Technical Summary: Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models

1. Problem Definition

Large Reasoning Models (LRMs) have demonstrated that allocating additional test-time compute to generate explicit intermediate reasoning traces (Chain-of-Thought, or CoT) improves performance on complex tasks. However, the prevailing assumption that "longer reasoning is consistently better" remains under-examined. While recent evidence suggests models can "overthink," existing literature largely treats this as an efficiency problem—aiming to reduce redundant computation while preserving the accuracy of full-length traces.

This paper argues that overthinking is also a reliability problem. A model may reach the correct answer early in its generation process but, upon continuing to reason, destabilize that correct trajectory, leading to a final incorrect prediction. The authors distinguish between two types of overthinking:

• Verbose Overthinking: The model continues reasoning beyond the point of correctness but maintains the correct final answer (redundant but harmless).
• Harmful Overthinking: The model reaches a correct intermediate state but subsequently deviates from it due to continued reasoning, resulting in an incorrect final prediction (destabilizing and harmful).

The core research question is: Once a model has reached the correct answer, does further reasoning refine the solution, or does it deviate from it?

2. Methodology

2.1 Formalizing Reasoning Sufficiency

The authors introduce a formal definition of question difficulty based on the minimum reasoning budget required for a specific model to generate the correct answer. Unlike prior work that uses average trace length as a proxy for difficulty (which is confounded by verbosity), this approach isolates the computation strictly necessary for correctness.

• Notation: Let $t = (u_1, \dots, u_N)$ be a reasoning trace.
• First Correct Index ($\tau_y$): The minimum index $i$ such that the prefix $t_{\le i}$ yields the correct answer $y$.
• Empirical Difficulty ($\hat{\kappa}$): The computational budget associated with $\tau_y$.

2.2 Prefix-Level Trajectory Evaluation Protocol

To measure the dynamics of reasoning after correctness is achieved, the authors propose a prefix-level evaluation protocol:

1. Forced Termination: For every prefix $t_{\le i}$ of a generated trace, a fixed termination template is appended to force the model to output a prediction at that intermediate step.
2. Extraction: An external extractor model ($A$) parses the partial trace to determine the predicted answer $\hat{y}_i$.
3. Classification:
   • Verbose: $A(t_{\le \tau_y}) = y$ and $A(t_{\le N}) = y$.
   • Harmful: $A(t_{\le \tau_y}) = y$ and $A(t_{\le N}) \neq y$.
4. Metrics: The Harmful Overthinking Rate ($H$) is calculated as the fraction of solved instances where the model reaches a correct prefix but fails at the final step.

2.3 Experimental Setup

• Models: Evaluated on multimodal LRMs (MM-Eureka, R1-VL, ThinkLite-VL, VL-Rethinker) and language-only LRMs (Qwen3, InternS1).
• Benchmarks:
  • Multimodal: AI2D, MathVerse, MathVision, MathVista, MMStar, VMCBench.
  • Language-only: AIME2025, GPQA.
• Strategies Compared:
  • No-CoT: Immediate answer without reasoning.
  • Actual Length: Default unconstrained CoT behavior.
  • Optimal Length: An oracle strategy that stops at the first correct prefix (used to quantify the potential gain from eliminating harmful overthinking).
  • Early Stopping: Adaptive inference stopping when predictions remain unchanged for $K$ steps.

3. Key Results

3.1 Reasoning Length vs. Difficulty

• Surprising Efficiency: Many instances considered "reasoning-intensive" require surprisingly little reasoning. Optimal stopping points are often concentrated near the beginning of the trajectory (frequently at zero utterances).
• Performance Gap: Stopping at the first correct prefix (Optimal Length) significantly outperforms the model's default behavior (Actual Length).
  • On multimodal benchmarks, Optimal Length improves accuracy by up to 21% (e.g., +23.3% on MathVision for R1-VL).
  • This gain exceeds the improvement gained from reasoning-oriented post-training over standard instruction-tuned models.
• Takeaway: Current LRMs are limited not only by their ability to reason but by their inability to stop at the right time.

3.2 Harmful Overthinking Dynamics

• Non-Monotonicity: Reasoning trajectories are not monotonic. After reaching a correct state ($z_i=1$), the probability of remaining correct drops rapidly as additional steps are generated, plateauing around 0.2 after ~100 steps.
• Answer Format Sensitivity: Harmful overthinking is more prevalent in free-form generation (24% rate) compared to multiple-choice (11% rate). The unconstrained output space in free-form tasks makes it easier for models to drift from a correct answer via unsupported revisions or reinterpretations.
• Ineffectiveness of Simple Early Stopping:
  • Standard early stopping (e.g., stopping when the answer stabilizes for $K$ steps) reduces average reasoning length but fails to mitigate harmful overthinking.
  • In some cases, aggressive early stopping increases the harmful overthinking rate (e.g., from 10.1% to 18.4% in one setting), suggesting that reducing verbosity does not automatically stabilize the reasoning trajectory.

3.3 Generalization to Language-Only Models

The phenomenon is not specific to multimodal reasoning. Language-only models on math-heavy benchmarks (AIME2025, GPQA) exhibit similar patterns:

• Default reasoning improves over No-CoT, but Optimal Length yields massive gains (e.g., Qwen3 on AIME2025: 58.3% $\to$ 91.7%).
• This confirms that harmful overthinking is a broader reliability risk inherent to the reasoning process itself, not just a byproduct of visual grounding.

3.4 Failure Analysis

An analysis of 4,842 harmful traces reveals the primary causes of correctness deviation:

• Logical Drift: The dominant failure mode on abstract reasoning benchmarks (e.g., >50% on MathVerse). Models introduce contradictions, irrelevant detours, or unjustified answer revisions.
• Visual Reinterpretation: The dominant failure mode on perception-heavy benchmarks (e.g., >60% on MathVision/AI2D). Models misread, invent, or over-interpret visual evidence after initially solving the problem correctly.
• Calculation Errors: Account for a small fraction (<10%) of harmful overthinking, indicating that arithmetic mistakes are not the primary driver of instability.

4. Key Contributions

1. Formalization: Defined overthinking via reasoning sufficiency, disentangling verbose (redundant) overthinking from harmful (destabilizing) overthinking.
2. Protocol: Introduced a prefix-level trajectory evaluation protocol to measure reasoning sufficiency and correctness instability dynamically.
3. Quantification: Demonstrated that harmful overthinking is systematic across multimodal and language-only benchmarks, often causing models to drift from early correct answers to incorrect final predictions.
4. Categorization: Identified that correctness deviations are driven primarily by logical drift and visual reinterpretation, rather than arithmetic limitations.

5. Significance and Claims

The paper challenges the monolithic view of test-time scaling as strictly beneficial. It posits that allocating the right amount of reasoning is often more critical than simply enabling reasoning.

• Reliability vs. Efficiency: The authors argue that overthinking is a reliability issue, not just an efficiency one. Current strategies focused on reducing verbosity (efficiency) do not address the core instability of the reasoning process.
• The "Stop" Problem: The results suggest that a major limitation of current LRMs is the inability to recognize when a solution is sufficient. The "Optimal Length" strategy (an oracle) shows that significant performance gains are possible if models could simply stop reasoning once the correct answer is found.
• Future Direction: The work calls for future research to focus not only on making models reason more but on helping them understand when reasoning is sufficient, thereby preventing the destabilization of correct trajectories.

The paper concludes that harmful overthinking is a fundamental reliability risk that requires specific mitigation strategies beyond simple early stopping or length reduction.