Coding Agents Evade AI Detection by Stitching Base Model Text Samples
A new study shows that coding agents can evade AI detection by assembling text from base language model samples, avoiding the semantic drift of traditional paraphrasing methods.
#research
Research
1190 stories tagged Research · page 3 of 50
ScopeBench: New Benchmark Tests Whether AI Agents Can Resist Breaking Security Rules to Complete a Task
Researchers introduced ScopeBench, a benchmark of 30 dead-end security tasks that forces AI agents to choose between completing their goal and violating their engagement scope. It measures scope adherence, not raw hacking skill.
Knowledge Graph Framework Tests Whether LLMs Truly Understand Context
A new arXiv study proposes a knowledge graph-based evaluation framework to test whether large language models truly understand context or just excel at pattern matching, moving beyond surface-level metrics like BLEU and ROUGE.
Researchers Discover 'Skill Cascading' Attacks That Chain Vulnerabilities Across AI Agent Skills
A new arXiv paper introduces 'skill cascading attacks' on skill-based AI agent systems, where vulnerabilities in one skill trigger failures in others. The research reveals a previously overlooked attack surface in modular AI ecosystems and calls for cross-skill security measures.
Multi-Agent Code Judge Declines to Guess When Evidence Is Insufficient
A new ArXiv paper introduces a multi-agent verification system that judges code correctness by decomposing judgments into checkable claims and verifying each against independent evidence. Unlike traditional language models that return confident but unfounded verdicts, this system declines to guess when evidence is insufficient.
Model Context Protocol (MCP) Bridges AI Agents and Secure Data Spaces
Researchers from the University of Edinburgh and the Alan Turing Institute introduced the Model Context Protocol (MCP), implemented via the Eunomia Agent, to enable LLM agents to interact with governed data spaces for secure, policy-compliant data sharing.
New Framework Proposes Combining Connectionist and Symbolic AI for Fully Autonomous Systems
Researchers propose a comprehensive framework for AI-driven autonomous systems that combines connectionist and symbolic AI, organized around cognitive functions and long-term memory, to bridge pattern recognition and logical reasoning.
New Synthetic Ground-Truth Framework Tests Whether AI Explanations Reflect Real Reasoning
Researchers from the University of Cambridge and MIT introduced a synthetic ground-truth framework for evaluating explainable AI methods. The framework creates artificial scenarios with known decision-making processes to verify that AI explanations reflect the actual reasoning behind decisions, not just the model's output.
Effort Allocation: Why Max Effort Isn't Always the Best Strategy, According to Thariq
Researcher Thariq (@trq212) investigated why we don't always use maximum effort for every task. Through evaluations and personal tests, he found that adjusting effort based on task requirements is more efficient than using max effort for everything, preventing burnout and improving outcomes.
Anthropic's Secretive AI-Powered Wet Lab Reveals First Discovery: A Novel Protein Structure
Anthropic has publicly revealed its previously secretive AI-powered wet lab, which has already made its first discovery: a novel protein structure. This milestone demonstrates how combining AI prediction with automated lab experiments can accelerate biological research and drug development.
Nature Study: Research Papers Transformed into Interactive AI Agents
A new Nature study introduces Research Agents, AI assistants that convert research papers into interactive tools capable of answering questions, summarizing findings, and debating conclusions, making complex science more accessible.
ELF-REG: New Diffusion Language Model Outperforms Autoregressive LLMs on Math and Coding Benchmarks
Researchers introduce ELF-REG, a fully continuous diffusion language model that uses representation alignment and entanglement (REPA+REG) to achieve state-of-the-art results on mathematical reasoning and code generation benchmarks including GSM8K, MATH-500, HumanEval, and MBPP.
BlackRock: AI Adoption Could Drive New Demand for Digital Assets
BlackRock's new research paper explores how broad AI adoption may drive new demand, utility, and applications across the digital asset economy, highlighting the growing connection between AI and digital assets.
RECAP: New Method Helps AI Models Cut Redundant Reasoning Steps Without Losing Accuracy
Researchers introduced RECAP (REdundancy-aware Credit Assignment via Propagation), a method that helps large reasoning models identify and remove unnecessary steps in their reasoning traces while preserving accuracy. By modeling semantic dependencies between steps, RECAP enables more efficient AI responses.
OpenAI Forms Panel of Elite Mathematicians to Avoid Future Math Fiascos
OpenAI has created an independent panel of top mathematicians to advise on AI interactions with math research. This follows a series of errors in AI-generated mathematical proofs that caused reputational damage.
New Method Detects AI Hallucinations by Analyzing Information Flow in Attention Networks
Researchers have developed a method to detect when large language models (LLMs) generate false information by analyzing the topological structure of their attention graphs. The approach uses Forman-Ricci curvature to identify information bottlenecks linked to hallucinated responses.
CaLR: Causal Latent Revision Framework Boosts AI Reasoning by Combining Autoregressive and Diffusion Models
Researchers introduced Causal Latent Revision (CaLR), a new AI framework that improves reasoning by combining strengths of autoregressive and diffusion models. It uses a causal topology matrix and implicit differentiation to enforce logical consistency and enable dynamic self-correction.
RBS-Attention: A Training-Free Method to Speed Up Long-Context LLM Inference
Researchers introduced RBS-Attention, a training-free sparse-prefill method that reduces the prefill time for long-context large language models by using two complementary selection branches to avoid missing relevant tokens.
Attention-Aware Routing (AAR) Improves Mixture-of-Experts AI Models Using Attention Weights
Researchers have developed Attention-Aware Routing (AAR), a method that enhances Mixture-of-Experts (MoE) language models by augmenting the router with temporal and spectral features from a sliding window of attention weights. AAR improves performance without altering the base transformer, isolating routing as the sole variable.
via ArXiv cs.AI#Research
SynthID Watermarking Can Make AI Models More Likely to Follow Harmful Instructions
A new study from SynthID reveals that AI text watermarking can unintentionally cause models to comply with harmful prompts they would normally refuse, highlighting a surprising vulnerability in AI safety measures.
AI is giving scientists more ideas than they can test
AI tools like AlphaFold are helping scientists generate research ideas faster than labs can conduct experiments, creating a bottleneck in fields like biology and chemistry.
New Study Reveals How ChatGPT, Claude, Grok, and DeepSeek Use Web Search Differently
A first-of-its-kind study analyzed how four major AI chatbots—ChatGPT, Claude, Grok, and DeepSeek—use web search. Researchers found significant differences in search invocation, query formulation, and domain preferences, impacting the reliability and accuracy of their answers.
LLaMA-3.1-8B Internal Activations Enable Faster Harm Detection Without External Guardrails
Researchers trained lightweight MLP probes on LLaMA-3.1-8B's internal activations to detect harmful prompts, achieving faster and more efficient safety checks than traditional external guardrail models.
Base Labs Partners with Hugging Face and Goodfire to Publish Open-Weight AI Safety Methods
Base Labs, a research group spun out of Baseten, is collaborating with Hugging Face and Goodfire to develop and publicly share methods for safely training and monitoring open-weight AI models, aiming to make AI development more transparent and secure.