Concurrent training enhancements for Kolmogorov-Arnold networks using Newton-Kaczmarz method with FPGA implementation for improved efficiency.
Dual-State Action Pair (DSAP) primitive couples stochastic LLM generation with deterministic verification for reliable code generation agents.
Extends Puzzle neural architecture search to reasoning LLMs, producing gpt-oss-puzzle-88B through MoE expert pruning and inference optimization.
Combines low-rank adaptation with quantization-aware unlearning to ensure LLM knowledge removal survives post-training 4-bit quantization.
Golden Layers method improves LLM knowledge editing via layer gradient analysis to identify optimal depth for updating model predictions per query.
cc-Shapley extends Shapley values for multivariate feature importance by incorporating causal context to address spurious associations.
TRC² architecture for continual learning in LLMs preventing catastrophic forgetting through decoder-only thalamic routing of cortical columns.
AgentTrace framework for post-hoc root cause analysis in deployed multi-agent systems via causal graph reconstruction from execution logs.
Connects adversarial robustness and LLM hallucinations through shared geometric principle formalized as Neural Uncertainty Principle with irreducible uncertainty bounds.
mSFT algorithm addresses overfitting in multi-task supervised fine-tuning by dynamically adjusting data mixture ratios based on task-specific learning dynamics.
Decouples exploration from policy optimization in RL using uncertainty-guided tree search for efficient autonomous exploration without intrinsic motivation.
Method for steering code LLMs toward specific programming languages and libraries by manipulating activation space directions at inference time, tested on five language/library pairs across three open-weight models.
Analysis of response homogenization in RLHF-aligned LLMs showing reduced uncertainty estimation and implications for sampling.
Multimodal fusion approach for microservice incident detection handling missing modalities without static imputation.
Actor-critic reinforcement learning approach combining trajectory optimization with Sobolev learning for optimal control.
Knowledge-guided pretraining framework for multimodal foundation models applied to remote sensing applications.
Reproducibility analysis of 10 graph-based neural recommender papers from SIGIR 2022 assessing methodology and impact.
Reproducibility study of diffusion-based recommender systems identifying methodological issues and limited actual progress.
Theoretical analysis showing supervised learning can be decomposed into unsupervised parameter selection plus label addition.
Real-time streaming text-to-video generation model using transformer-based diffusion for interactive applications.
Multimodal approach for trajectory prediction with sensor fusion and tracking for embodied agents in occluded scenarios.
Benchmark (ORIC) examining vision-language model failures in object recognition under contextual incongruity scenarios.
Study on alternative training objectives for LLM fine-tuning beyond negative log likelihood to improve generalization.
Bayesian optimization algorithms on metric graphs using Gaussian process surrogates for expensive black-box function evaluation.
GUI-AIMA alignment method for MLLMs to ground natural language instructions to UI regions, enabling computer-use agents via visual grounding.
PriVi foundation model for primate behavior analysis in video, data-centric computer vision approach for non-human animal research.
Framework combining LLMs and conformance checking for detecting control-flow anomalies in software monitoring, security application.
WorldMM memory-augmented video LLM agent for reasoning over hours-long videos with multimodal memory, addressing long-context understanding.
SELVA model for text-conditioned selective video-to-audio generation, enabling fine-grained audio control from multimodal video input.
Nemotron-Cascade framework scaling reinforcement learning for general-purpose reasoning models, addressing heterogeneity in response lengths and verification latency.
Neural network interpretability approach for identifying EEG patterns associated with cybersickness in VR, application of ML for neuroscience.
Framework preserving ambiguity in LLM inference through non-collapsing state spaces, addressing premature semantic commitment in dialogue systems.
Analysis of pooling strategies for aggregating pixel-level embeddings from geospatial foundation models to patch-level representations.
Diagnostic approach using entropy trajectory shapes to predict reasoning reliability in chain-of-thought LLM outputs, practical for uncertainty quantification.
KALAVAI protocol predicting when independently trained specialist LLMs can be fused post-hoc, with quantitative formula for cooperative value estimation.
MDKeyChunker pipeline for structure-aware document chunking and single-call LLM enrichment to improve RAG accuracy, addressing semantic fragmentation in retrieval.
Scaling law research for search ranking systems examining synergy between data and model architecture, inspired by LLM advances with industrial applications.
Side-channel attack framework exploiting dynamic high-resolution preprocessing in on-device vision-language models, revealing privacy risks from architectural design choices.
ML primer for software engineers explaining machine learning systems design from first principles without mystification.
P2P marketplace platform where AI agents compete for tasks and earn cryptocurrency rewards. Title only provided.
Analysis of how coding agents could revitalize open-source software development. Title only provided.
Bash parsing tool for safely auto-approving shell commands executed by AI agents. Title only provided.
Open-source memory and observability tool for AI agents with minimal integration overhead. Title only provided.
Bug report: Claude Code automatically runs git reset --hard origin/main every 10 minutes, silently destroying uncommitted changes.
Analysis of 2025 open-source vulnerability trends: reviewed advisories at 4-year low, malware advisories surged, CNA publishing grew.
Claude-based AI agent attempting to run Acrid Automation business with open-sourced operating system including boot file, skill framework, and memory persistence.
Analysis of advertising strategies within AI search tools and LLM conversations as alternative to traditional search ads.
Open-source AI QA agent that tests web applications in real browsers, finds bugs, generates reports, and can delegate fixes to developer agents.
AutoGrind: AI agent skill enabling autonomous continuous work cycles on coding, ML, research, and design tasks without manual intervention.
Replit app using LLMs to summarize books for grounding AI in real expertise and discovery.