---
title: "Discussion on \"The Grounding Gap: Why In-Context Reinforcement Learning Still Isn’t Learning\" | Hashnode"
description: "Discussion on \"The Grounding Gap: Why In-Context Reinforcement Learning Still Isn’t Learning\". The Growing Discrepancy Between AI Hype and Reality"
url: "https://hashnode.com/posts/the-grounding-gap-why-in-context-reinforcement-learning-still-isnt-learning/68f75f451d6a8895d864b45e"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [The Grounding Gap: Why In-Context Reinforcement Learning Still Isn’t Learning](https://ai-cosmos.hashnode.dev/the-grounding-gap-why-in-context-reinforcement-learning-still-isnt-learning?utm_source=hashnode&utm_medium=feed)

The Growing Discrepancy Between AI Hype and Reality

Oct 21, 2025 · 7 min read · ai-cosmos.hashnode.dev  
Read the article

-   [#icrl](https://hashnode.com/tag/icrl)
-   [#ai](https://hashnode.com/tag/ai)
-   [#ai-research](https://hashnode.com/tag/ai-research)
-   [#ai-myths](https://hashnode.com/tag/ai-myths)
-   [#ai-literacy](https://hashnode.com/tag/ai-literacy)

## Written by

[Gerard Sans](https://hashnode.com/@gerardsans)

@gerardsans · London, UK

  
[View profile](https://hashnode.com/@gerardsans)

AI/ML Google Developer Expert, AI/ML Cloud Champion, Angular GDE

I help developers succeed in Artificial Intelligence and Web3; Former AWS Amplify Developer Advocate. I am very excited about the future of the Web and JavaScript. Always happy Computer Science Engineer and humble Google Developer Expert. I love sharing my knowledge by speaking, training and writing about cool technologies. I love running communities and meetups such as Web3 London, GraphQL London, GraphQL San Francisco, mentoring students and giving back to the community.

## Responses 1

Markdown supported · ⌘↵ to post

[Zbigniew Łukasiak](https://hashnode.com/@zby)

[Oct 27, 2025](https://hashnode.com/posts/the-grounding-gap-why-in-context-reinforcement-learning-still-isnt-learning/68f75f451d6a8895d864b45e/comment/68ffdb27155cd47038b780b1)

I agree that In Context Learning is not really learning (in fact I would just call it reasoning by analogy).

But:

"In practice, they describe automation of prompt engineering: feeding outputs back as inputs, applying heuristic scoring functions, selecting better completions through sampling.

This is useful! It genuinely improves performance on many benchmarks. But it doesn’t transcend the pretraining distribution. The model is still generating next-token predictions, just with more elaborate scaffolding around the generation process."

How do you define the "pretraining distribution" that it is not changed by changing the generator? How it could not change but still give us better results?

[Gerard Sans](https://hashnode.com/@gerardsans)

[Dec 16, 2025](https://hashnode.com/posts/the-grounding-gap-why-in-context-reinforcement-learning-still-isnt-learning/68f75f451d6a8895d864b45e/comment/68ffdb27155cd47038b780b1#reply-6941a4fc16134634dc6dffd6)

Here’s the key: after pretraining, the model’s latent space is frozen. Post-training and prompting don't change this landscape; they change how we navigate it. Iterating over prompts, for ICL or any task, is the process of finding the optimal starting point and pathway through that fixed space. We're not changing the distribution's mass; we're just learning to traverse it more effectively to reach better results.

Related articles: \[1\] [ai-cosmos.hashnode.dev/understanding-ai-in-2025-i…](https://ai-cosmos.hashnode.dev/understanding-ai-in-2025-its-still-all-about-the-next-token) \[2\] [ai-cosmos.hashnode.dev/the-ai-trinity-what-everyo…](https://ai-cosmos.hashnode.dev/the-ai-trinity-what-everyone-gets-wrong-about-modern-ai-systems) \[3\] [ai-cosmos.hashnode.dev/the-attention-bottleneck-a…](https://ai-cosmos.hashnode.dev/the-attention-bottleneck-ai-failure-modes-explained)
