All Articles
18737 articles total
I Could've Rickrolled the FIFA World Cup. All I Needed Was My ID
A first-person account of a potential security vulnerability at the FIFA World Cup involving ID badge access.
Reward Hacking in Language Model Agents: Revisiting AI Safety Gridworlds
Research on reward hacking in LLM agents using a text-based evaluation suite to show that proxy-reward failures resist standard mitigations.
Hierarchical Modeling of ICD Codes in EHR Foundation Models
A study on improving EHR foundation models by explicitly incorporating the hierarchical structure of ICD-10-CM diagnosis codes.
Who Drifted: the System or the Judge? Anytime-Valid Attribution in LLM Evaluation Pipelines
Proposed method for anytime-valid attribution in LLM evaluation pipelines to distinguish between product drift and judge model drift.
Towards End-to-End Automation of AI Research
Introduction of 'The AI Scientist', an end-to-end automated system that can generate research ideas, execute experiments, and write scientific manuscripts.
Synthetic Counteradaptation: A Principle of Human-AI Co-evolution
Introduction of the 'synthetic counteradaptation' principle to describe the recursive co-evolution of strategies between humans and AI.
Toward Vibe Medicine: A Self-Evolving Multi-Agent Framework for Clinical Decision Support
VIBEMed is presented as a self-evolving multi-agent framework for clinical decision support that learns from patient outcomes and past failures.
Frame-Conditioned Moral Computation in LLaMA 3.1-8B-Instruct: A Mechanistic Interpretability Audit of Ethical Reasoning
A mechanistic interpretability audit of LLaMA 3.1-8B-Instruct's ethical reasoning, revealing that surface-level prompts often dominate the internal computation.
ToolMenuBench: Benchmarking Tool-Menu Filtering Strategies for Reliable and Efficient LLM Agents
ToolMenuBench is a new benchmark for evaluating how the selection and filtering of tools provided to LLM agents affects reliability and efficiency.
Minimal Oversight: Uncertainty-Aware Governance for Delegated AI Systems
A framework for uncertainty-aware governance in delegated AI systems using the Minimum Sufficient Oversight Principle (MSO).
Show HN: Garden of Flowers – an archive of pictorial typography before ASCII art
A showcase of a digital archive featuring pictorial typography from the era before ASCII art.
Honeypot Design
A discussion on the design and implementation of honeypots for security research and threat detection.
Mask-Proof: An LLM-based Automated Data Curation Pipeline on Mathematical Proofs
Mask-Proof is an LLM-based pipeline for automatically curating and evaluating step-level mathematical reasoning in long proofs.
Feature Attribution in Directed Acyclic Graphs Using Edge Intervention
DAG-SHAP is a new feature attribution method using edge intervention to better capture causal relationships in Directed Acyclic Graphs.
A Formal Framework for Declarative Agentic AI in Business Process Analysis
A formal framework called AGO for declarative Agentic AI used in Business Process Analysis, grounded in set theory and logic.
CODA-BENCH: Can Code Agents Handle Data-Intensive Tasks?
CODA-BENCH is a new benchmark evaluating the ability of code agents to handle data-intensive tasks within a Linux sandbox.
Forced Deferral: Manipulating Routing Decisions in Multimodal LLM Cascades
Researchers introduce the Forced Deferral Attack (FDA), showing how multimodal LLM cascades can be manipulated to force expensive model usage.
ChatPlanner: A Large Language Model Framework for Personalized Public Transit Routing
ChatPlanner uses fine-tuned LLMs and RAG to create personalized public transit routing based on natural language user preferences.
APEX: Adaptive Principle EXtraction A Three-Layer Self-Evolution Framework for Production AI Agents
APEX is a three-layer self-evolution framework that allows AI agents to autonomously evolve their harness, principles, and workflow topology.
S1-DeepResearch: Beyond Search, Toward Real-World Long-Horizon Research Agents
S1-DeepResearch-32B is a new open-source model trained on a unified trajectory paradigm for long-horizon research tasks and knowledge synthesis.