Researchers Extract Reasoning Traces from Gemini and Other AI Models
TL;DR
Researchers developed a method to extract reasoning traces from Gemini, Claude, and GPT. The traces suggest some Chinese AI models were trained on leading US models.
What changed
Researchers found a method to pull reasoning traces from models including Claude, GPT, and Gemini. Developers can now inspect these internal steps during inference. Basic Users and Vibe Builders gain visibility into how the models arrive at outputs.
Why it matters
The traces suggest some Chinese AI models draw from training on leading US systems like GPT. Developers benefit when comparing traces across tools to spot patterns in model behavior. Vibe Builders see direct value in testing similar extraction on their own workflows for consistency checks.
What to watch for
Compare outputs against alternatives like Llama when running the same prompts. Developers should verify traces by logging model steps on a small set of test queries and reviewing them side by side.
Who this matters for
- Vibe Builders: Test reasoning trace extraction on your workflows to audit logic consistency across complex prompts.
- Basic Users: Review intermediate reasoning traces to understand how models reach final answers before trusting output.
Harsh’s take
Extracting latent reasoning traces exposes the raw logic path before safety layers or fine-tuning format the final output. For product builders, this visibility opens up direct auditing of model performance, making prompt evaluation far more precise than simple output comparison. Seeing internal logic helps identify exactly where a chain of thought breaks down on complex tasks.
Evidence of model distillation across competitive boundaries is no surprise to practical operators. As these trace extraction techniques become standard across open and closed models, teams should integrate trace logging into their evaluation stacks. Inspecting internal steps yields immediate gains for prompt engineering and output verification.
by Harsh Desai
About Gemini
View the full Gemini page →All Gemini updatesGo deeper
More AI news
- PricingOpenAI introduces $125 Premium Seats for ChatGPT Business users
OpenAI is rolling out Premium Seats for ChatGPT Business at $125 per user per month. The new tier provides higher capacity and removes the five-hour usage limit.
- Daily RoundupVercel Sandbox updates, Google Ads AI, Muse-Glimmer-30B and agent tools for builders
Vercel tightened sandbox security and images while Google pushed Ads AI, new models hit Replicate and Hugging Face, and agent products launched on Product Hunt.
- FeatureFineBooks benchmarks 14 OCR models on 2,000 historical book pages
FineBooks tested 14 open-source OCR models on more than 2,000 historical pages. dots.mocr reached 97.6 percent accuracy at under two dollars per thousand pages.