New Method Extracts AI Models’ Reasoning Traces, Fuels Training Debate

In one line: Researchers demonstrated a way to pull internal "reasoning traces" from major LLMs — and used them to raise questions about the training origins of some Chinese AI.

Key points

• Researchers devised a method to extract so-called reasoning traces from Claude, GPT, and Gemini.
• These traces offer a window into the internal processing a model runs through before producing an answer.
• Based on their analysis, the researchers say some Chinese AI models may have been trained on leading US models — a claim, not a confirmed finding.

Why it matters

Tools that observe a model's inner workings sit at the heart of AI interpretability research. Such techniques can also serve to trace a rival model's training lineage, adding fresh evidence to ongoing debates over model distillation and intellectual property.

Read more

• A New Trick Reveals AI Models’ Inner Thoughts — Wired