SymLang framework uses language-guided program synthesis with symmetry constraints to discover governing equations from noisy experimental data.
Analysis of fairness constraints in ML systems showing when enforcing fairness leads to worse outcomes for affected groups.
Self-evolving NLP extraction system that adapts to domain-specific taxonomies and emerging terminology in specialized fields like medical and legal.
Privacy analysis of DNA foundation model embeddings showing inversion attacks can reconstruct genomic sequences from shared embeddings.
Theoretical analysis of post-training linear autoregressive models with policy gradients, proving convergence properties under margin conditions.
Graph-based reinforcement learning framework for power distribution network resilience incorporating topological features for outage management.
Method for discovering interpretable audio attributes using multimodal LLMs for low-resource audio classification with high reliability.
Analysis of safety alignment in small language models using weak supervision and Self-MOA approach for reducing human annotation costs while maintaining usefulness.
Token caching optimization for vision-language navigation models accounting for visual and semantic dynamics to reduce inference cost.
Token merging optimization for Segment Anything Model to improve inference speed while preserving segmentation quality and handling SAM's attention architecture.
Agricultural vision competition focused on data-centric AI and model generalization under real-world distribution shifts rather than model design alone.
Research on generalization of RL-trained vision-language mobile agents for GUI automation. Addresses lack of standardized benchmarks and open-source RL systems for interactive task learning.
Method for preserving LLM safety alignment during fine-tuning by constraining safety-critical tokens, addressing alignment drift.
Knowledge-grounded NL2SQL system handling heterogeneous SQL dialects with semantic correctness and dialect-specific syntax compliance.
Transformer variant decomposing residual stream into token and context components for interpretable language modeling.
Framework for auditable fine-tuning and inference of proprietary LLMs on cloud platforms with cryptographic verification.
Monograph on probabilistic inference and learning theory using Stein's method with applications to variational gradient descent.
Lightweight on-device adaptation framework for speech enhancement models addressing dynamic acoustic scene changes with frozen backbone.
World model learning approach using symmetry exploration to capture physical invariances and conservation laws for extrapolative generalization.
Tool for verifying and explaining RL policies for multi-bridge network maintenance with formal safety guarantees and interpretability.
Generative-reconstructive-discriminative network with ROI attention for industrial surface defect detection and localization.
Re-evaluation of LiRA membership inference attacks under realistic assumptions, questioning prior effectiveness claims with realistic threat models.
Systematic comparison of four training objectives (cross-entropy, prototype, triplet, AP loss) for out-of-distribution detection in image classification.
Security analysis of large vision-language models vulnerable to semantic slot filling attacks that elicit unsafe outputs.
Hierarchical multi-agent system for Kubernetes autoscaling addressing resource waste through coordinated pod and node scaling policies.
arXiv paper on Staged Multi-Agent Training (SMAT) for co-adaptive exoskeleton control, using curriculum learning to mirror human motor adaptation.
arXiv paper on silicon photonics acceleration for diffusion model inference, targeting energy efficiency of UNet and attention mechanisms.
arXiv paper on physics-based reinforcement learning for data-driven exoskeleton control using joint-moment prediction instead of lab-based inverse dynamics.
arXiv research evaluating synthetic data for baggage trolley detection in airport logistics systems.
arXiv research on federated learning with compression for non-convex optimization on heterogeneous distributed data.
arXiv research on ML-driven microarchitectural techniques addressing memory bottleneck in modern computing systems.
arXiv paper on scaling Mixture-of-Experts model training using Megatron Core, addressing systems challenges in sparse model architectures across memory, communication, and computation.
Report on Chinese AI companies distributing 8 billion yuan in coupons during Lunar New Year for agentic AI apps. Market analysis of agent deployment in China.
Postmortem of Tess.Design, AI image marketplace with artist royalties (50%). Launched May 2024, shut down January 2026 with learnings on ethical AI models.
Study evaluating 14 AI agents across 2 benchmarks on 12 metrics across 4 reliability dimensions. Finds recent capability gains yield only small improvements in actual reliability compared to accuracy scores.
Open source agent framework forking OpenAI's Symphony, using Claude Code for autonomous implementation of Linear board issues. AI agents with LLM integration.
Open source model-agnostic AI code review tool with full control over model choice and costs. Alternative to Claude Code Review.
Andrej Karpathy thought piece on autonomous AI agents conducting frontier research across compute clusters. Speculative/fictional framing of agentic research systems.
Security research on model artifact integrity during local LLM inference in llama.cpp. Creates llm-inference-tampering project targeting inference-layer attacks.
Framework for defining requirements and specifications for AI systems beyond testing/evals. Addresses gap between eval scores and actual user satisfaction in AI products.
PUG: tool that converts messy API documentation into structured CLI tools and MCP servers using LLMs for AI agents.
TLAi+ Benchmarks: dataset and benchmark suite for evaluating LLMs on TLA+ formal specification tasks with diverse problem types.
Autonoma: AI agents that automatically generate test suites and find bugs by navigating applications without manual test scripts.
Rainy Updates: deterministic dependency review and upgrade tool for Node monorepos with CI/CD integration and automated fix PRs.
Essay on limitations of open weights models without open training data, discussing post-training challenges for trillion parameter models.
AI-powered technical interview prep tool simulating realistic interviewer interactions with WebRTC and Socket.io.
Nvidia planning to launch NemoClaw, an open-source AI agent platform for enterprise software companies to dispatch agents.
Plannotator: open source tool for manual code review and feedback loops for autonomous agents. OSS framework for agent improvement via human feedback.
CLI tool using Claude to analyze project codebases and generate customized Claude Code configurations. Integrates with Claude CLI for code-specific setup.
Analysis of SRAM-centric AI accelerators (Cerebras, Groq, d-Matrix) vs GPUs for inference, focusing on near-compute vs far-compute memory tradeoffs.