New Research Reveals Why AI Reasoning Models Overthink: The Termination Circuit
Qwen3-1.7B knows the answer 70% earlier but keeps thinking due to a localized verification gate.
Researcher Chandram Dutta found that in Qwen3-1.7B, the answer to GSM8K problems is often settled around 30% of the chain of thought, yet the model continues thinking. The decision to stop (emission of the </think> token) is driven by a localized set of MLP layers near the end of the network — layer 27 contributes about 40% of the logit, and the top 8 components (mostly MLPs) account for 71%. This small set of MLPs, called the termination circuit, acts as a verification gate: it stops only when the written answer matches the internally computed value. Causal interventions show that splicing the model's own correct answer sentence into the middle of its CoT triggers </think> as the greedy next token 94% of the time, while wrong answers or the right answer at the wrong time fail. Ablating just the top 4 MLPs in this circuit (at sentence boundaries) made the model unable to stop. The article also notes that no single direction could reliably fire the gate early; the best direction recovered only about 20% of the effect from patching the full internal state.
- In Qwen3-1.7B, the median answer to GSM8K problems is settled at ~30% of the chain of thought; the remaining 70% is overthinking.
- The decision to emit </think> is driven by a localized set of MLP layers (layer 27 contributes ~40% of the logit), not attention heads.
- The termination circuit acts as a verification gate: splicing the correct answer sentence into the CoT triggers stopping 94% of the time, but wrong values or timing fail; ablating these MLPs prevents stopping entirely.
Why It Matters
Understanding the termination circuit could enable more efficient reasoning models by allowing early exit without sacrificing accuracy.