Proposes MASPO algorithm improving gradient utilization and sample efficiency in RL for LLM reasoning tasks beyond GRPO limitations.
Theoretical framework for training modular LLMs by combining domain-specific expert models robustly without heuristic dataset weighting.
Investigates weight regularization techniques in parameter-efficient continual learning with low-rank adapters for pre-trained models.
Theoretical analysis of normalization strategies and their impact on expressivity of Transformer-based time series models.
Studies canonicalization of representation spaces across independently trained multimodal contrastive models for consistency.
Develops anytime-valid statistical watermarking method to distinguish machine-generated text from human content in LLMs.
Proposes privacy-preserving federated split learning with intermediate representation protection for distributed ML training.
Addresses variance issues in asynchronous RL training for LLMs using policy-gradient methods like REINFORCE and GRPO on reasoning tasks.
arXiv paper on federated learning with incremental data under limited communication. Addresses catastrophic forgetting in privacy-preserving distributed learning scenarios.
Framework distinguishing weak verification (self-consistency, proxy rewards) from strong verification (human inspection) in LLM reasoning loops.
Reverso foundation model for efficient zero-shot time series forecasting across diverse domains.
FAMOSE framework uses ReAct paradigm with LLM agents for autonomous feature engineering in tabular data.
Black-box adversarial attacks on large vision-language models using fine-grained detail targeting to overcome gradient-free optimization challenges.
Framework for multi-round human-AI collaboration using counterfactual harm and complementarity principles to ensure conversational AI reliably improves decision quality.
Margin-aware reward modeling framework with self-refinement for RLHF/RLAIF alignment pipelines, reducing reliance on human preference data through augmentation.
Efficient KV cache prefetching for LLM inference using GPU-native media ASICs, addressing bandwidth limitations in remote cache reuse scenarios.
MobCache framework using LLMs for scalable large-scale human mobility simulation with caching optimization.
Study showing AI safety datasets overrely on obvious triggering cues and fail to reflect realistic adversarial attacks.
SEMAS self-evolving multi-agent system for industrial IoT predictive maintenance with real-time anomaly detection.
Empirical study of adversarial code comments manipulating LLM vulnerability detection across Python, JavaScript, and Java.
Large-scale deanonymization attack using LLM agents with internet access to re-identify pseudonymous online profiles.
Using reference-guided LLM-evaluators as soft verifiers to improve LLM alignment in non-verifiable domains.
Comparison of simple baselines against code evolution techniques across mathematical bounds, agentic scaffolds, and ML competitions.
Hybrid-Gym environment for training coding agents on diverse software engineering tasks beyond single GitHub issues.
NeST selective neuron tuning approach for parameter-efficient LLM safety alignment without full fine-tuning overhead.
MALLVi multi-agent framework combining LLMs and vision for closed-loop robotic manipulation with environmental feedback.
LLM-WikiRace benchmark for evaluating long-term planning, reasoning, and knowledge navigation in language models over Wikipedia.
Multi-objective optimization and quantum hybridization of Allegro interatomic potential model for molecular property prediction.
DeepContext framework for stateful monitoring of multi-turn LLM conversations to detect adversarial intent drift and bypass safety guardrails.
LLM4Cov: offline agent learning framework for hardware verification testbench generation using execution-aware LLM agents without online reinforcement learning.
Phantom: automated agent hijacking attack via structural template injection, bypassing LLM safety measures with higher success rates and transferability.
Greedy multi-path verification algorithm accelerating speculative decoding by optimizing token acceptance in draft models.
PRIMO model quantifying predictive importance of modalities in multimodal LLMs when data is incomplete or asynchronous.
Cross-lingual text classification methods for analyzing multilingual social media discourse across multiple languages.
Diffusion-based method for generating high-dimensional samples from moment constraints with maximum entropy guarantees.
Privacy-preserving mechanism enabling verification that LLM inference providers run the correct model without replacing it with a weaker one.
Framework for detecting temporal contamination in LLM backtesting by identifying post-cutoff knowledge leakage during training.
Analysis of representation collapse in Transformer-based neural machine translation models using angular dispersion metrics.
AI agent for medical diagnosis that uses LLMs to ask follow-up questions and reason over differential diagnoses iteratively.
Taxonomy for uncertainty quantification in long-form LLM outputs to detect hallucinations, addressing limitations of existing short-form methods.
Study of position and label biases in LLM multiple-choice question answering via synthetic benchmark evaluation.
Open-source Python package implementing debiased machine learning via Riesz regression for causal parameter estimation.
Decentralized optimization method with adaptive stepsizes for multi-agent networks using three-operator splitting.
Adaptive regularization framework for maintaining LLM safety during fine-tuning while preserving utility.
Unsupervised anomaly detection framework using conditional flow matching for autonomous vehicle safety validation.
Online meta-learning approach for geospatial discovery using latent concepts with strategic sampling under constraints.
Multi-agent autonomous framework for designing, implementing, and verifying numerical PDE solvers without manual tuning.
Pruning technique for diffusion language models reducing inference cost by reconsidering attention sink preservation.
Risk-aware decision making algorithms for restless bandits incorporating downside risk mitigation.
Benchmark for evaluating physical safety risks of LLMs controlling robotic systems like drones with threat classification.