[2605.07984] Where's the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions

[2605.07984] Where's the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions

arXiv - AI 3 min read

About this article

Abstract page for arXiv paper 2605.07984: Where's the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions

Computer Science > Machine Learning arXiv:2605.07984 (cs) [Submitted on 8 May 2026] Title:Where's the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions Authors:Nicole Ma, Nick Rui View a PDF of the paper titled Where's the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions, by Nicole Ma and 1 other authors View PDF HTML (experimental) Abstract:We study planning site formation in language models -- where internal representations of structurally-constrained future tokens form during the forward pass, and whether they causally drive generation. Using rhyming-couplet completion as a clean test of forward-looking constraint, we apply two lightweight methods (linear probing and activation patching) across Qwen3, Gemma-3, and Llama-3 at more than ten scales. Probing shows that future-rhyme information is linearly decodable at the line boundary, with signal that strengthens with scale in all three families. Activation patching reveals that only Gemma-3-27B causally relies on this encoding, exhibiting a handoff in which the causal driver migrates from the rhyme word to the line boundary around layer 30. Every other model we test conditions on the rhyme word throughout generation, with near-zero causal effect at the line boundary despite strong probe signal. We localize the Gemma-3-27B handoff to five attention heads through two-stage path patching that recover ~90% of the rhyme-routing capacity at ...

Originally published on May 11, 2026. Curated by AI News.

Related Articles

Researchers asked ChatGPT, Gemini and Claude which jobs are most exposed to AI. The chatbots wildly diagree
Llms

Researchers asked ChatGPT, Gemini and Claude which jobs are most exposed to AI. The chatbots wildly diagree

A study reveals that AI models disagree on which jobs are most vulnerable to automation, highlighting the unreliability of AI-generated e...

AI Tools & Products · 4 min ·
I stopped treating ChatGPT like Google — and everything suddenly clicked
Llms

I stopped treating ChatGPT like Google — and everything suddenly clicked

I stopped using ChatGPT like Google and started treating it like a thinking partner — here’s why that simple shift made the AI dramatical...

AI Tools & Products · 8 min ·
Hackers abuse Google ads, Claude.ai chats to push Mac malware
Llms

Hackers abuse Google ads, Claude.ai chats to push Mac malware

AI Tools & Products · 6 min ·
Llms

Does Claude dream of electric gavels? A federal case with Kansas connections sets an AI precedent.

AI Tools & Products ·
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