All Articles
16523 articles total
Statistically Grounded Sparse-Feature Interventions for Activation-Space Control in Large Language Models
Develops a transparent SAE-feature steering pipeline for behavioral control in LLMs using a statistically grounded Borda consensus approach.
Logic-Guided Data Extraction with Answer Set Programming and Large Language Models
Combines LLM-based data extraction with Answer Set Programming (ASP) to ensure logical consistency and reduce the number of LLM calls required for complex extraction tasks.
Geometry-Guided Constraint Learning for LLM Safety Classification
Explores the use of Sparse Autoencoders (SAE) to optimize safety classification in LLMs, suggesting safety boundaries admit low-dimensional linear descriptions in SAE feature space.
Rethinking Uncertainty Evaluation in Large Language Models
Critiques current uncertainty evaluation in LLMs, arguing that calibration is insufficient and proposing the C1 metrics to measure coherent probabilistic beliefs.
Spectral-LSH: Sub-Quadratic Prompt Compression via Krylov-Projected Locality-Sensitive Hashing
Introduces Spectral-LSH, a training-free prompt compression method using Krylov-projected Locality-Sensitive Hashing to reduce prefill attention costs.
Beyond Tracking or Shortcut: Composition-Bounded Predictive States in Poker Autoregressive Models
Analyzes opponent-range representation in poker autoregressive models, finding that most predictive information stems from visible betting patterns rather than hidden states.
git's –end-of-options Flag
Discussion regarding the use of the --end-of-options flag in git to explicitly separate options from positional arguments.
FineServe: A Fine-Grained Dataset and Characterization of Global LLM Serving Workloads
FineServe introduces a multi-model LLM serving workload dataset and generator to better benchmark routing and scheduling in production environments.
Hybrid LSTM-Graph Neural Framework for Robust Financial Fraud Detection and Adversarial Resilience
FraudShield AI is a hybrid LSTM-Graph Neural framework designed to detect sophisticated financial fraud by combining temporal sequences and relational context.
OpenEvoShield: Dual Non-Stationary Continual Defense for Open-World Multi-Agent System Attacks
OpenEvoShield is a co-evolutionary continual defense framework designed to protect multi-agent LLM systems from dynamic adversarial attacks.
Benchmarking Confidential GPU Inference on NVIDIA H100 under Intel TDX
A benchmark study on the performance overhead of confidential GPU inference using NVIDIA H100 GPUs under Intel TDX.
FormulaSPIN: Self-Play Fine-Tuning for Natural Language to Spreadsheet Formula Generation
FormulaSPIN introduces a self-play fine-tuning framework that improves natural language to spreadsheet formula generation without additional data.
Information Discernment in Large Language Models
The Learn2Discern (L2D) framework reveals that LLMs struggle with source and truth discernment when integrating external knowledge.
NEXUS: Structured Runtime Safety for Tool-Using LLM Agents
NEXUS is a structured runtime safety monitor for tool-using LLM agents that uses a graded escalation policy to block or revise risky actions.
Stochastic Primal-Dual Decoding for Multiobjective Generative Recommender Systems
A new stochastic primal-dual decoding layer for generative recommender systems to balance relevance with multi-objective constraints at inference time.
LISA: Linear-Indexed Sparse Attention for Efficient Long-Context Reasoning
LISA is a plug-and-play attention replacement module that reduces inference complexity from O(n^2) to O(nM) for long-context reasoning models.
Don't Blame the Large Language Model: How Agent Harness Evolution Shapes Coding Agent Quality
A study isolates the impact of 'agent harnesses' (middleware) on coding agent quality, finding that rapid iterations in the harness can cause significant regressions regardless of the underlying LLM.
Auto-AEG: Scalable Data Construction for Open-Vocabulary Audio Event Grounding
The Auto-AEG pipeline enables scalable data construction for open-vocabulary audio event grounding, introducing the AEGBench dataset for evaluating temporal localization in LALMs.
Human Grounded Evaluation of Large Language Models for Optical Network Automation
HuGLEN is introduced as a stepwise evaluation pipeline that combines LLM-as-a-judge with expert ratings to optimize the trade-off between quality and efficiency for network automation.
Enhancing Rubric-based RL via Self-Distillation
The CriPO method enhances rubric-based RL through on-policy self-distillation to solve exploration failures (unexplored and suppressed criteria) in LLM training.