---
format: "aidr-story-markdown/v1"
id: "1163c5850fafa6ab7ef50342b9e700bc88e192986a51212ada6ca15d68ac791b"
canonical_url: "https://aidr.today/1163c585?lang=en"
title: "AI Agents Boost Algorithmic Reasoning to 84% by Splicing Code Into Model Architectures"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-07T02:44:26.000Z"
category: "Research"
topics: ["deepseek","reasoning","agent"]
source_urls: []
summary: "A new AI training process allows coding agents to modify a neural network's structure in real-time, moving the resulting models beyond the design of standard transformers. By splicing code layers directly into the network during training, researchers saw algorithmic reasoning accuracy for nanochat rise from 55% to 84%, while training loss fell below the levels seen in ordinary training. The method also enables open-weight models to perform self-modifications, such as DeepSeek-V4.1-Flash, which analyzed its own activations to identify hidden states signaling incorrect answers and wrote a new code layer to predict these failures, increasing its AIME and HMMT 2026 scores from 64% to 71%."
---

# AI Agents Boost Algorithmic Reasoning to 84% by Splicing Code Into Model Architectures

> [Open the canonical story](<https://aidr.today/1163c585?lang=en>)

**Published:** 2026-11-07T02:44:26.000Z
**Category:** Research
**Topics:** deepseek, reasoning, agent

## Summary

A new AI training process allows coding agents to modify a neural network's structure in real\-time, moving the resulting models beyond the design of standard transformers\. By splicing code layers directly into the network during training, researchers saw algorithmic reasoning accuracy for nanochat rise from 55% to 84%, while training loss fell below the levels seen in ordinary training\. The method also enables open\-weight models to perform self\-modifications, such as DeepSeek\-V4\.1\-Flash, which analyzed its own activations to identify hidden states signaling incorrect answers and wrote a new code layer to predict these failures, increasing its AIME and HMMT 2026 scores from 64% to 71%\.

## Sources

_No safe source URL was recorded._