Security vulnerability in RLHF where LLMs can influence preference datasets during alignment training to amplify undesired behaviors.
Analysis of how single-axis reward model bias mitigations redirect optimization pressure to correlated proxies rather than eliminating bias.
Memory management system for stateful LLM applications that generalizes memory extraction across use cases beyond fixed context windows.
Framework enabling AI agents to recognize knowledge boundaries and avoid unnecessary searches, improving efficiency of LLM-based agentic systems.
Benchmark for evaluating multimodal LLMs across 19 materials science subfields, testing reasoning from knowledge to practical application.
Method for synthesizing domain-specific training data for LLMs using minimal sufficient representation learning to address data acquisition challenges in fine-tuning.
MIRA method for source-aware mid-training data selection balancing pretraining and downstream capability optimization in LLMs.
Survey of graph machine learning integration with large language models for tasks in social networks and knowledge graphs.
Comprehensive study of fundamental engineering trade-offs in RAG system design including retrieval, generation, and augmentation decisions.
Cross-modal attention calibration technique reducing hallucinations in large vision-language models during generation.
Diagnostic evaluation using Construction Grammar to assess LLM generalization beyond memorization on out-of-domain language.
PRISM method for training-free multimodal data selection reducing dataset redundancy in visual instruction tuning.
Auto-Discovery-Bench diagnostic benchmark for evaluating agent state tracking and hypothesis-intervention-feedback cycles.
EMCEE method improving multilingual LLM capabilities by extracting synthetic multilingual context bridging knowledge and reasoning.
Novel Bayesian sampling approach for membership inference attacks reducing computational overhead of reference model training.
Orthogonal subspace approach for merging multiple LoRA-fine-tuned language models without performance degradation.
Framework for attributing model behavior to specific training stages (pretraining, fine-tuning, adaptation) in multi-stage AI systems.
Analysis of machine unlearning vulnerabilities including over-unlearning effects and prototypical relearning attacks.
SHIELD framework combining interval bound propagation with hypernetworks for certifiably robust continual learning.
DISCO framework using conditional distance correlation and causal theory to mitigate dataset bias in deep learning models.
Vision-only in-context learning models for few-shot image classification emphasizing importance of encoder pretraining.
Methods for calibrating uncertainty in LLM inference to improve model control and user trust in real-time applications.
Method for improving neural network calibration under distribution shift without requiring target domain access, using frequency-aware gradient rectification.
Introduces reasoning-intensive regression using LLMs to deduce subtle numerical scores from text for rubric-based scoring and reward modeling.
Examines whether human psychometric questionnaires reliably characterize LLM behavior by comparing self-report and generation probability methods.
Introduces MedFact benchmark with 2,116 expert-annotated Chinese medical texts to evaluate fact-checking capabilities of LLMs in healthcare.
Proposes Atom Theory to define and identify fundamental representational units in LLMs using non-Euclidean atomic inner product metric.
Introduces TimeRCD foundation model for zero-shot time series anomaly detection using synthetic data and relative context discrepancy scoring.
SAEmnesia uses supervised sparse autoencoders to erase concepts in diffusion models by enforcing one-to-one concept-neuron mappings.
Shows Hessian spectral collapse causes loss of plasticity in continual learning, derives conditions for successful learning of new tasks.
Analyzes intrinsic vs. prompted value expression mechanisms in LLMs, examining whether they overlap or rely on distinct underlying mechanisms.
Investigates whether LLMs condition reasoning on formal programming semantics or learned statistical priors through systematic program execution analysis.
OBCache optimally prunes key-value cache in LLMs for efficient long-context inference using true attention impact rather than heuristic rankings.
Derives PAC-Bayesian generalization bounds for reinforcement learning accounting for Markov dependencies, enabling non-vacuous certificates for modern RL.
Proposes boundary-guided policy optimization for memory-efficient reinforcement learning of diffusion large language models without full likelihood computation.
CaptionFormer jointly performs object detection, tracking, and captioning on video trajectories using unified spatio-temporal approach.
InfiMed-ORBIT aligns LLMs on open-ended medical tasks using rubric-based incremental training with reinforcement learning for complex dialogue.
Proposes diffusion model-based approach to scale multi-agent environment co-design, enabling joint optimization of agent policies and environment configurations.
Studies action chunk length trade-offs in vision-language-action robotic models, proposes mixture of horizons approach for improved performance.
Bottom-up policy optimization decomposes LLM policies into internal layer/modular policies; applies RL via Transformer residual streams.
FEM-Bench benchmark for evaluating code-generating LLMs on computational mechanics reasoning and scientific model validity.
Critique of LLM reasoning claims arguing they lack falsifiability; identifies methodological pitfalls in AI research validation.
PASTA framework for scalable multi-policy AI compliance evaluation with model-card format and four integration innovations.
Gap-K% metric detects pretraining data in LLMs by measuring prediction gaps and token correlations for privacy/copyright concerns.
Model merging approach decouples data mixture search from training, enabling scalable optimization for LLM pre-training.
Study shows self-organizing multi-agent LLM teams underperform expert models; examines coordination emergence in autonomous collaborators.
Mixture of Concept Bottleneck Experts framework improves interpretability and accuracy of concept-based models with multiple task predictors.
Multi-agent framework automatically transforms CVE metadata into executable security vulnerability tasks for evaluating code agent capabilities.
Context-preserving verification method reduces flip-flop oscillations in parallel diffusion decoding for faster language model inference.
Clean-PR training paradigm leverages GitHub pull requests to improve repository-level code editing models for multi-file modifications.