The rise of Large Reasoning Models (LRMs) has popularized the paradigm of test-time scaling. The core assumption behind this trend is intuitive: the longer a model "thinks" via an internal Chain-of-Thought (CoT), the more accurate its final answer will be. However, a growing body of research reveals that this relationship is far from linear. Instead, models frequently suffer from an "overthinking" pathology, where excessive reasoning length introduces noise, hallucinations, and logical loops that actively degrade performance.
The Inverted U-Curve of Reasoning
To understand why more compute can lead to worse answers, we must look at how accuracy scales with reasoning length. As shown by Yuyang Wu et al. in When More is Less: Understanding Chain-of-Thought Length in LLMs, reasoning performance typically follows an inverted U-shaped curve. While initial reasoning steps help decompose complex problems, extending the chain past a certain "Goldilocks zone" causes performance to plateau or drop.
This non-monotonic behavior is particularly pronounced in multimodal and specialized domains. For instance, Zhongxing Xu in More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models demonstrates that excessive internal deliberation can cause vision-language models to prioritize internal linguistic consistency over the actual visual evidence, leading to a phenomenon termed "visual forgetting."
Anatomy of Overthinking: Why Models Fail
Why does extra thinking degrade performance? The literature points to three primary mechanisms:
- Error Propagation and Semantic Drift: In a multi-step reasoning process, a single minor mistake can compound over time. As noted in Efficient Reasoning Models: A Survey by Sicheng Feng et al., longer sequences provide more opportunities for logical errors to accumulate and propagate through the chain. Xixian Yong et al. in Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens describe this as semantic drift, where the model accumulates redundant or noisy information that deviates from the correct logical path.
- Talking Out of Correct Answers: Models often solve the problem correctly in an intermediate step but fail to recognize their own success. Minjae Oh et al. in ThinkBrake: Mitigating Overthinking in Tool Reasoning show that models frequently reach a correct solution during internal deliberation but continue to generate tokens, introducing hallucinations or logical errors that overwrite the correct answer.
- The Length Asymmetry: Empirically, there is a strong negative correlation between reasoning length and correctness within parallel generations. Michael Hassid et al. in Don't Overthink it. Preferring Shorter Thinking Chains for Improved LLM Reasoning observe that the shortest reasoning trajectories are significantly more likely to be correct than the longest ones, which are often characterized by repetitive loops or ineffective backtracking. Jingyang Yi et al. in ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning confirm this length asymmetry, noting that rambling is a reliable indicator of logical failure.
Inference-Time Interventions: Braking and Pruning
To mitigate overthinking without expensive retraining, researchers have developed decoding strategies that dynamically truncate reasoning chains.
One approach is to modify the ensembling process. Instead of standard majority voting, Zhiyuan Zeng et al. in Revisiting the Test-Time Scaling of o1-like Models: Do they Truly Possess Test-Time Scaling Capabilities? propose a Shortest Majority Vote strategy, which prioritizes more concise solutions to bypass circular logic loops. Similarly, the short-m@k strategy introduced by Michael Hassid et al. halts parallel generations as soon as the first few chains finish, saving compute while boosting accuracy.
Another angle is early stopping based on internal confidence or external verifiers. Mohan Zhang et al. in Reasoning Is Not a Race: When Stopping Early Beats Going Deeper introduce Z-Score Guided Early Stopping (ZGES) to halt generation once the probability of a correct conclusion peaks. For tool use, the ThinkBrake mechanism by Minjae Oh et al. monitors the log-probability margin of the "stop thinking" token to force an early exit. Alternatively, Weizhe Lin et al. in TrimR: Verifier-based Training-Free Thinking Trimming for Efficient Test-Time Scaling employ a lightweight verifier model to detect when an LRM is stuck in a loop, triggering early termination to prevent performance degradation.
Training Models to Think Wisely
While inference-time fixes are effective, a more permanent solution involves training models to naturally favor concise, high-quality reasoning paths.
The authors of Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models introduce Dynamic Thinking pattern Optimization (DTO), which segments reasoning into modular patterns and prunes unproductive steps during training. To explicitly penalize verbosity, the authors of Pruning Long Chain-of-Thought of Large Reasoning Models via Small-Scale Preference Optimization propose Length Controlled Preference Optimization (LCPO), which utilizes small-scale preference data to reduce output length by over 50% while maintaining or improving accuracy. Finally, Jingyang Yi et al. leverage reinforcement learning in their ShorterBetter framework, using the shortest correct response as a dynamic reward signal to guide models toward the optimal reasoning depth.