Pipis et al. (2025) — ARXIV-2512.12895
Key takeaways
- Open reasoning models frequently loop under greedy or low-temperature decoding, repeating the same span indefinitely. Temperature matters; larger models loop less; and distilled students loop significantly even when the teacher they were distilled from rarely does.
- The pattern points at the mismatch between the training distribution and what the model actually learned — errors in learning — rather than at decoding alone.
- A synthetic graph-reasoning task isolates two mechanisms. The first is risk aversion from hardness of learning: when the correct progress-making step is hard to learn and an easy cyclic step is available, the model puts relatively more mass on the cycle and gets stuck. That is the one LIT-119 invokes.
Standing in the anthology
The explanation LIT-119 gives for SOTA-127: a 90M model shown chain-of-thought traces beyond its capacity learned the cheaper pattern, repetition, and stopped producing function calls. Filed for the mechanism; its own remedies are not yet drawn into a practice.