AI-generated summaries

Today's ML research,
without the noise.

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

48 Papers today
8h Update frequency
7 Days of history
Reliable Self-Harm Risk Screening via Adaptive Multi-Agent LLM Systems
Meghana Karnam, Ananya Joshi
NLP Large Language Models Theory
  • Introduces a principled framework for multi-agent decision-making in self-harm risk screening.
  • Implements adaptive sampling strategies to efficiently allocate resources based on case complexity.
  • Demonstrates significant reductions in false positive rates while maintaining recall across datasets.
  • Provides a foundation for auditing and deploying AI systems in safety-critical behavioral health settings.
Read more
Preserve Support, Not Correspondence: Dynamic Routing for Offline Reinforcement Learning
Zhancun Mu, Guangyu Zhao, Yiwu Zhong, Chi Zhang
Reinforcement Learning Generative Models Multimodal
  • DROL introduces a dynamic routing mechanism that allows for flexible action selection in offline RL.
  • The method preserves local action support rather than fixed correspondence to a teacher action.
  • DROL demonstrates improved performance on multimodal benchmarks while maintaining efficient inference.
  • The routing mechanism enables ownership of action regions to shift during training, enhancing learning dynamics.
Read more
Unsupervised Learning of Inter-Object Relationships via Group Homomorphism
Kyotaro Ushida, Takayuki Komatsu, Yoshiyuki Ohmura, Yasuo Kuniyoshi
Computer Vision Theory Robotics
  • Proposes a novel unsupervised learning method based on group homomorphism.
  • Integrates object segmentation and motion extraction in a single framework.
  • Demonstrates the ability to segment objects and understand their interactions without labeled data.
  • Introduces algebraic constraints to achieve disentangled representations.
Read more
Reinforcing privacy reasoning in LLMs via normative simulacra from fiction
Matt Franchi, Madiha Zahrah Choksi, Hal Triedman, Helen Nissenbaum
NLP Large Language Models Reinforcement Learning
  • Introduces normative simulacra from fiction as a method to enhance LLM privacy reasoning.
  • Utilizes a two-stage training process combining supervised fine-tuning and reinforcement learning.
  • Implements a composite reward function to evaluate privacy reasoning based on contextual norms.
  • Demonstrates improved alignment of LLM outputs with human privacy expectations across multiple benchmarks.
Read more
Absorber LLM: Harnessing Causal Synchronization for Test-Time Training
Zhixin Zhang, Shabo Zhang, Chengcan Wu, Zeming Wei, Meng Sun
Large Language Models Efficient ML NLP
  • Absorber LLM preserves causal relationships between historical contexts and future inferences.
  • The method reduces memory consumption during inference while maintaining model performance.
  • Causal synchronization is introduced as a mechanism for effective context absorption.
  • Empirical results show superior performance over traditional transformers and linear-time models.
Read more
Logistic Bandits with $ ilde{O}( ext{sqrt}{dT})$ Regret without Context Diversity Assumptions
Seoungbin Bae, Dabeen Lee
Theory Optimization
  • SupSplitLog is the first algorithm for logistic bandits achieving $ ilde{O}( ext{sqrt}{dT})$ regret without context diversity assumptions.
  • The algorithm improves the dependence on dimension d compared to existing methods.
  • SupSplitLog employs a novel sample-splitting technique for constructing estimators.
  • The method can adapt to provide a regret bound based on a data-dependent complexity measure.
Read more
LAF-Based Evaluation and UTTL-Based Learning Strategies with MIATTs
Yongquan Yang
Theory Optimization Interpretability
  • Introduces the EL-MIATTs framework to handle ambiguous true targets in ML.
  • Develops LAF-based evaluation algorithms for coherent model assessment.
  • Proposes UTTL-based learning strategies for effective model training under uncertainty.
  • Analyzes the structural properties of task-specific MIATTs and their impact on evaluation and learning.
Read more
Fairness under uncertainty in sequential decisions
Michelle Seng Ah Lee, Kirtan Padh, David Watson, Niki Kilbertus, Jatinder Singh
Reinforcement Learning Theory Optimization
  • Introduces a taxonomy of uncertainties in sequential decision-making: model, feedback, and prediction uncertainty.
  • Formalizes uncertainties using counterfactual logic and reinforcement learning techniques.
  • Demonstrates potential harms of naive policies that ignore unobserved outcomes.
  • Shows how uncertainty-aware exploration can improve fairness metrics in decision systems.
Read more
mcdok at SemEval-2026 Task 13: Finetuning LLMs for Detection of Machine-Generated Code
Adam Skurla, Dominik Macko, Jakub Simko
Large Language Models NLP
  • Development of mcdok system for detecting machine-generated code in multiple programming languages.
  • Adaptation of the mdok approach for code detection, focusing on binary and multi-class classification tasks.
  • Use of various large language models (LLMs) tailored for better code understanding.
  • Competitive results in all subtasks, indicating the effectiveness of the proposed methods.
Read more
How Much Is One Recurrence Worth? Iso-Depth Scaling Laws for Looped Language Models
Kristian Schwethelm, Daniel Rueckert, Georgios Kaissis
NLP Large Language Models Efficient ML
  • Introduces a joint scaling law for looped language models with a recurrence-equivalence exponent φ = 0.46.
  • Demonstrates that additional recurrences in looped LMs lead to increased validation loss at matched training compute.
  • Establishes a five-axis evaluation suite to analyze the performance of looped LMs across different tasks.
  • Finds that looped models prefer wider architectures with fewer training tokens compared to non-looped models.
Read more
Drug Synergy Prediction via Residual Graph Isomorphism Networks and Attention Mechanisms
Jiyan Song, Wenyang Wang, Chengcheng Yan, Zhiquan Han, Feifei Zhao
Graph Learning
  • Introduces ResGIN-Att, a novel model for predicting drug synergy.
  • Integrates molecular features, genomic profiles, and drug interactions for enhanced predictions.
  • Employs residual connections to mitigate over-smoothing in deep learning layers.
  • Demonstrates competitive performance on benchmark datasets, showcasing robustness and generalization.
Read more
Sharpness-Aware Poisoning: Enhancing Transferability of Injective Attacks on Recommender Systems
Junsong Xie, Yonghui Yang, Pengyang Shao, Le Wu
Optimization Theory
  • Introduces Sharpness-Aware Poisoning (SharpAP) to enhance the transferability of injective attacks on recommender systems.
  • Addresses the limitations of using fixed surrogate models for generating poisoned data.
  • Implements a min-max-min tri-level optimization framework to optimize poisoned data against the worst-case victim model.
  • Demonstrates significant improvements in attack transferability through comprehensive experiments on real-world datasets.
Read more
Iterative Model-Learning Scheme via Gaussian Processes for Nonlinear Model Predictive Control of (Semi-)Batch Processes
Tai Xuan Tan, Alexander Mitsos, Eike Cramer
Optimization
  • Introduces GP-MLMPC, an iterative model-learning scheme for NMPC using Gaussian Processes.
  • Demonstrates significant performance improvements in batch process control with limited initial data.
  • Achieves an 83% reduction in tracking error and a 17-fold increase in final product mass after eight iterations.
  • Utilizes uncertainty quantification from GPs to enforce chance constraints for safe operation.
Read more
Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics
Weizhi Nie, Zhen Qu, Weijie Wang, Chunpei Li, Ke Lu, Bingyang Zhou, Hongzhi Yu
Large Language Models Time Series Interpretability
  • Introduces a LLM-guided framework for sepsis early warning that enhances interpretability.
  • Combines spatiotemporal feature extraction with clinical reasoning prompts to improve prediction accuracy.
  • Achieves superior AUC scores compared to traditional models, indicating better predictive performance.
  • Provides interpretable trajectories that assist clinicians in understanding physiological deterioration.
Read more
Tempered Sequential Monte Carlo for Trajectory and Policy Optimization with Differentiable Dynamics
Heng Yang
Reinforcement Learning Optimization Robotics
  • Introduces a KL-regularized approach to trajectory and policy optimization that leverages differentiable dynamics.
  • Develops a novel tempered sequential Monte Carlo (TSMC) method for efficient sampling from multimodal distributions.
  • Combines sampling-based exploration with gradient-based optimization to enhance performance in trajectory and policy optimization tasks.
  • Demonstrates the effectiveness of TSMC through experiments that show superior performance compared to existing methods.
Read more
Fast Neural-Network Approximation of Active Target Search Under Uncertainty
Bilal Yousuf, Zsofia Lendek, Lucian Busoniu
Robotics Optimization Efficient ML
  • Introduces a CNN-based approximation for Active Search (AS) and Intermittent Active Search (ASI) to enhance computational efficiency.
  • Utilizes a multi-channel grid representation to encode critical information for decision-making.
  • Demonstrates that the CNN achieves detection rates comparable to AS and ASI while significantly reducing computation time.
  • Validates the approach through extensive simulations with both uniform and clustered target distributions.
Read more
Multi-Task Optimization over Networks of Tasks
Julian Hatzky, Thomas Bartz-Beielstein, A. E. Eiben, Anil Yaman
Optimization Robotics Graph Learning
  • Introduction of MONET, a graph-based multi-task optimization algorithm.
  • MONET addresses scalability issues of existing multi-task optimization methods.
  • Combines individual and social learning strategies for knowledge transfer.
  • Empirical results show MONET outperforms MAP-Elites variants across multiple domains.
Read more
Kernel Contracts: A Specification Language for ML Kernel Correctness Across Heterogeneous Silicon
Cooper Veit
Theory
  • Introduces a formal specification language for ML kernel contracts to address implicit agreements in computations.
  • Defines twelve contract classes based on empirical evidence, covering various failure modes.
  • Establishes a three-state calibration requirement for testable contracts.
  • Demonstrates the application of the framework through three case studies of kernel failures.
Read more
Geometric Monomial (GEM): a family of rational 2N-differentiable activation functions
Eylon E. Krause
Optimization Theory Efficient ML
  • GEM is a family of C2N-smooth activation functions that improve upon ReLU's limitations.
  • The introduction of E-GEM and SE-GEM variants allows for greater flexibility and performance optimization.
  • GEM outperforms GELU in several benchmarks, particularly in deep CNNs and transformers.
  • The smoothness parameter N plays a crucial role in determining the effectiveness of the activation function based on the architecture used.
Read more
The Path Not Taken: Duality in Reasoning about Program Execution
Eshgin Hasanov, Md Mahadi Hassan Sibat, Santu Karmaker, Aashish Yadavally
Large Language Models
  • Current benchmarks for LLMs focus too narrowly on single execution paths, limiting their evaluation of program understanding.
  • The proposed duality framework introduces forward and backward reasoning tasks to better assess LLMs' causal understanding of program execution.
  • DEXBENCH, the new benchmark, comprises 445 paired instances that facilitate a more robust evaluation of LLMs.
  • Results show that dual-path reasoning can reveal limitations in models that perform well in isolation but struggle under joint evaluation.
Read more
Sum-of-Checks: Structured Reasoning for Surgical Safety with Large Vision-Language Models
Weiqiu You, Cassandra Goldberg, Amin Madani, Daniel A. Hashimoto, Eric Wong
Computer Vision Multimodal Interpretability
  • Introduction of Sum-of-Checks framework for structured surgical safety assessment.
  • Framework decomposes CVS criteria into expert-defined reasoning checks.
  • Demonstrated improvement in accuracy and transparency of LVLM-based assessments.
  • LVLMs show reliable performance on observational checks but variability on anatomical evidence.
Read more
Do Not Imitate, Reinforce: Iterative Classification via Belief Refinement
Mahdi Kallel, Johannes Tölle, Ahmed Hendawy, Carlo D'Eramo
Reinforcement Learning Computer Vision Efficient ML
  • Introduction of Reinforced Iterative Classification (RIC) as a novel approach to classification using reinforcement learning.
  • RIC allows for iterative refinement of predictions, improving calibration and reducing overconfidence in model outputs.
  • The framework provides a natural mechanism for adaptive computation, dynamically allocating resources based on input complexity.
  • Empirical results show that RIC achieves competitive accuracy while enhancing calibration on multiple image classification benchmarks.
Read more
Channel-Free Human Activity Recognition via Inductive-Bias-Aware Fusion Design for Heterogeneous IoT Sensor Environments
Tatsuhito Hasegawa
Time Series Multimodal
  • Introduces a channel-free HAR framework that adapts to heterogeneous sensor environments.
  • Utilizes metadata for improved structural information recovery during activity recognition.
  • Demonstrates strong robustness against channel perturbations and improved performance with metadata conditioning.
  • Maintains competitiveness with traditional channel-fixed models while enabling cross-dataset transfer learning.
Read more
From Local to Cluster: A Unified Framework for Causal Discovery with Latent Variables
Zongyu Li
Graph Learning Theory Efficient ML
  • L2C framework bridges local structure learning and cluster-level causal discovery.
  • Automatically discovers clusters from local causal patterns without manual assignment.
  • Handles latent variables effectively without assuming causal sufficiency.
  • Theoretical guarantees of soundness, atomic completeness, and computational efficiency.
Read more
Toward Efficient Membership Inference Attacks against Federated Large Language Models: A Projection Residual Approach
Guilin Deng, Silong Chen, Yuchuan Luo, Yi Liu, Songlei Wang, Zhiping Cai, Lin Liu, Xiaohua Jia, Shaojing Fu
NLP Large Language Models Federated Learning
  • ProjRes is the first projection residuals-based passive MIA specifically designed for FedLLMs.
  • The method achieves near 100% accuracy in inferring data membership, outperforming existing techniques.
  • ProjRes operates without the need for shadow models or auxiliary classifiers, enhancing efficiency.
  • The study reveals significant privacy vulnerabilities in FedLLMs, necessitating a reassessment of their security frameworks.
Read more
Droplet-LNO: Physics-Informed Laplace Neural Operators for Accurate Prediction of Droplet Spreading Dynamics on Complex Surfaces
Ganesh Sahadeo Meshram, Partha Pratim Chakrabarti, Suman Chakraborty
Theory Efficient ML Optimization
  • Introduction of PI-LNO, a novel neural network architecture for droplet dynamics.
  • Achieves significant speedup in predictions compared to traditional CFD methods.
  • Demonstrates superior accuracy with a mean R2 score of 0.9009 across various conditions.
  • Utilizes a physics-regularized loss function to ensure physically feasible predictions.
Read more
On the Properties of Feature Attribution for Supervised Contrastive Learning
Leonardo Arrighi, Julia Eva Belloni, Aurélie Gallet, Ivan Gentile, Matteo Lippi, Marco Zullich
Computer Vision Interpretability
  • Supervised Contrastive Learning (SCL) enhances feature attribution quality compared to Cross-Entropy (CE) loss.
  • Models trained with SCL show improved robustness and generalization capabilities.
  • Grad-CAM-based feature attributions from SCL-trained models are more faithful and continuous.
  • Lower contrastivity in SCL models indicates a more stable feature attribution across classes.
Read more
A Nationwide Japanese Medical Claims Foundation Model: Balancing Model Scaling and Task-Specific Computational Efficiency
Nanae Aratake, Taisei Tosaki, Yuji Okamoto, Eiichiro Uchino, Masaki Nakamura, Nobutomo Matsui, Akiko Hatakama, Yasushi Okuno
Efficient ML
  • Systematic evaluation across five model scales revealed task-dependent performance ceilings.
  • Disease prediction improved with larger models, while medication prediction saturated at a smaller size.
  • Optimal model sizes can lead to significant reductions in pretraining time without sacrificing performance.
  • Task-specific capacity ceilings are essential for efficient resource allocation in model development.
Read more
Hidden Failure Modes of Gradient Modification under Adam in Continual Learning, and Adaptive Decoupled Moment Routing as a Repair
Yuelin Hu, Zhenbo Yu, Zhengxue Cheng, Wei Liu, Li Song
Optimization Theory
  • Identification of the 'attenuate-then-adapt conflict' in gradient modification under Adam.
  • Demonstration that traditional methods lead to increased forgetting in continual learning tasks.
  • Introduction of Adaptive Decoupled Moment Routing as a solution to mitigate identified failures.
  • Empirical validation showing significant performance improvements over existing methods.
Read more
Reliability Auditing for Downstream LLM tasks in Psychiatry: LLM-Generated Hospitalization Risk Scores
Shevya Pandya, Shinjini Bose, Ananya Joshi
Large Language Models NLP Interpretability
  • LLMs are sensitive to clinically insignificant variables, affecting psychiatric risk assessments.
  • Prompt design significantly influences model outputs, necessitating controlled methodologies in AI healthcare applications.
  • A structured audit framework can identify reliability issues in LLMs, particularly in psychiatric contexts.
  • Increased output variability correlates with the addition of irrelevant features, highlighting predictive instability.
Read more
Focus Session: Hardware and Software Techniques for Accelerating Multimodal Foundation Models
Muhammad Shafique, Abdul Basit, Muhammad Abdullah Hanif, Alberto Marchisio, Rachmad Vidya Wicaksana Putra, Minghao Shao
Multimodal Efficient ML Optimization
  • Presents a multi-layered methodology for accelerating multimodal foundation models (MFMs).
  • Integrates hardware and software optimization techniques to enhance energy efficiency and performance.
  • Employs advanced techniques such as mixed-precision quantization, structural pruning, and model cascading.
  • Demonstrates effectiveness on medical MFMs and code generation tasks.
Read more
FairyFuse: Multiplication-Free LLM Inference on CPUs via Fused Ternary Kernels
Fei Zuo, Xiaoyan Xi, Quanyi Zeng, Feiyu Wang, Ho Fai Leung
NLP Large Language Models Efficient ML
  • FairyFuse is the first ternary-weight GEMV kernel on x86 CPUs that eliminates floating-point multiplications.
  • The system achieves a 29.6× speedup over FP32 by optimizing memory bandwidth usage through fused execution.
  • End-to-end evaluation shows FairyFuse outperforms llama.cpp Q4_K_M by 1.24× while maintaining high model quality.
  • The findings suggest that CPUs are more suitable than GPUs for extreme quantization in LLMs.
Read more
Performance Anomaly Detection in Athletics: A Benchmarking System with Visual Analytics
Blessed Madukoma, Prasenjit Mitra
Time Series
  • The system processes a large dataset of athletic performances to identify potential doping violations.
  • Eight detection methods are utilized, including statistical and machine learning techniques.
  • Trajectory-based methods outperform others in balancing detection and false alarm rates.
  • The system supports expert-driven investigations with an interactive visual analytics interface.
Read more
ARFBench: Benchmarking Time Series Question Answering Ability for Software Incident Response
Stephan Xie, Ben Cohen, Mononito Goswami, Junhong Shen, Emaad Khwaja, Chenghao Liu, David Asker, Othmane Abou-Amal, Ameet Talwalkar
Time Series NLP Multimodal
  • ARFBench provides a comprehensive evaluation framework for TSQA in software incident response.
  • Frontier VLMs significantly outperform existing baselines in TSQA tasks.
  • Hybrid TSFM-VLM models show promise for specialized time series question answering.
  • A model-expert oracle approach demonstrates complementary strengths between models and human experts.
Read more
Learning Coverage- and Power-Optimal Transmitter Placement from Building Maps: A Comparative Study of Direct and Indirect Neural Approaches
Çağkan Yapar
Optimization
  • Introduces a large dataset (RadioMapSeer-Deployment) for transmitter placement with dual labels.
  • Identifies an asymmetric trade-off between coverage and power in transmitter placements.
  • Compares indirect heatmap-based and direct score-map models for transmitter placement.
  • Demonstrates significant speed improvements in predictions using neural network models.
Read more
Protect the Brain When Treating the Heart: A Convolutional Neural Network for Detecting Emboli
Andrea Angino, Ken Trotti, Diego Ulisse Pizzagalli, Rolf Krause, Tiziano Torre, Stefanos Demertzis
Computer Vision
  • Introduction of a 2.5D U-Net architecture for real-time GME detection.
  • Development of a custom annotation tool for creating a specialized dataset.
  • Demonstration of high segmentation accuracy and robust detection capabilities.
  • Integration of the model into surgical protocols for real-time patient monitoring.
Read more
Estimating Tail Risks in Language Model Output Distributions
Rico Angell, Raghav Singhal, Zachary Horvitz, Zhou Yu, Rajesh Ranganath, Kathleen McKeown, He He
NLP Large Language Models Efficient ML
  • Introduces a method for estimating the probability of harmful outputs in language models using importance sampling.
  • Demonstrates that unsafe model versions can be created to enhance the likelihood of harmful outputs, allowing for efficient sampling.
  • Achieves accurate estimates of harmful output probabilities with significantly fewer samples than traditional methods.
  • Reveals the sensitivity of model outputs to input perturbations, indicating that query-level estimates are crucial for understanding deployment risks.
Read more
Data-Driven Open-Loop Simulation for Digital-Twin Operator Decision Support in Wastewater Treatment
Gary Simethy, Daniel Ortiz Arroyo, Petar Durdevic
Time Series
  • Introduction of CCSS-RS, a novel data-driven simulator for wastewater treatment decision support.
  • Model effectively handles irregular and missing sensor data, crucial for real-world applications.
  • Demonstrated significant predictive accuracy improvements over existing models.
  • Case studies validate the model's operational utility in real-time decision-making.
Read more
Sink-Token-Aware Pruning for Fine-Grained Video Understanding in Efficient Video LLMs
Kibum Kim, Jiwan Kim, Kyle Min, Yueqi Wang, Jinyoung Moon, Julian McAuley, Chanyoung Park
Computer Vision Large Language Models Efficient ML
  • Identifies sink tokens as a critical barrier to fine-grained video understanding.
  • Proposes Sink-Token-aware Pruning (SToP) to enhance existing pruning methods.
  • Demonstrates significant performance improvements across diverse benchmarks.
  • Validates the effectiveness of SToP in maintaining visual grounding during pruning.
Read more
A Hybridizable Neural Time Integrator for Stable Autoregressive Forecasting
Brooks Kinch, Xiaozhe Hu, Yilong Huang, Martine Dyring Hansen, Sunniva Meltzer, Nathaniel Donald Hamlin, David Sirajuddin, Eric C. Cyr, Nathaniel Trask
Time Series Theory Efficient ML
  • Introduces a hybrid autoregressive transformer embedded in a mixed finite element framework for stable forecasting.
  • Proves preservation of discrete energies and uniform gradient bounds, addressing the exploding gradient problem.
  • Achieves a 65× reduction in model parameters compared to existing models while maintaining high forecasting accuracy.
  • Demonstrates real-time surrogate modeling capabilities with significant speedup over traditional simulation methods.
Read more
Generalizing Numerical Reasoning in Table Data through Operation Sketches and Self-Supervised Learning
Hanjun Cho, Gahyun Yoo, Hanseong Kim, Jay-Yoon Lee
NLP Large Language Models
  • Identifies three failure modes in numerical reasoning: reasoning inefficiency, data scarcity for logical supervision, and header dependency.
  • Introduces operation sketches to enhance contextual reasoning and reduce reliance on surface-level patterns.
  • Combines header anonymization and self-supervised learning to improve data efficiency and robustness.
  • Demonstrates superior performance of TaNOS over traditional SFT methods in both in-domain and cross-domain settings.
Read more
Conditional anomaly detection with soft harmonic functions
Michal Valko, Branislav Kveton, Hamed Valizadegan, Gregory F. Cooper, Milos Hauskrecht
Graph Learning
  • Introduction of a non-parametric method for conditional anomaly detection using soft harmonic functions.
  • Regularization techniques to avoid detection of isolated and fringe points.
  • Development of a backbone graph for efficient computation in large datasets.
  • Demonstration of the method's efficacy on synthetic, UCI, and real-world datasets.
Read more
SCM: Sleep-Consolidated Memory with Algorithmic Forgetting for Large Language Models
Saish Sachin Shinde
Large Language Models NLP Theory
  • Introduces a unified memory architecture for conversational agents that mimics biological memory processes.
  • Implements multi-dimensional value tagging for richer memory prioritization.
  • Achieves perfect recall in multi-turn conversations while significantly reducing memory noise.
  • Describes a complete system design including synchronization and visualization tools.
Read more
Conditional anomaly detection using soft harmonic functions: An application to clinical alerting
Michal Valko, Hamed Valizadegan, Branislav Kveton, Gregory F. Cooper, Milos Hauskrecht
Graph Learning Theory Efficient ML
  • Introduces a non-parametric approach for conditional anomaly detection using soft harmonic functions.
  • Focuses on identifying unusual patient-management decisions to prevent medical errors.
  • Incorporates regularization to handle isolated and fringe points in the data.
  • Demonstrates effectiveness on a real-world electronic health record dataset.
Read more
PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training
Harsh Kumar, Rahul Maity, Tanmay Joshi, Aman Chadha, Vinija Jain, Suranjana Trivedy, Amitava Das
Large Language Models NLP Theory
  • Introduction of the Stealth Pretraining Seeding (SPS) threat model for LLMs.
  • Development of the PermaFrost-Attack framework to study latent conceptual poisoning.
  • Introduction of three geometric diagnostics for analyzing adversarial influence in LLMs.
  • Empirical evidence showing persistent unsafe behaviors induced by SPS across multiple model families.
Read more
Zero-Shot Morphological Discovery in Low-Resource Bantu Languages via Cross-Lingual Transfer and Unsupervised Clustering
Hillary Mutisya, John Mugane
NLP
  • Novel multi-method approach combining transfer learning and unsupervised clustering for morphological analysis.
  • Discovered 2,455 noun class labels in Giriama, significantly increasing the available morphological data.
  • Identified two previously undocumented morphological patterns in Giriama.
  • Achieved 78.2% lemmatization accuracy on known paradigms and high segmentation rates.
Read more
Distance-Misaligned Training in Graph Transformers and Adaptive Graph-Aware Control
Qinhan Hou, Jing Tang
Graph Learning
  • Distance-misaligned training highlights the mismatch between task-relevant information and model communication strategies.
  • The preferred graph-distance bias varies with task locality, indicating the need for adaptive control.
  • An oracle adaptive controller outperforms fixed bias settings, demonstrating the importance of task-specific distance targets.
  • Distance-resolved diagnostics can effectively identify over-globalizing and under-reaching failures in Graph Transformers.
Read more
Even More Guarantees for Variational Inference in the Presence of Symmetries
Lena Zellinger, Antonio Vergari
Theory Optimization
  • Establishes sufficient conditions for exact recovery of the mean using FKL and α-divergences.
  • Extends previous results on robust variational inference under target symmetries.
  • Provides guidelines for selecting variational families based on the derived conditions.
  • Highlights potential optimization failures when sufficient conditions are not satisfied.
Read more