'Don't Overthink It': a new survey maps the science of making reasoning models stop wasting tokens
A new arXiv survey systematizes the fast-growing literature on efficient R1-style reasoning models — techniques for cutting chain-of-thought length, compressing reasoning paths, and deciding when thinking harder actually helps. Overthinking has become a measurable tax, and a research field has formed around not paying it.
The survey names the problem the deployment world already pays for: reasoning models that think too much. R1-style models buy accuracy with tokens, but the marginal token is often worthless — models re-derive, second-guess, and pad long after the answer is determined. At production scale that surplus thinking is a direct compute cost and latency floor, which is why efficiency went from footnote to field.
The techniques catalogued share a theme: reasoning length should be a learned, input-dependent decision. Path compression, length-aware training objectives, adaptive early exit, and hybrid fast/slow routing all attack the same failure — treating every problem as if it deserved maximum deliberation. The evidence that fixed thinking budgets are systematically wrong is now overwhelming enough to survey.
It matters because efficient reasoning is where research meets margin. The same weeks this survey landed, an open model demonstrated that post-training alone could lift agentic performance past a larger sibling — capability per token, not capability per parameter, is the axis the field now optimizes. The overthinking literature is the theory of the price cuts everyone is living through.
arXiv — Don't Overthink It: a survey of efficient R1-style large reasoning models → · Sebastian Raschka — LLM research papers: the 2026 list →