Danny Ayers 🌳

What if I told you that your AI agent isn't "stupid"—it’s just working without a map? 🗺️ We’ve all seen it: You give an LLM a complex task. It starts strong. But by step 7, it’s hallucinating actions and ignoring rules you gave it 30 seconds ago. The fix isn't better reasoning. It's better representation. 🧵👇 2/ Here is the invisible problem with current prompting (even Chain-of-Thought): It asks the model to solve a physics problem without drawing a free-body diagram first. The AI is trying to hold the entire state of the world—variables, rules, constraints—in its "head" (latent space) while simultaneously trying to solve the puzzle. 3/ Cognitive science calls this a "Mental Model." When humans solve hard problems, we don't just guess. Engineers sketch blueprints. 📐 Programmers write pseudo-code. 💻 Scientists define variables. 🧪 We stabilize the problem before we attempt the solution. 4/ A fascinating new paper, "Model-First Reasoning LLM Agents" (Kumar & Rana, 2025), just dropped. Their hypothesis? Hallucination is a symptom of reasoning over an unstable internal model. If the foundation is shaky, the logic collapses. 5/ So, they propose a radical (but simple) shift: Model-First Reasoning (MFR). It splits the job into two distinct phases: Phase 1: 🛑 The Architect. Build the model. List entities, states, and hard constraints. Do not solve. Phase 2: ✅ The Builder. Now execute the plan using only the model from Phase 1. 6/ Think of it like building LEGOs. The "Chain-of-Thought" approach is grabbing a handful of bricks and guessing where they go. The "MFR" approach is opening the instruction manual, sorting the pieces by color, and then starting step 1. 7/ The Experiment: They tested this on "constraint-heavy" tasks—like medical scheduling or complex logistics. The results? 📉 Drastic drop in rule violations. 🚫 Fewer "implicit assumptions" (guessing). 🧠 Plans that actually make sense from start to finish. 8/ Why does this work? It’s called "Soft Symbolic Grounding." By forcing the LLM to write down the rules first, you turn a fuzzy language task into a structured logic task. If the agent tries to cheat later, the model acts as a guardrail. "Sorry, you can't move that. Your model says it's heavy." 9/ (I know, it sounds obvious now that you see it. But how many of our prompts actually do this?) We usually say: "Solve this step-by-step." We should be saying: "Define the rules of the game. Then play it." 10/ 🛠️ The Toolkit: Try this Prompt Today Don't just take my word for it. Next time you have a complex reasoning task, inject this block at the start: "Before solving, explicitly construct a model of the problem. List:Relevant Entities State Variables (what changes?) Actions (what is allowed?) Constraints (what is forbidden?) Once defined, generate the plan strictly adhering to this model." 11/ The Big Takeaway: We've been obsessed with the process of reasoning (CoT, ReAct). We forgot about the object of reasoning. Structure precedes solution. If you want reliable agents, stop letting them improvise the world. Make them build it first. 🏗️ 12/ This paper is a brilliant bridge between classical AI (which loved explicit models) and modern GenAI (which loves vibes). Read the full paper: "2512.14474v1.pdf" I’m curious: Have you tried separating "setup" from "solution" in your prompts? What happened? Let's discuss below! 👇

What if I told you an AI's 'mind' isn't a chaotic mess of code, but has a hidden 3D map of meaning that almost perfectly mirrors our own psychology? Would you believe me? A new paper just dropped, and it completely rewired my brain about how AI works. 🤯 🧵 A thread... You know that frustrating moment when you're trying to steer an AI? You ask it to be more "creative," and it suddenly starts making factual errors. These weird side-effects, or "off-target effects," have always been treated as random bugs in the system. Except they're not random at all. Researchers dug into the embedding matrices of LLMs (the part of the model that gives meaning to words) to map out the 'geometry' of concepts. They defined directions for 28 different semantic scales, like kind-cruel, strong-weak, or beautiful-ugly. And here's the first bombshell: The way the AI organized words along these scales was wildly similar to how humans do. An AI's internal 'rating' of a word like "peace" on a "good-bad" scale almost perfectly matched data from human surveys. The AI has human-like intuition. But that's not even the interesting part. They took all 28 of these distinct conceptual directions and used PCA to see if there was a deeper structure. And something weird happened. All that complexity collapsed. It turns out, you can explain most of the meaning with just THREE core dimensions: Evaluation (good vs. bad) Potency (strong vs. weak) Activity (active vs. passive) If this sounds familiar, it's because it's the exact same 3-factor structure psychologists discovered in human language 70 years ago! (I know, right? My mind was blown too.) Here's where it all clicks. This is the key to the whole mystery. Concepts aren't stored in neat little boxes. In our minds, and in the AI's, concepts that are related are geometrically close. The direction for 'kind' is naturally aligned with 'beautiful.' 'Strong' points in a similar direction to 'big.' They are entangled meaningfully. And this solves the puzzle of "random" side effects! When you nudge a word in the AI's mind to be more 'beautiful,' you're also dragging it along the 'kind' axis because their vectors are aligned. The side effect isn't random—it's predictable geometry. The size of the effect is proportional to the angle between the concepts. Now, you might think the solution is to force these concepts apart to get more control. That's what some previous methods tried to do. But this research says that's a terrible idea. It's like trying to teach a child that 'soft' has nothing to do with 'gentle.' You don't gain control; you just break their understanding of the world. The entanglement is a feature, not a bug. Here’s what this means for you: Next time you see an AI give a weirdly 'off' response, don't just think 'bug.' You're getting a glimpse of its hidden conceptual map. That "off" association is a clue to how it connects ideas. You're seeing the geometric relationships in its mind in real-time. This research completely changes how I think about AI. It’s not some alien intelligence operating on unknowable logic. It's a mirror, reflecting the deep, hidden, and sometimes weird structures of our own collective language and thought. Makes you wonder what other human patterns are hiding in the code... If you want your mind blown too, go read the paper. Incredible work by Austin C. Kozlowski, Callin Dai, & Andrei Boutyline. Paper: "Semantic Structure in Large Language Model Embeddings" arXiv:2508.10003v1