KAIST and Naver find written reasoning steps map to distinct internal patterns in LLMs
Researchers at South Korea's KAIST and Naver AI Lab report that large language models keep calculation, formula retrieval, and deduction as separable processes inside their internal states, not merely as rhetorical stages in visible chain-of-thought text. Testing Qwen2.5-7B, Qwen3-8B, and Gemma4-31B on math tasks, the team found the clearest distinction in middle layers. Reporting on the work, The Decoder highlights a widening safety gap: monitors that read only written reasoning may miss hidden computation performed internally. Operators should not treat chain-of-thought transcripts as a complete audit trail. Important caveats remain unsettled, including whether these patterns can catch errors or steer generation mid-run, and whether focusing on middle layers overlooks reasoning dynamics in earlier or later layers.
KAIST and Naver find written reasoning steps map to distinct internal patterns in LLMs
Researchers at South Korea's KAIST and Naver AI Lab report that calculation, formula retrieval, and deduction remain separable in model internal states, with the strongest signal in middle layers. The study tested Qwen2.5-7B, Qwen3-8B, and Gemma4-31B on math tasks, though whether these findings can catch errors or steer generation mid-run remains open.
Key takeaway
Visible chain-of-thought is a lossy view of internal reasoning, so safety monitoring cannot rely on written steps alone.
What happened
According to The Decoder, researchers at KAIST and Naver AI Lab found that calculation, formula retrieval, and deduction remain separable in model internal states, with the strongest signal in middle layers.
The study evaluated Qwen2.5-7B, Qwen3-8B, and Gemma4-31B on math tasks. The Decoder reports the work shows models process more than their visible chain of thought reveals, though whether the patterns can catch errors or steer generation mid-run remains open.
Evidence
KAIST and Naver AI Lab researchers report calculation, formula retrieval, and deduction remain separable in model internal states, with the strongest signal in middle layers.
The Decoder · attributed
Researchers at South Korea's KAIST and Naver AI Lab report that calculation, formula retrieval, and deduction remain separable in model internal states, with the strongest signal in middle layers.
The study tested Qwen2.5-7B, Qwen3-8B, and Gemma4-31B on math tasks.
The Decoder · attributed
The study tested Qwen2.5-7B, Qwen3-8B, and Gemma4-31B on math tasks, though whether these findings can catch errors or steer generation mid-run remains open.
Reasoning steps like calculation, formula retrieval, and deduction are clearly separable in a model's internal states, especially in the middle layers.
The Decoder · attributed
Reasoning steps like calculation, formula retrieval, and deduction are clearly separable in a model's internal states, especially in the middle layers.
Models process more than their visible chain of thought reveals, creating a gap between observable outputs and internal computation.
The Decoder · attributed
That matters for AI safety, because models process more than their visible chain of thought reveals.
Why it matters
Safety researchers and model operators may need interpretability tools that monitor internal states rather than relying solely on visible chain-of-thought outputs to detect hidden reasoning.
Limits and uncertainties
Whether these internal patterns can catch errors or steer generation mid-run remains open.
The study's focus on middle layers may not capture reasoning patterns in early or late layers.
Practical implications
Do not treat chain-of-thought transcripts as a complete safety audit trail for math-heavy agents.
Evaluate whether internal-state monitoring, especially in middle layers, can complement visible reasoning logs in production guardrails.
What to watch
Research testing whether separable internal patterns can catch errors or steer generation mid-run.
Follow-up work on reasoning signals in early and late layers beyond the three evaluated model families.
