/Tech4h ago

Google's Zorik Gekhman and Jonathan Herzig find chain-of-thought reasoning helps LLMs retrieve stored factual knowledge

Story Overview

Google researchers Zorik Gekhman and Jonathan Herzig show that flipping on chain-of-thought reasoning lets models pull more facts straight from their weights, even on straightforward single-hop questions. The trick works across Gemini variants and Qwen3-32B, lifting pass@k scores on SimpleQA-Verified and EntityQuestions without any weight changes.

41561718011.9K

#218

Original post

Rishabh Agarwal@agarwl_#218inTech

Nice blog studying why reasoning helps LLMs recover the knowledge stored in their weights.

https://research.google/blog/thinking-to-recall-how-reasoning-unlocks-parametric-knowledge-in-llms/

Thomas Dimson@turtlesoupy

@eliebakouch @agarwl_ I've noticed that recently too! Okay @agarwl_ it looks like I CAN repro consistently and it isn't noise but reasoning-off. Reasoning makes it find you. Screenshot from my session. Interesting result; reasoning helps discover knowledge?

1:21 PM · Jun 28, 2026 · 11.9K Views

Developer Impact

Reasoning tokens create extra runway

Even dummy phrases like repeated "Let me think" give models a computational buffer that improves recall, though real factual priming during reasoning recovers most of the gains.

Open Question

Hallucinations in the trace raise final risk

Any fabricated intermediate fact sharply increases the chance the final answer is wrong, so filtering for clean reasoning paths becomes a practical lever at inference time.

Sentiment

Sentiment building, check back later.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.