Library

Knowledge Base

Browse our curated collection of insights, research, and practical frameworks.

Your Agent Didn't Fail at the Tool Call. It Failed at the Plan.

Sun et al. (2026) introduce Agent Planning Benchmark: 4,209 multimodal cases across 12 models. GPT-5 holistic plan correctness is 74.5% vs GPT-4o's 19.5%. Extra tools drop GPT-5 to 67.0% and spike tool-use errors from 2.3% to 15.8%. Constraint violation remains 16%+ even at the frontier.

The Lost in the Middle Problem: Why LLMs Ignore What You Put in the Centre of Your Prompt

The Lost in the Middle Problem: Why LLMs Ignore What You Put in the Centre of Your Prompt

Liu et al. (2023) demonstrated that language model performance on multi-document QA degrades significantly when relevant information appears in the middle of a long context, even when models nominally support that context length. This 'lost in the middle' effect has direct, actionable consequences for how you structure RAG outputs, system prompts, and few-shot examples.

About LLM Wisdom

LLM Wisdom is a curated knowledge library dedicated to the rapidly evolving world of large language models. We collect, distill, and present the most valuable insights, research, and practical guidance from across the AI landscape — so you don't have to spend hours scouring forums, papers, and social feeds for signal through the noise.

This library is built for practitioners and curious minds alike: engineers shipping LLM-powered products, researchers tracking the frontier, founders exploring new applications, and anyone who wants to build a genuine, working understanding of how these models behave, what they're capable of, and where the field is heading. Whether you're looking for a deep study on retrieval-augmented generation, a reusable prompting framework, or an expert interview that surfaces hard-won lessons from the trenches, you'll find it organized and accessible here.

Every piece is sourced thoughtfully. We surface original work from practitioners and researchers — always linking back to the author and their original post — and rewrite dense or scattered material into clear, structured editorial pieces that respect the source while making the knowledge easier to absorb and act on.

LLM Wisdom is built and maintained by a small team of AI practitioners who felt the gap between the explosion of LLM content online and the lack of a calm, trustworthy place to learn from it. We're independent and reader-supported, and we're always looking for new voices to feature. If you publish work in this space — or know someone who does — we'd love to hear from you.

Contact Us

Have a question, a piece of work you'd like us to feature, or just want to say hello? We'd love to hear from you.

Email: hello@llmwisdom.com

Social: Twitter | LinkedIn | GitHub