AI Research

🥹 0. TL;DR

Large language models write fluent explanations even when they’re wrong. Verifying their reasoning usually requires another LLM slow, expensive, and circular.

We needed something different:

A miniature reasoning critic <50 KB trained on synthetic reasoning mistakes, able to instantly detect broken reasoning in much larger models.

The Tiny Critic:

• trains on GSM8K-style reasoning traces generated by DeepSeek or Mistral
• uses FrontierLens, and Visual Policy Maps (VPMs) to convert reasoning into canonical numerical features
• is just a logistic regression with ~30 parameters
• runs in microseconds
• plugs into any agent
• dramatically improves InitAgent, R1-Loops, and research-planning stability

This post tells the full story how we built it, why it works, and what we learned about the shape of reasoning.

How a self-evolving AI learns to reflect, score, and rewrite its own reasoning

🧪 Summary

What if an AI could think not just solve problems, but reevaluate its beliefs in the face of new information?

In this post, we introduce a system that does exactly that. At the core of our pipeline is a lightweight scoring model called MR.Q, responsible for evaluating ideas and choosing the best ones. But when it encounters a new domain, a new goal, or a shift in task format, it doesn’t freeze it adapts.

📖 Summary

Imagine drowning in a sea of research papers, each holding a fragment of the knowledge you need for your next breakthrough. How does an AI system, striving for self-improvement, navigate this information overload to find precisely what it needs? This is the core challenge our Document Intelligence pipeline addresses, transforming chaotic documents into organized, searchable knowledge.

In this post we combine insights from Paper2Poster: Towards Multimodal Poster Automation from Scientific Papers and Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training to build an AI document profiler that transforms unstructured papers into structured, searchable knowledge graphs.

🧪 Summary

“What if AI systems could learn how to improve themselves not just at the level of weights or prompts, but at the level of strategy itself? In this post, we show how to build such a system, powered by symbolic rules and reflection.

The paper Symbolic Agents: Symbolic Learning Enables Self-Evolving Agents introduces a framework where symbolic rules guide, evaluate, and evolve agent behavior.

Summary

Chain-of-thought is powerful, but which chain? Short explanations work for easy tasks, long reflections help on hard ones, and code sometimes beats them both. What if your model could adaptively pick the best strategy, per task, and improve as it learns?

The Adaptive Reasoning Model (ARM) is a framework for teaching language models how to choose the right reasoning format direct answers, chain-of-thoughts, or code depending on the task. It works by evaluating responses, scoring them based on rarity, conciseness, and difficulty alignment, and then updating model behavior over time.

🔧 Summary

The General Reasoner paper shows how we can train LLMs to reason across domains using diverse data and a generative verifier. In this post, I walk through our open-source implementation showing how we built a modular reasoning agent capable of generating multiple hypotheses, evaluating them with an LLM-based judge, and selecting the best answer.

---

🧠 What We Built

We built a GeneralReasonerAgent that:

• Dynamically generates multiple hypotheses using different reasoning strategies (e.g., cot, debate, verify_then_answer, etc.)
• Evaluates each pair of hypotheses using either a local LLM judge or our custom MR.Q evaluator
• Classifies the winning hypothesis using rubric dimensions
• Logs structured results to a PostgreSQL-backed system

All of this was integrated with our existing stephanie framework, which includes: