  • Identification of distinct model capabilities based on task requirements.
  • Strong associations found between anxiety disorders and relational uncertainty.
Read more
Physics-guided Convolutional Neural Network for Domain Growth Prediction in Systems with Conserved Kinetics
Vijay Yadav, Madhu Priya, Manish Dev Shrimali, Prabhat K. Jaiswal
Theory Efficient ML
  • Introduction of an attention-based, physics-guided CNN for modeling phase separation dynamics.
  • The model incorporates conservation constraints to ensure physical consistency during long-term predictions.
  • Demonstrates accurate predictions of domain growth and preservation of mixture composition.
  • Validates the model against known growth laws, confirming its effectiveness for critical and off-critical mixtures.
Read more
Blackwell Approachability and Gradient Equilibrium are Equivalent
Brian W. Lee, Nika Haghtalab, Michael I. Jordan, Ryan J. Tibshirani
Theory Optimization
  • GEQ is algorithmically equivalent to Blackwell Approachability, allowing for the use of GEQ oracles in BA problems.
  • The paper identifies necessary and sufficient conditions for achieving GEQ, expanding its theoretical foundation.
  • Efficient reductions between GEQ and other frameworks like regret minimization and calibration are established.
  • The results imply that GEQ algorithms are as powerful as classical regret minimization and calibration algorithms.
Read more
Target-Aware Bandit Allocation for Scalable Surrogate Optimization in Chemical Space
Mohammad Haddadnia, Yuvan Chali, Abhilash Jayaraj, Constance Kraay, Joana Reis, Felix Strieth-Kalthoff, Haribabu Arthanari
Optimization
  • Introduction of BOBA, a bandit-guided surrogate optimization framework.
  • Elimination of full-library inference by adaptive computation allocation.
  • Demonstration of the importance of uncertainty-aware bandit strategies.
  • Establishment of a tunable tradeoff between optimization performance and inference cost.
Read more
Quantization in Federated Learning: Methods, Challenges and Future Directions
Farwa Ikram, Dipanwita Thakur, Antonella Guzzo, Giancarlo Fortino
Federated Learning Efficient ML
  • Introduces a novel taxonomy of quantization methods specific to Federated Learning.
  • Analyzes the interaction between quantization and core FL behaviors such as client drift and convergence stability.
  • Identifies open research gaps and provides design guidelines for deploying quantized FL.
  • Establishes quantization as a critical factor affecting the efficiency and robustness of FL systems.
Read more
Topology-Informed Neural Networks for Flood Detection in Optical and Synthetic Aperture Radar Imagery
Sophia Li, Max Zhao, Raghu G. Raj, Tianyu Chen
Computer Vision Time Series Interpretability
  • Introduces Topological Data Analysis (TDA) for flood detection, enhancing interpretability and robustness.
  • Utilizes the SEN12-FLOOD dataset, which includes both optical and SAR imagery for comprehensive flood monitoring.
  • Achieves a notable accuracy improvement in flood detection by combining topological features with traditional neural network architectures.
  • Demonstrates the effectiveness of transfer learning and a lightweight Gaussian topological embedding in improving model performance.
Read more
EVOM: Agentic Meta-Evolution of Actor-Critic Architectures for Reinforcement Learning
Boyun Zhang, Chao Wang, Kai Wu
Reinforcement Learning Large Language Models Optimization
  • EVOM automates the design of actor-critic architectures, addressing high evaluation costs and open-ended design challenges.
  • The framework utilizes a bi-level optimization approach with an inner loop for weight training and an outer loop for architecture evolution.
  • An LLM-based design agent generates and refines architecture programs, enhancing the search process.
  • Experimental results show EVOM outperforms existing methods, including manual designs and LLM-guided searches.
Read more
fTNN: a tensor neural network for fractional PDEs
Qingkui Ma, Hehu Xie, Xiaobo Yin
Theory Optimization Efficient ML
  • Introduction of fTNN, a tensor neural network for solving fractional PDEs.
  • Development of a deterministic integration framework for the fractional Laplacian.
  • Construction of boundary-singularity-aware trial functions for better accuracy.
  • Design of a spatiotemporally separable neural network for time-dependent PDEs.
Read more
Mesh-RL: Coupled subgrid reinforcement learning
Behnam Gheshlaghi, Bahador Rashidi, Shahin Atakishiyev
Reinforcement Learning
  • Mesh-RL introduces a spatial domain-decomposition framework for reinforcement learning.
  • The framework enforces boundary-consistent TD updates for improved value propagation.
  • Mesh-RL accelerates learning without modifying the reward function or requiring explicit planning.
  • Empirical results show significant improvements in convergence speed and learning stability.
Read more
Discovering Millions of Interpretable Features with Sparse Autoencoders
XinYang He, Wei Wang, Bing Zhao, Xuan Ren, WenBo Li, WeiXu Qiao, Hu Wei, Lin Qu
NLP Large Language Models Interpretability
  • Introduction of Qwen3-Instruct SAE, a comprehensive suite of Sparse Autoencoders for the Qwen3 model family.
  • Layer-wise SAEs are trained at key activation sites to enhance interpretability.
  • Evaluation reveals distinct sparsity-fidelity trade-offs across different layers.
  • Demonstrated utility in a case study for steering model behavior.
Read more
Algorithmic Foundations of Deep Learning: Complexity-Theoretic Rates and a Characterization of Universal Approximation
Anastasis Kratsios, Simone Brugiapaglia, Bum Jun Kim, Gregory Cousins, Haitz Sáez de Ocáriz Borde
Theory
  • Introduces a complexity-theoretic perspective on neural network expressivity.
  • Establishes a circuit-to-neural-network compilation theorem linking computational complexity to neural network architecture.
  • Characterizes universal approximation in feedforward neural networks based on the presence of non-affine nonlinearities.
  • Demonstrates improved approximation rates for complex functions compared to classical approximation theory.
Read more
A General Framework for Learning Algebraic Properties from Cayley Graphs using Graph Neural Networks
Tal Weissblat
Graph Learning
  • Development of a property-independent GNN framework for learning algebraic properties from Cayley graphs.
  • Successful case studies on abelianity, nilpotency, and solvability of finite groups.
  • Expanded dataset includes a broader range of finite groups for comprehensive evaluation.
  • Demonstrates the potential of GNNs to extract significant algebraic information from graph representations.
Read more
Data-Free Reservoir Features for Efficient Long-Horizon Cold-Start Continual Learning
Augustinas Jučas, Yangchen Pan
Computer Vision Efficient ML
  • CIRCLE introduces a data-free frozen-feature design for cold-start exemplar-free class-incremental learning.
  • The model combines fixed random feature extraction with an ensemble of streaming linear discriminant analysis heads.
  • CIRCLE shows superior performance at long task horizons (50-500 task splits) compared to traditional methods.
  • The approach eliminates the need for replay, task-boundary information, and backbone backpropagation.
Read more
Hallucination in World Models is Predictable and Preventable
Nicklas Hansen, Xiaolong Wang
Generative Models Reinforcement Learning Robotics
  • Introduction of MMBench2, a large dataset for visual world modeling with ground-truth actions and rewards.
  • Identification of three distinct hallucination modes in generative world models.
  • Development of three predictors for detecting hallucinations without additional training.
  • Implementation of a coverage-aware training method that enhances rollout fidelity.
Read more
Equivariance and Augmentation for Bayesian Neural Networks
Miaowen Dong, Axel Flinth, Jan E. Gerken
Theory
  • The paper establishes a theoretical framework for understanding how data augmentation induces equivariance in variational Bayesian inference.
  • Three novel symmetrization techniques are introduced to improve the equivariance properties of BNNs.
  • The study shows that starting from an invariant prior allows the variational distribution to maintain its invariance during training.
  • Numerical experiments validate the theoretical results, with orbit expansion outperforming baseline methods.
Read more
Embedding Foundation Model Predictions in Discrete-Choice Models with Structural Guarantees
Yingshuo Wang, Xian Sun, Yanhang Li, Zhichao Fan, Zexin Zhuang
Theory
  • Introduces a two-stage adapter for embedding foundation model predictions in discrete-choice models.
  • Preserves structural guarantees of multinomial logit models while enhancing prediction accuracy.
  • Demonstrates significant accuracy improvements (up to 12.8 percentage points) across various datasets.
  • Maintains cost monotonicity and produces realistic willingness-to-pay estimates.
Read more
Decision-Aligned Evaluation of Uncertainty Quantification
Annika Schneider, Tommy Rochussen, Joshua Stiller, Vincent Fortuin
Theory
  • Introduces decision-alignment as a formal criterion for evaluating UQ metrics.
  • Identifies that many traditional UQ metrics are misaligned with downstream decision utilities.
  • Proposes prior-weighted utility metrics that better capture the value of models for decision-making.
  • Demonstrates through experiments that prior-weighted metrics align more closely with real-world decision utilities.
Read more
Finding Stationary Points by Comparisons
Helin Wang, Chenyi Zhang, Xiwen Tao, Yexin Zhang, Tongyang Li
Optimization Theory
  • Developed an algorithm for finding ϵ-stationary points using a comparison oracle with improved query complexity.
  • Introduced a quantum algorithm that reduces query complexity significantly in the quantum oracle model.
  • Demonstrated the relevance of the findings to practical optimization problems in machine learning.
  • Identified the need for further research on lower bounds in the context of comparison-based optimization.
Read more
Rethinking Training & Inference for Forecasting: Linking Winner-Take-All back to GMMs
Qiyuan Wu, Katie Z Luo, Bharath Hariharan, Wei-Lun Chao, Mark Campbell
Robotics Time Series Theory
  • Identifies a fundamental mismatch in training objectives for trajectory forecasting models.
  • Proposes two effective post-hoc treatments to improve mode probability assignments.
  • Demonstrates that the WTA training objective can lead to over-segmentation of trajectory predictions.
  • Provides a unified perspective on GMMs and K-means clustering in the context of forecasting.
Read more
Error-Conditioned Neural Solvers
Haina Jiang, Liam Wang, Peng-Chen Chen, Min Seop Kwak, Seungryong Kim, Brian Bell, Jeong Joon Park
Optimization Theory Efficient ML
  • ENS utilizes the PDE residual as an input rather than an optimization target, enabling better error correction.
  • The framework achieves significant improvements in prediction accuracy, especially in ill-conditioned scenarios.
  • ENS demonstrates robustness to initialization and generalizes effectively under distribution shifts.
  • Theoretical and empirical analysis reveals the limitations of existing hybrid methods in achieving accurate solutions despite low residuals.
Read more
Structure Before Collapse: Transient semantic geometry in next-token prediction
Yize Zhao, Isabel Papadimitriou, Christos Thrampoulidis
NLP Large Language Models Theory
  • Neural Collapse theory predicts that one-hot supervision erases latent semantic structure, yet language models learn such structure.
  • The authors introduce Representational Similarity Analysis (RSA) to track semantic structure in learned representations.
  • Three synthetic languages are identified where models recover structured semantic geometry despite one-hot supervision.
  • The semantic geometry is transient, emerging early in training before collapsing to a symmetric ETF.
Read more
Batch-Invariant Spectral Intelligence for Robust and Explainable Insect Authentication
Majharulislam Babor, Giacomo Rossi, Annalisa Altavilla, Oliver Schlüter, Marina M.-C. Höhne
Interpretability
  • Introduction of the Batch-Invariant Spectral Network (BISN) for insect authentication.
  • BISN effectively suppresses batch-specific spectral variations before learning species features.
  • Achieved a mean accuracy of 0.93 in classifying insect species across different batches.
  • Utilized explainable AI to confirm model reliance on relevant biochemical absorption regions.
Read more
Designing Reward Signals for Portable Query Generation: A Case Study in Industrial Semantic Job Search
Ping Liu, Qianqi Shen, Jianqiang Shen, Wenqiong Liu, Rajat Arora, Yunxiang Ren, Chunnan Yao, Dan Xu, Baofen Zheng, Wanjun Jiang, Andrii Soviak, Kevin Kao, Jingwei Wu, Wenjing Zhang
Reinforcement Learning NLP Optimization
  • Introduces a novel RLAIF framework for generating portable job search queries.
  • Identifies and addresses the issue of reward-hacking leading to verbatim copying.
  • Demonstrates that robust reward shaping is more impactful than the choice of RL optimizer.
  • Implements a rule-based reward floor to improve query generation quality.
Read more
Just how sure are you? Improving Verbalized Uncertainty Calibration in Medical VQA
Eren Senoglu, Federico Toschi, Nicolo Brunello, Andrea Sassella, Mark James Carman
Multimodal Large Language Models Computer Vision
  • Proposes a novel training-based framework for verbalized confidence calibration in Medical VQA.
  • Introduces a composite loss function that combines multiple calibration and regularization techniques.
  • Demonstrates a 60% reduction in calibration error and a 26% improvement in discrimination across benchmarks.
  • Validates the method on two different model architectures, showing robustness and effectiveness.
Read more
Recovering Governing Equations from Solution Data: Identifiability Bounds for Linear and Nonlinear ODEs
Yang Pan, Helmut Bölcskei
Theory
  • Introduces Hausdorff distance as a metric for comparing ODEs, capturing worst-case separation of solutions.
  • Establishes identifiability bounds for linear and nonlinear ODEs, detailing conditions for unique identification.
  • Provides quantitative analysis of sample complexity, determining the number of observations needed for reliable recovery.
  • Fills a significant gap in the theoretical understanding of learning governing equations from solution data.
Read more
Multipath Adaptive Gated Bottleneck Latent ODE with Raman Data Fusion for Cell Culture Process Forecasting
Johnny Peng, Thanh Tung Khuat, Ellen Otte, Katarzyna Musial, Bogdan Gabrys
Time Series
  • Introduction of a novel adaptive framework for bioprocess forecasting.
  • Combination of GB-Latent ODE and MP-JIT-FT for generating multiple plausible future trajectories.
  • Integration of Raman spectroscopy data to enhance model training and forecasting accuracy.
  • Demonstrated superior performance on real-world bioreactor data compared to traditional methods.
Read more
Cross-Head Attention Uplift Network with Inverse Propensity Score under Unobserved Confounding
Haoran Zhang, Chuanpu Li, Yuxin Fu, Bin Tong, Guan Wang, Bo Zheng, Feng Zhou
Theory
  • Introduction of CHAUN for improved uplift modeling through attention mechanisms.
  • Theoretical proof of ITE identifiability with true propensity scores despite unobserved confounding.
  • RA-IPS method to optimize propensity weights and mitigate selection bias.
  • Empirical validation showing up to 25.6% improvement in QINI scores over state-of-the-art models.
Read more
Epiphany-Aware KV Cache Eviction Without the Attention Matrix
Steven Kolawole, Virginia Smith
Large Language Models Efficient ML NLP
  • EPIKV scores tokens based on internal representation changes rather than attention weights, improving eviction quality.
  • The method allows for a 16× longer feasible context length compared to traditional attention-based scoring methods.
  • EPIKV matches or exceeds the performance of existing attention-based baselines on MATH-500 and AIME-2024 benchmarks.
  • The approach runs up to 2.8× faster than attention-based eviction methods at equal budget.
Read more
Automating Potential-based Reward Shaping with Vision Language Model Guidance
Henrik Müller, Daniel Kudenko
Reinforcement Learning NLP Multimodal
  • Introduction of VLM-PBRS, a framework that automates potential-based reward shaping using VLM feedback.
  • Utilization of smaller, cost-effective VLMs to generate preference labels, reducing computational burden.
  • Empirical validation showing improved sample efficiency and robustness to reward hacking in RL environments.
  • Demonstration of the relationship between VLM label accuracy and learning efficiency.
Read more
Reasoning Quality Emerges Early: Data Curation for Reasoning Models
Hongyi Henry Jin, Wenhan Yang, Meysam Ghaffari, Carlos Morato, Baharan Mirzasoleiman
NLP Large Language Models Efficient ML
  • Introduces a new method for data curation that relies on initial reasoning tokens rather than strong reasoning models.
  • Demonstrates that challenging examples can be identified based on loss metrics from early reasoning tokens.
  • Establishes a correlation between loss patterns and gradient similarity during fine-tuning.
  • Achieves up to 1.7% performance improvement over existing baselines while being 91% more token efficient.
Read more
Explaining Temporal Graph Neural Networks via Feature-induced Information Flow
Ping Xiong, Thomas Schnake, Klaus-Robert Müller, Shinichi Nakajima
Graph Learning Interpretability Time Series
  • Introduction of a novel Event Relevance (ER) method for explaining ETGNNs.
  • Extension of the Normalized Relevance Measure (NRM) framework to handle complex neural architectures.
  • Demonstration of superior qualitative and quantitative performance over existing explanation methods.
  • Focus on capturing the entire information flow, including intermediate event-induced variables.
Read more
SSM Adapters via Hankel Reduced-order Modeling: Injection Site Determines Task Suitability in Long-Context Fine-Tuning
Omanshu Thapliyal
NLP Large Language Models Efficient ML
  • Introduces Hankel Reduced-order Model (HRM) adapters for parameter-efficient fine-tuning.
  • HRM outperforms LoRA variants on LongBench tasks with significant accuracy improvements.
  • Demonstrates consistent performance across diverse configurations in state-tracking and language modeling.
  • Provides a computationally efficient method for integrating temporal memory into frozen transformer models.
Read more
GEOALIGN: Geometric Rollout Curation for Robust LLM Reinforcement Learning
Ting Zhou, Zhenqing Ling, Yiyang Zhao, Ying Shen, Daoyuan Chen
Reinforcement Learning Large Language Models
  • Introduction of GEOALIGN to address directional inconsistency in online RL for LLMs.
  • GEOALIGN operates as a lightweight plug-in for rollout curation, enhancing training stability.
  • The method improves performance in dialogue alignment and mathematical reasoning tasks.
  • GEOALIGN demonstrates resilience against controlled reward corruption.
Read more
AIGP: An LLM-Based Framework for Long-Term Value Alignment in E-Commerce Pricing
Chennan Ma, Yanning Zhang, Siqi Hong, Xiuchong Wang, Fei Xiao, Keping Yang
NLP Large Language Models Reinforcement Learning
  • AIGP integrates LLM-based reasoning with long-term business value alignment for dynamic pricing.
  • The framework employs a Long-Term Value Estimator (LTVE) trained via offline reinforcement learning.
  • AIGP achieved a +13.21% increase in GMV and +7.59% in ROI over traditional pricing models.
  • The model provides interpretable pricing rationales, enhancing decision transparency.
Read more
When Does Quality-Aware Multimodal Fusion Matter? A Leakage-Safe Diagnostic for Decision-Level Dependence
Jaden Moon, Arvind Pillai, Andrew Campbell
Multimodal
  • Introduces a leakage-safe diagnostic to assess the influence of quality scores on multimodal predictions.
  • Finds that permuting reliability scores does not significantly degrade model performance in most cases.
  • Demonstrates that quality-aware fusion is effective only when quality signals accurately predict the reliability of modalities.
  • Highlights the importance of distinguishing between correlation and causation in multimodal system performance.
Read more
High-Probability PL-SGD with Markovian Noise: Optimal Mixing and Tail Dependence
Dhruv Sarkar, Aprameyo Chakrabartty, Vaneet Aggarwal
Optimization Theory
  • Establishes a linear dependence on mixing time for high-probability PL-SGD, closing the gap with previous quadratic bounds.
  • Introduces a lag-blocking argument to derive uniform high-probability guarantees under geometric mixing.
  • Extends results to heavy-tailed Markovian gradients with a new clipped block method that addresses Markovian bias.
  • Demonstrates optimality of results through matching lower bounds for both light-tailed and heavy-tailed scenarios.
Read more
Graph Neural Networks Applications Across Domains: All Insights You Need
Abderaouf Bahi
Graph Learning
  • GNNs have become the default model for data with relational structures, moving beyond niche applications.
  • The paper categorizes twelve application domains, detailing graph construction methods and architecture performance.
  • Common challenges across domains include issues with heterophily, temporal graphs, and deployment discrepancies.
  • Over-smoothing, robustness, and explainability are highlighted as critical factors for GNN adoption.
Read more
Heavy-Ball Q-Learning with Residual Weighting Correction
Donghwan Lee
Reinforcement Learning Theory Optimization
  • Introduces a corrected heavy-ball Q-learning method with theoretical guarantees for faster convergence.
  • Utilizes a switched linear system representation to analyze Q-learning dynamics.
  • Establishes conditions for acceleration based on the common eigenvector of mean mappings.
  • Extends the method to linear function approximation with analogous results.
Read more
Localizing RL-Induced Tool Use to a Single Crosscoder Feature
Andrii Shportko, Shubham Bhokare, Ahmed Zeyad A Alzahrani, Bowen Cheng, Gustavo Mercier, Jessica Hullman
NLP Large Language Models Reinforcement Learning
  • Introduces Dedicated Feature Crosscoders (DFC) to isolate RL-specific features for tool use.
  • Demonstrates a +31.1% improvement in tool correctness through RL fine-tuning.
  • Identifies capability spillover, allowing frozen models to benefit from RL fine-tuning without retraining.
  • Shows that steering a single A-exclusive feature can significantly enhance tool-calling performance.
Read more
Asymptotically Optimal Learning for Parametric Prophet Inequalities
Jung-hun Kim, Anna Grebennikova, Vianney Perchet
Theory Optimization
  • Characterization of optimal full-information asymptotic competitive ratios for parametric families.
  • Development of a confidence-based dynamic-programming policy for online learning.
  • Achieving optimal competitive ratios using only online observations without offline samples.
  • Derivation of distribution-specific convergence rates for various reward distributions.
Read more
A Multi-Fidelity Convolutional Autoencoder-Transfer Learning Framework for Guided-Wave-Based Damage Diagnosis Using Large Simulated and Limited Experimental Datasets
Santosh Kapuria, Abhishek
Efficient ML Time Series Theory
  • Introduces a multi-fidelity transfer learning framework for GWSHM.
  • Utilizes lightweight physics-based simulations to generate synthetic datasets.
  • Achieves superior damage localization and sizing accuracy compared to CNN-based methods.
  • Demonstrates strong generalization capabilities on unseen data.
Read more
Effective Covariance Dynamics in Solvable High-Dimensional GANs
Andrew Bond, Zafer Doğan
Generative Models Theory Optimization
  • Introduces effective covariance dynamics for multi-feature GANs with structured latent covariance.
  • Establishes a solvable region in learning-rate and noise space for successful GAN training.
  • Demonstrates a signal-boosting mechanism where weak coordinates can be lifted above the learnability threshold.
  • Validates theoretical findings through numerical simulations and empirical experiments on benchmark datasets.
Read more
How Good Can Linear Models Be for Time-Series Forecasting?
Lang Huang, Jinglue Xu, Luke Darlow
Time Series
  • Ridge regression, when carefully tuned, can outperform complex models like transformers and MLPs in time-series forecasting.
  • Optimal lookback periods are highly dataset-specific and often non-monotonic with respect to forecast horizons.
  • Local normalization strategies consistently yield better forecasting accuracy than global normalization.
  • The study reveals that different time series within the same dataset may require distinct hyperparameter settings.
Read more
Implementation of reinforcement learning in chemical reaction networks: application to phototaxis as curiosity-driven exploration
Ruyi Tang, Grégoire Sergeant-Perthuis, David Colliaux
Reinforcement Learning Robotics Theory
  • Integration of reinforcement learning with biochemical reaction dynamics for modeling phototaxis.
  • Formulation of phototaxis as a subjective POMDP to address sensory ambiguity.
  • Use of inverse reinforcement learning to derive phototactic policies from experimental data.
  • Demonstration that tumbling serves as an information-acquisition strategy rather than random noise.
Read more
Reinforcement Learning without Ground-Truth Solutions can Improve LLMs
Yingyu Lin, Qiyue Gao, Nikki Lijing Kuang, Xunpeng Huang, Kun Zhou, Tongtong Liang, Zhewei Yao, Yi-An Ma, Yuxiong He
Reinforcement Learning Large Language Models Optimization
  • RiVER enables training of LLMs without ground-truth solutions using score-based optimization.
  • The framework addresses challenges of scale and frequency dominance in reinforcement learning.
  • Significant performance improvements were observed in both ALE ratings and exact-solution benchmarks.
  • Calibrated reward shaping enhances the effectiveness of feedback in training LLMs.
Read more
Sketched Linear Contrastive Learning: Approximation, Optimization, and Statistical Scaling
Ziyan Chen, Zhongzhu Zhou, Ding-Xuan Zhou
Theory Optimization Multimodal
  • Introduces a theoretical framework for understanding scaling laws in contrastive learning.
  • Derives a risk decomposition that clarifies the contributions of approximation, optimization, and sampling errors.
  • Demonstrates that contrastive learning requires learning interactions between two views, affecting scaling behavior.
  • Provides an explicit scaling law related to sketch dimension, sample size, and optimization horizon.
Read more
Learning Probabilistic Filters with Strictly Proper Scoring Rules
Eviatar Bach, Ricardo Baptista, Jochen Bröcker, Bohan Chen, Andrew Stuart
Theory Time Series Optimization
  • Introduction of the Proper Scoring Ensemble Filter (PSEF) for Bayesian filtering.
  • Utilizes transformer-based architecture for analysis mapping of forecast ensembles.
  • Training based on strictly proper scoring rules enhances probabilistic accuracy.
  • Theoretical foundation proves minimization of population objective aligns with true filtering distribution.
Read more
Otter Weather: Skillful and Computationally Efficient Medium-Range Weather Forecasting
Cristiana Diaconu, Jonas Scholz, Aliaksandra Shysheya, Stratis Markou, Payel Mukhopadhyay, Miles Cranmer, Richard E. Turner
Efficient ML Time Series
  • Otter Weather democratizes high-performance weather forecasting by significantly reducing training costs.
  • The deterministic model outperforms traditional NWP by 9.6% at a 24-hour lead time while using less than 3.5 A100-days for training.
  • Otter-XL achieves a 9.7% improvement in probabilistic forecasting over the IFS ENS baseline with a 30 A100-day budget.
  • The model demonstrates a 100-fold reduction in compute compared to resource-intensive architectures.
Read more
Finding the Time to Think: Learning Planning Budgets in Real-Time RL
Aneesh Muppidi, Firas Darwish, Dylan Cope, João F. Henriques, Jakob Nicolaus Foerster
Reinforcement Learning
  • Introduction of variable-delay real-time RL, allowing agents to choose deliberation time based on state.
  • Development of a lightweight gating policy that selects state-dependent planning budgets.
  • Empirical characterization of the trade-off between planning quality and inference time.
  • Demonstration of the gating policy's effectiveness across multiple real-time environments.
Read more
Transformer-Based Classification of Bacterial Raman Spectra with LOOCV
Jamile Mohammad Jafari, Thomas Bocklitz
Theory
  • Transformer models significantly outperform conventional machine learning methods in classifying bacterial Raman spectra.
  • The study utilized a nested leave-one-replicate-out cross-validation framework for rigorous model evaluation.
  • Transformers demonstrated robust performance on raw Raman spectra without preprocessing.
  • Improved class separation was observed in the latent feature space learned by the transformer model.
Read more
The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
Alex Iacob, Andrej Jovanović, William F. Shen, Daniel Burkhardt, Meghdad Kurmanji, Nurbek Tastan, Lorenzo Sani, Niccolò Alberto Elia Venanzi, Ambroise Odonnat, Zeyu Cao, Bill Marino, Xinchi Qiu, Nicholas D. Lane
Theory Optimization NLP
  • Introduction of the Red Queen Gödel Machine (RQGM) for recursive self-improvement with evolving evaluators.
  • Controlled utility evolution allows dynamic adaptation of evaluation criteria across epochs.
  • Empirical results show RQGM improves performance in coding tasks, scientific writing, and proof grading.
  • Co-evolved evaluators provide cheaper and more effective evaluation signals compared to static benchmarks.
Read more