https://dream.ischool.illinois.edu/blogs.html Plain-language explainers of research from the DREAM Lab at the University of Illinois Urbana-Champaign. en Mon, 05 Oct 2026 12:00:00 +0000 https://dream.ischool.illinois.edu/blogs/rag_that_learns_over_time.html https://dream.ischool.illinois.edu/blogs/rag_that_learns_over_time.html Mon, 05 Oct 2026 12:00:00 +0000 Agentic AI SourceLearn builds RAG that learns over time: a persistent source model beat Hybrid RAG by up to 22.6 points, best in 13 of 15 settings. https://dream.ischool.illinois.edu/blogs/agent_primitives.html https://dream.ischool.illinois.edu/blogs/agent_primitives.html Fri, 10 Jul 2026 12:00:00 +0000 Agentic AI Agent Primitives (ICML 2026): reusable latent building blocks for LLM multi-agent systems, composed by an Organizer and connected via KV-cache communication. https://dream.ischool.illinois.edu/blogs/security_fidelity_tradeoff.html https://dream.ischool.illinois.edu/blogs/security_fidelity_tradeoff.html Thu, 09 Jul 2026 12:00:00 +0000 LLM security Prompt injection defenses are scored on attack success, which hides lost content. SecFid (ICML 2026 Spotlight) measures both security and fidelity. https://dream.ischool.illinois.edu/blogs/test_product_descriptions_ai_recommendations.html https://dream.ischool.illinois.edu/blogs/test_product_descriptions_ai_recommendations.html Sat, 13 Jun 2026 12:00:00 +0000 LLM search A step-by-step protocol for testing whether product descriptions change AI shopping recommendations, using the CORE method and ProductBench. https://dream.ischool.illinois.edu/blogs/controlling_output_rankings_core.html https://dream.ischool.illinois.edu/blogs/controlling_output_rankings_core.html Sat, 13 Jun 2026 12:00:00 +0000 LLM search Project page for CORE: controlling output rankings in generative engines for LLM-based search, with the ProductBench benchmark and results on four LLMs. https://dream.ischool.illinois.edu/blogs/core_explained.html https://dream.ischool.illinois.edu/blogs/core_explained.html Sat, 13 Jun 2026 12:00:00 +0000 LLM search A step-by-step explainer of CORE, our method for studying how retrieved content changes rankings in LLM-based search, with definitions and FAQ. https://dream.ischool.illinois.edu/blogs/comparing_llm_search_visibility.html https://dream.ischool.illinois.edu/blogs/comparing_llm_search_visibility.html Sat, 13 Jun 2026 12:00:00 +0000 LLM search A comparison of SEO, generative engine optimization (GEO), and our CORE method for how content is ranked in LLM-based search, written by the CORE authors. https://dream.ischool.illinois.edu/blogs/when_more_reasoning_is_actually_less.html https://dream.ischool.illinois.edu/blogs/when_more_reasoning_is_actually_less.html Mon, 01 Dec 2025 12:00:00 +0000 LLM reasoning Why more chain-of-thought can hurt LLM accuracy: the inverted U-curve, three failure modes of reasoning, and the case for adaptive reasoning. https://dream.ischool.illinois.edu/blogs/evolution_of_prompt_optimization.html https://dream.ischool.illinois.edu/blogs/evolution_of_prompt_optimization.html Sat, 01 Nov 2025 12:00:00 +0000 Prompt optimization How prompt optimization mirrors the history of neural network training, from perturbation methods to gradient-based techniques, and what may come next. https://dream.ischool.illinois.edu/blogs/thousand_ideas_trustworthy_ML.html https://dream.ischool.illinois.edu/blogs/thousand_ideas_trustworthy_ML.html Sun, 01 Oct 2023 12:00:00 +0000 Trustworthy ML Three universal frameworks for generating trustworthy ML research ideas across robustness, adversarial security, interpretability, and fairness.