TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Ask HN: Is reasoning just some iterative context refinement?

2 pointsby sschmitt4 days ago
context = &quot;you are a useful model. you evaluate your own output for N steps after the initial user input&quot;<p>prompt = &quot;here comes a prompt by the user&quot;<p>context += prompt<p>for _ in range(N): context += evaluate_llm(context)

1 comment

Reubend4 days ago
Are you asking in terms of how the current crop of &quot;reasoning models&quot; are implemented, or are you asking more philosophically about the nature of true reasoning?<p>Calling it &quot;refinement&quot; is dismissive. It&#x27;s generating new information, which is in many cases well beyond the scope of the original prompt.<p>Reasoning models today are just a marketing spin on chain-of-thought techniques that benefit from reinforcement learning.