Research map / Large language models / Reasoning and chain of thought

Wait, Wait, Wait... Why Do Reasoning Models Loop?

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies why reasoning models loop, attributing looping to errors in learning (hardness-driven risk aversion and temporally correlated errors) and showing temperature reduces looping but is only a stopgap, using open reasoning models and a synthetic star-graph task.

Problem

Reasoning models often get stuck repeating text in chains of thought, especially under greedy decoding or low temperature; it is unclear why this happens and whether temperature addresses the root cause.

Approach

Attributes looping to systematic errors in learning (imperfect fit of the teacher/training distribution). Shows, via a proposition and a synthetic random-walk star-graph task, that hard-to-learn progress actions spread probability mass so easy cyclic actions win under greedy decoding, and that Transformers have an inductive bias toward temporally correlated errors that produce loops. Temperature r

Key results

• Looping at low temperature is common: Qwen-1.5B loops in 76% of responses vs 37% for Qwen-32B at temperature 0 on AIME; OpenThinker3-1.5B loops 30% vs 4% for teacher QwQ-32B.
• On the star-graph task at temperature 1 with exploration, the learned model's response length is 108.1 vs 24.4 for a perfect learner (about 4.4x).
• Temporal correlation: consecutive root visits share the top-predicted child 98.7% of the time under greedy decoding.

Surprising findings

• Phi-4 reasoning models loop more on easy than hard AIME problems, often repeating answer-presentation phrases after finding the answer.
• Adding a tiny fraction (0.001) of looping traces in training makes the reset probability rise toward 1 as the loop persists.

Limitations

• Synthetic star-graph task is simple and may not map to real reasoning models.
• Looping mechanisms in real models are only supported by qualitative examples and a single prefix analysis.
• Phi-4 family deviates from the hard-problem pattern and is only preliminarily explained.

Experimental setup

Datasets: AIME 2024 AIME 2025 GPQA Diamond

Benchmarks: AIME 2024 AIME 2025 GPQA Diamond

Metrics: looping fraction response length accuracy

Baselines: greedy decoding perfect learner repetition penalty

Builds on: star graph hardness (Bachmann and Nagarajan 2024) neural text degeneration distillation of reasoning traces

Closest papers by meaning

• Reasoning with Latent Thoughts: On the Power of Looped Transformers (ICLR 2025 · similarity 0.89)
• On the Inductive Bias of Stacking Towards Improving Reasoning (NeurIPS 2024 · similarity 0.88)
• Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision (NeurIPS 2026 · similarity 0.87)
• Prioritize the Process, Not Just the Outcome: Rewarding Latent Thought Trajectories Improves Reasoning in Looped Language Models (ICML 2026 · similarity 0.87)
• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.87)
• Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning (NeurIPS 2025 · similarity 0.87)