MedicalPatchNet: self-explainable architecture for chest X-ray classification using patch-based independent classification and aggregation for transparency.
PeruMedQA benchmark dataset of Peruvian medical exam questions in Spanish. Evaluates LLM performance on non-English, Latin American medical domain tasks.
Investigates sim-to-real gap in visual navigation by comparing simulator-trained and real-world-trained policies. Demonstrates simulator policies can match real-world performance.
Preference-aligned audio captioning framework using RLHF with CLAP-based reward model trained on human-labeled preferences. Addresses gap between supervised learning and real preferences.
DivEye detector for AI-generated text using diversity metrics. Improves detection of synthetic text while providing interpretability over black-box classifiers.
Theoretical analysis proving minimax convergence rates for learning pairwise interactions in single-layer attention models. Rate independent of embedding dimension and token count.
Automated data generation framework for multi-step bimanual mobile manipulation tasks. Uses imitation learning to reduce costly human teleoperation data collection.
Framework for generating multimodal datasets with controllable mutual information between modalities. Enables systematic study of MI estimators and multimodal self-supervised learning.
Largest multilingual scaling laws study with 774 experiments across 400+ languages. Introduces Adaptive Transfer Scaling Law (ATLAS) for monolingual and multilingual pretraining.
Sequential transducer model for recommendation systems handling ultra-long user histories. Explores memorization with transformer-like architecture at scale.
Likelihood-free inference approach adapting domain support for stochastic systems. Addresses misspecified support in robotics agent deployment scenarios.
Cosmos-Predict2.5 foundation model for world simulation unifying text/image/video generation. Leverages vision-language model for grounded physical AI predictions.
Theoretical proof that optimal learning rates transfer across widths in MLPs with μP parameterization. Shows learning rate converges to nonzero constant at infinite width.
LLM agent for symbolic equation discovery using multi-step scientific reasoning. Guides model through hypothesis formation, data analysis, and equation validation.
Evaluates HiFloat low-bit formats on Ascend NPUs for LLM inference. Compares INT8 and 4-bit floating-point for efficiency-accuracy tradeoffs.
Evolutionary System Prompt Learning method jointly improves LLM contexts and weights via reinforcement learning. Enables autonomous self-improvement for agentic systems.
Multi-agent LLM framework for robotic manipulation with closed-loop visual feedback. Integrates language and vision models for task planning in dynamic environments.
LLM agents diagnose and repair infeasible supply chain optimization models. Demonstrates closed-loop agent task decomposition for operations research problems.
Research on prompt injection vulnerabilities in LLM agents via skill files. Identifies security risks in agent supply chains and skill-based attacks.
Economic analysis of AGI's impact on labor and growth. Argues human verification becomes the bottleneck as AI decouples cognition from biology.
Multi-agent Claude orchestrator for GitHub/Linear issues using LanceDB embeddings and code knowledge graphs. Real workflow integration.
Personal project using LLMs to reverse-engineer MP3 player firmware. Limited technical depth provided.
Enterprise autonomous agent framework for workflow automation. Lacks implementation details and technical depth.
Self-hosted open-source API gateway supporting multiple LLM providers with per-token limits, statistics, and OpenAI compatibility.
Developer tool using repo history to recommend code reviewers with awareness of availability and expertise.
Discussion of UI design patterns optimized for LLM-native workflows post-foundational models.
Open-source LLM inference engine optimizing memory efficiency and cold starts for serverless deployments.
Minimal post about hedge fund experiment staffed by AI employees using paper money. No technical details.
Bloomfilter service allows AI agents to register ICANN domains via single API call using on-chain payments. MCP server compatible.
Study analyzing gender and social stereotypes in Spanish-language LLMs using 4,156 test questions from Latin American researchers.
WebMCP Core tool converts websites into Model Context Protocol definitions for AI agents. Open source CLI with playground and A/B testing.
Anthropic reportedly weakens AI safety principles amid competitive pressure. News article about company policy shift.
Production-grade open-source agent operating system written in Rust with 137K LOC, 14 crates, comprehensive testing.
Skills and MCP servers for Claude Code to generate videos programmatically using Remotion and FFmpeg.
Comparative analysis of web framework token efficiency for AI agent code generation.
Rails engine for building and monitoring LLM agents in production with cost tracking, retries, circuit breakers, and observability dashboard.
Self-hosted observability server exposing logs, database, and metrics as 75 MCP tools. Works with Claude Code and Cursor.
Analysis of multi-agent workflow failures, identifying three engineering patterns for reliable agent systems. Technical guidance on agent design.
Open-source bundle of 8 MCP servers for homelab services (Proxmox, Grafana, Ollama, etc). 40 tools total, Python implementation.
Nkmc virtual filesystem allows AI agents to call APIs using standard Unix commands (ls, cat, grep). Minimal details provided.
Analysis of how tool use and notation reduce task complexity for LLMs rather than increasing model capability. Examines agent design patterns.
Developer built LLM comment detector for HackerNews after being flagged for excessive AI-assisted posting. Personal experience account.
Deff tool streamlines review of AI-generated code changes. Surfaces diffs with vim motion support for faster comprehension.
Clerk invoicing app built with AI agents in 7 days. Uses natural language chat for invoice generation and PDF parsing.
Claude Code MCP integration for stateless GPU provisioning across cloud providers with conversational control and cost optimization.
Edictum is a runtime governance library for LLM agents that enforces safety contracts at tool-call boundaries. Tested on 6 frontier models across 17,420 interactions, identifying a 'GAP' where models refuse harmful text requests but execute them via tool calls.
Tldraw moves test suite to closed source to prevent AI-assisted reimplementation of open source libraries. Discusses implications for open source projects with commercial models.
Unworldly is a tamper-proof audit trail system for AI agents with real-time behavior monitoring, file/shell command tracking, and HIPAA/ISO 42001 compliance. Records and replays agent sessions.
Tesseract is a 3D architecture editor desktop app with built-in MCP server for AI-assisted code visualization. Enables Claude integration to display codebase analysis visually rather than in text.
Opinion on documentation quality for both AI agents and humans. Argues against segregating workflows between human and AI use, advocating unified documentation standards.