[2603.28371] Coherent Without Grounding, Grounded Without Success: Observability and Epistemic Failure

[2603.28371] Coherent Without Grounding, Grounded Without Success: Observability and Epistemic Failure

arXiv - AI 4 min read

About this article

Abstract page for arXiv paper 2603.28371: Coherent Without Grounding, Grounded Without Success: Observability and Epistemic Failure

Computer Science > Computers and Society arXiv:2603.28371 (cs) [Submitted on 30 Mar 2026] Title:Coherent Without Grounding, Grounded Without Success: Observability and Epistemic Failure Authors:Camilo Chacón Sartori View a PDF of the paper titled Coherent Without Grounding, Grounded Without Success: Observability and Epistemic Failure, by Camilo Chac\'on Sartori View PDF HTML (experimental) Abstract:When an agent can articulate why something works, we typically take this as evidence of genuine understanding. This presupposes that effective action and correct explanation covary, and that coherent explanation reliably signals both. I argue that this assumption fails for contemporary Large Language Models (LLMs). I introduce what I call the Bidirectional Coherence Paradox: competence and grounding not only dissociate but invert across epistemic conditions. In low-observability domains, LLMs often act successfully while misidentifying the mechanisms that produce their success. In high-observability domains, they frequently generate explanations that accurately track observable causal structure yet fail to translate those diagnoses into effective intervention. In both cases, explanatory coherence remains intact, obscuring the underlying dissociation. Drawing on experiments in compiler optimization and hyperparameter tuning, I develop the Epistemic Triangle, a model of how priors, signals, and domain knowledge interact under varying observability. The results suggest that neithe...

Originally published on March 31, 2026. Curated by AI News.

Related Articles

Llms

I think we’re about to have a new kind of “SEO”… and nobody is talking about it.

More people are asking ChatGPT things like: “what’s the best CRM?” “is this tool worth it?” “alternatives to X” And they just… trust the ...

Reddit - Artificial Intelligence · 1 min ·
Llms

Why would Claude give me the same response over and over and give others different replies?

I asked Claude to "generate me a random word" so I could do some word play. Then I asked it again in a new prompt window on desktop after...

Reddit - Artificial Intelligence · 1 min ·
Anthropic essentially bans OpenClaw from Claude by making subscribers pay extra | The Verge
Llms

Anthropic essentially bans OpenClaw from Claude by making subscribers pay extra | The Verge

The popular combination of OpenClaw and Claude Code is being severed now that Anthropic has announced it will start charging subscribers ...

The Verge - AI · 4 min ·
Llms

wtf bro did what? arc 3 2026

The Physarum Explorer is a high-speed, bio-inspired neural model designed specifically for ARC geometry. Here is the snapshot of its curr...

Reddit - Artificial Intelligence · 1 min ·
More in Llms: This Week Guide Trending

No comments

No comments yet. Be the first to comment!

Stay updated with AI News

Get the latest news, tools, and insights delivered to your inbox.

Daily or weekly digest • Unsubscribe anytime