Source-linked AI summary
The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis
Zihao Wei, Liang Pang, Jiahao Liu, Wenjie Shi, Jingcheng Deng, Shicheng Xu, Zenghao Duan, Fei Sun, Huawei Shen, Xueqi Cheng
TL;DR
Longer reasoning trajectories can improve LLM performance but also cause overthinking, and fine-grained evidence about when reasoning becomes redundant remains limited. The paper analyzes length and semantic dynamics to identify an instance-specific RCP and uses termination-token rank monitoring for early exit. Across the reported reasoning-dynamics findings, post-RCP computation is characterized as redundant, motivating self-termination for more efficient test-time scaling.
Problem
Longer reasoning trajectories can yield diminishing returns or performance degradation, but how intermediate steps affect answers and when reasoning becomes redundant remains insufficiently explored.
Method
The paper analyzes reasoning through thinking-content compensation and semantic path convergence, then detects each instance's RCP online with RCPD by monitoring termination-token rank dynamics.
Results
The analysis identifies post-RCP redundancy as the primary cause of overthinking and finds that RCPD can detect the completion boundary for early truncation while maintaining accuracy.
Takeaways & Limitations
Reliable self-termination at the RCP provides a basis for efficient test-time scaling by avoiding computation after reasoning has converged.
Takeaways & Limitations
The evaluation focuses on Qwen3 and DeepSeek-R1 and does not establish how universal RCP-like transitions are across closed-source, tool-augmented, or multimodal models.
Abstract
from arXiv · showhide
Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., </think>). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.
1 Introduction
The paper examines why longer reasoning trajectories can produce diminishing returns by tracking how reasoning length and semantic representations evolve. It identifies an instance-dependent Reasoning Completion Point (RCP) and proposes online detection to truncate redundant computation.
- The study addresses insufficient fine-grained evidence about how intermediate reasoning steps affect final answers and when additional tokens become redundant.
- Thinking-content compensation initially makes generated content increasingly concise as reasoning expands, but later reasoning length grows while content remains unchanged, producing redundancy.
- Semantic trajectories move from active exploration to a stable neighborhood with repetitive oscillations, indicating that the solution hypothesis has stabilized.
- The RCP marks the end of compensation and onset of semantic convergence, separating necessary reasoning from post-convergence redundant computation on each instance.
- RCPD detects this latent boundary during decoding by monitoring termination-token rank dynamics, enabling early truncation without additional computational overhead while maintaining accuracy.
2 Reasoning Length Dynamics
The paper studies reasoning as coupled thinking and content phases using controlled stepwise truncation. It finds an early compensation regime and a late regime where further thinking expands without changing stabilized content.
- Dynamics Setup: Stepwise Truncation: Generation is modeled as P(y, t|x) = P(t|x)P(y|x, t), separating a thinking trajectory t that conditions content generation y.
- Dynamics Setup: Stepwise Truncation: The protocol truncates thinking after each sentence-level reasoning step, injects an end-of-thinking delimiter, and immediately measures resulting content and correctness.
- Observations: Early truncation forces the model to offload more reasoning into content: correct answers often require substantially longer content, whereas incorrect answers cluster at very short content.
- Observations: With sufficient thinking, content retracts to concise outputs and accuracy stabilizes; further ineffective reasoning is expressed primarily as longer thinking.
- Observations: After the RCP, content length remains in a narrow stable band even as the thinking budget increases by over 100 steps, showing that extra thinking no longer changes content volume.
3 Reasoning Semantic Dynamics
Reasoning Semantic Dynamics treats induced content as a trajectory in latent semantic space, revealing early active exploration followed by convergence into a stable neighborhood with repetitive oscillations. This convergence explains why additional thinking becomes largely redundant.
- 3.1 Dynamics Setup: Semantic Lens: The semantic trajectory is estimated by sampling multiple continuations per truncated thinking prefix, embedding them, and tracking their per-step centroid.A higher-dimensional PCA Gaussian proxy and terminal reference distribution support quantitative convergence analysis, while the 2D projection is used for visualization.
- Observation 3: Semantic Exploration: Semantic representations initially undergo rapid shifts, indicating active exploration of feasible solution paths.Large centroid jumps and widely varying Dglobal(k) show that induced content can change markedly between adjacent truncation steps.
- Observation 4: Reasoning Convergence: After the transition, semantic representations converge toward a localized neighborhood and exhibit repetitive oscillations.97.3% of post-transition trajectory points remain inside the 95% attraction basin, while the centroid largely stays within that region.
- Observation 4: Reasoning Convergence: Dglobal(k) declines toward a low plateau after the turning point, indicating that the induced content distribution approaches its terminal state.The non-zero residual floor reflects an entropy gap between the instantaneous state Qk and the aggregate convergence basin Q∞, consistent with dynamic equilibrium rather than singular collapse.
- 3.2 Observations: Once content semantics have converged, further thinking steps occur after the induced distribution has stabilized, making overthinking largely redundant.The semantic view therefore complements length statistics by identifying the mechanistic transition from active semantic change to semantic stationarity.
4 Reasoning Completion Point
The Reasoning Completion Point (RCP) marks the instance-specific boundary between active reasoning and converged reasoning. It is defined by simultaneous length stabilization and semantic convergence, after which additional computation is largely redundant.
- 4 Reasoning Completion Point: The RCP is the critical boundary separating Pre-RCP active reasoning from Post-RCP converged reasoning.Before the boundary, semantic paths evolve and content remains volatile; afterward, semantic state and content length stabilize, producing diminishing returns and possible redundant rationalization.
- 4 Reasoning Completion Point: The RCP is defined as the earliest truncation step k_RCP where content length stabilization and semantic convergence are simultaneously achieved.Content stabilization requires Δcontent(k) to remain below ε_c, while convergence requires Dglobal(k) ≤ ε_D.
- 4 Reasoning Completion Point: The thresholds ε_c and ε_D are determined from late-step tail fluctuations, making the RCP an instance-specific stopping criterion.Beyond k_RCP, additional thinking no longer materially alters the induced content’s length or conceptual essence, so mitigation becomes an online detection problem.
5 Online Overthinking Elimination
The paper turns offline RCP analysis into an online detector that monitors termination-token rank dynamics to stop redundant reasoning. Across reasoning benchmarks and models, RCPD reduces generation while preserving competitive accuracy and adapts better than fixed budgets.
- 5.1 Offline RCP Extraction: Answer emergence typically coincides with kRCP, while a sharp drop in </think> rank provides an online-accessible convergence signature.String-based answer detection is brittle because final answers vary in form, motivating the intrinsic rank proxy.
- 5.1 Offline RCP Extraction: RCPD uses offline kRCP supervision to fit interpretable rank-dynamics rules, then forces </think> online when the detector triggers.At each sentence boundary, it queries the next-token distribution and tracks the current termination-token rank with a short history.
- 5.3 Experimental Evaluation: RCPD is evaluated on AIME24, AIME25, and GPQA-D with Qwen3 and DeepSeek-R1 using token count, accuracy, and RPCR.Comparisons include Full Reasoning, Budget Force, No-Think, DEER, and the training-based S-GRPO baseline.
- 5.3 Experimental Evaluation: RCPD substantially reduces tokens versus full decoding while maintaining competitive performance, rivals S-GRPO without fine-tuning, and outperforms it on GPQA-D but not AIME24.These results indicate that post-RCP reasoning adds computation without accuracy gains.
- 5.3 Experimental Evaluation: Under equal computation, Budget Force degrades because static truncation can stop during Pre-RCP, whereas RCPD adapts to instance-specific complexity.Qualitative analyses associate Post-RCP reasoning with cognitive inertia, repeated rationalization, and redundant loops that RCPD truncates.
6 Related Work
Related work spans post-training, prompt-based, and early-exit approaches to efficient reasoning. The paper positions RCPD within overhead-free early exiting, contrasting intrinsic-state monitoring with methods that require extra training or static prompts.
- Related-work taxonomy: Efficient reasoning work is organized into post-training, prompt-based, and early-exit methods.This taxonomy follows and extends prior classifications of reasoning-efficiency approaches.
- Post-training Based Methods: Post-training methods reduce reasoning through variable-length CoT supervision, length-aware reinforcement learning, or latent-space reasoning.These approaches alter training or reasoning representations rather than detecting an instance-specific stopping point during ordinary decoding.
- Prompt-based Methods: Prompt-based methods encourage concise chains, but static prompts can fail to adapt to diverse and complex inputs, compromising depth and precision.Their effectiveness depends directly on prompt design.
- Comparison framing: Table 1 distinguishes inference-time methods above the dashed line from training-based methods below it and reports Tok, CR, and RPCR.S-GRPO is the training-based baseline, while missing results are marked with “-” and externally sourced results with S-GRPO* when applicable.
- Early Exit Methods: Early-exit methods either add monitors or training costs, or use intrinsic model states for overhead-free termination such as DEER.The paper aligns RCPD with the latter efficient paradigm.
7 Conclusion
The paper characterizes LLM reasoning as active reasoning followed by convergence, separated by an instance-specific RCP. It argues that detecting this boundary enables efficient self-termination and mitigates post-RCP overthinking.
- 7 Conclusion: Reasoning Dynamics comprise Thinking-Content Compensation and Semantic Path Convergence, separating active reasoning from a converged phase at the RCP.The conclusion identifies post-RCP redundancy as the primary cause of overthinking within the studied setting.
- 7 Conclusion: A simple detector can identify the RCP and support reliable self-termination for more efficient test-time scaling.The conclusion frames stopping at the convergence boundary as the practical implication of the analysis.
8 Limitations
The analysis and evaluation focus on Qwen3 and DeepSeek-R1 because these open-weight models expose the delimiters and decoding signals needed for RCPD. The broader universality of RCP-like transitions remains untested.
- 8 Limitations: The study primarily evaluates Qwen3 and DeepSeek-R1 across multiple sizes and benchmarks because they expose explicit thinking delimiters and intervention signals.This scope supports controlled RCP analysis but does not establish applicability across all model families.
- 8 Limitations: The paper does not run the full RCP-analysis suite on closed-source APIs, tool-augmented agents, or multimodal models.Whether RCP-like phase transitions are universal beyond these systems remains outside the paper’s scope.
9 Potential Risks
RCPD is presented as an efficiency-focused method with no inherent implementation or deployment risks identified in the supplied discussion.
- RCPD targets reasoning-process efficiency and reduced computational overhead.The discussion frames the method as mitigating overthinking and unnecessary resource consumption.
B Additional Evaluation Details
The evaluation compares RCPD with standard, no-thinking, static, trigger-based, and training-based early-exit strategies across AIME24, AIME25, and GPQA-D. An oracle-calibrated static baseline still cannot match RCPD’s instance-adaptive compute–accuracy trade-off, while temporal consistency improves stopping robustness.
- Evaluation Protocol: RCPD is evaluated on AIME24, AIME25, and GPQA-D, containing 30, 30, and 198 problems, respectively.
- Baselines: The comparison includes Full, No-Think, Budget Force, DEER, and S-GRPO as representative inference-time or training-based strategies.Budget Force statically truncates reasoning; DEER uses reflective cues and confidence thresholds; S-GRPO is training-based.
- Baseline Calibration: An oracle iso-computational calibration gives Budget Force a globally tuned budget matching RCPD’s average token usage, despite that budget being unavailable at deployment.
- Static Budgets to Adaptive Stopping: Accuracy improves only within a narrow saturation region while token usage keeps growing, so one fixed budget cannot reliably stop near saturation across instances.RCPD instead uses online convergence signals for a more consistent compute–accuracy trade-off.
- Stopping Robustness: RCPD outperforms the single-threshold </think>-5 rule by using short-horizon temporal rank patterns that reduce sensitivity to noise.The comparison treats delimiter rank as a meaningful convergence proxy while showing that temporal consistency improves robustness.
C Implementation Details of RCPD Stopping Criteria
RCPD stops decoding when termination-token rank dynamics indicate saturation, progressive convergence, or a sustained low-rank plateau. Its semantic analyses and supplementary examples support post-RCP stability, while counterfactual tests reveal that post-RCP reasoning often rationalizes injected answers rather than genuinely re-verifying them.
- Stopping Criteria: RCPD triggers termination when the </think> rank satisfies any symbolic stopping condition distilled from probe-derived decision boundaries.The final stop signal is the disjunction of the defined conditions.
- Immediate Saturation: Immediate saturation uses Rt ≤5 as evidence that reasoning has collapsed into a definitive conclusion.
- Progressive Convergence: Progressive convergence detects a monotonic ladder pattern in descending rank values, representing directed convergence toward the solution.
- Sustained Plateau: Sustained plateaus identify an entropic floor through short-term stability and long-term persistence, signaling low-rank repetition without new information.The conditions filter noise and detect prolonged looping or stalling in a high-probability region.
- Additional Visualizations: Additional cases show 94.3% and 96.5% of post-RCP points inside the 95% basin, supporting recurrence across different problems.
- Counterfactual Validation: Injecting an incorrect intermediate answer produced the injected answer in 49 of 64 trials, while only 11 recovered the correct answer.The authors interpret this as confirmation bias: post-RCP generation seldom genuinely re-verifies and instead rationalizes earlier tokens.