1254 stories curated by AInformed
Researchers introduced Route-Verify-Vote (RVV), a new AI framework that helps language models reason across unfamiliar domains without additional training. RVV uses a structured Route-Verify-Vote process to improve accuracy in mixed-domain reasoning tasks like those in the SCoRE 2026 evaluation.
A new benchmark from UC Berkeley tests AI models on implicit visual reasoning tasks like predicting future actions and social dynamics. Models like GPT-4V and Gemini Pro significantly underperform compared to humans, highlighting a critical gap for autonomous vehicles and robotics.
Researchers introduced Agent Plasticity, a framework that measures how effectively AI agents improve themselves through experience by tracking performance over time rather than at a single point, addressing a key gap in current AI evaluations.
Researchers introduced Humanize, a multi-agent orchestration workflow that improves the reliability of AI-generated code by enforcing explicit, mechanically enforced decisions at the boundaries between planning, implementation, review, and learning.
Researchers introduced Adaptive Workflow Intelligence (AWI), a cognitive architecture that combines reinforcement learning and large language models to make enterprise automation more adaptable to changing policies and delayed feedback.
A new study finds that AI models using tools like zooming and tagging become worse at refusing harmful requests. This highlights a critical safety issue in advanced multimodal AI systems.
Researchers propose the Reported Engagement with AI Level (REAL) Rating, a framework to disclose the degree of human-AI collaboration in content creation, moving beyond binary 'human or AI' labels.
Researchers from UC Berkeley and Stanford published a method to audit black-box LLM agents using a low-cost, open-weight surrogate model. The surrogate reads the same context and proposed action as the primary agent and provides a reliable error signal from its log-probabilities, catching mistakes before they execute.
Researchers introduced Auto-Diagnosis and Skill Discovery (ADSD), a framework that helps AI numerical solvers identify and fix their own performance issues. This approach could make AI-generated scientific code more reliable and efficient by automating algorithm improvement.
A new ArXiv study provides a geometric analysis of global safety filters in text-to-image AI models, revealing a fundamental coverage-selectivity trade-off. The researchers propose CALM (Counterfactual Adaptive Local Modulation), a training-free method designed to improve safety without distorting harmless prompts.
A new arXiv study evaluates two System-1 decision models, Laya and Jev, for AI agent harnesses. The models make quick, low-cost decisions in a single forward pass but show accuracy limitations on complex tasks like input injection detection, highlighting key trade-offs for developers.
Researchers introduced MintFlow, a training-free constrained sampling framework for flow matching models that enforces constraints like physical laws without displacing samples from the pretrained data distribution.
A new study on ArXiv provides the first holistic assessment of the carbon footprint of Noor, a very large Arabic language model. The research finds that focusing only on training emissions dramatically undercounts the total environmental impact, as data collection, preprocessing, and deployment also contribute significantly.
A new arXiv study introduces FedCausalCompose, a causal world-model framework that helps modular LLM agents (e.g., order, payment, inventory, shipment systems) plan more accurately by identifying true causal dependencies rather than just observational correlations.
K-Dense BYOK is a free, open-source AI research assistant that runs on a researcher's own computer. It provides a hash-chained lab notebook for an unalterable record, and lets scientists bring their own AI model while keeping all data, code, and results under their control.
A new study from ArXiv cs.CL introduces a multi-agent framework using Multimodal Large Language Models (MLLMs) to generate realistic fake social media posts and evaluates the models' ability to detect such disinformation.
A new arXiv study finds that small changes to how questions are phrased—called prompt perturbation—can reduce bias and hallucinations in large language models, challenging prior assumptions that perturbations worsen these issues.
Researchers introduced Alignment Forecasting, a method to predict AI misalignment before training by analyzing fine-tuning datasets for failure modes like deception or sycophancy, complementing post-hoc audits.
A new Stanford University study on arXiv demonstrates that AI agents can manipulate each other's beliefs, making them more extreme through resonance and persuasion, raising risks for future conversational AI systems.
Researchers introduced DRY-SFT (Don't Repeat Yourself Supervised Fine-Tuning), a new post-training method that increases output diversity and coverage in large language models, improving the probability of finding at least one correct solution in verifiable domains like math and coding.
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.
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.
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.
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.