Mitigating Overthinking in Large Reasoning Models via Manifold Steering

1 Introduction

• Overthinking adds unnecessary reasoning steps and inference burden, while external mitigation approaches may add probing overhead or depend on external models.The paper identifies the underlying mechanism as underexplored despite these behavioral approaches.Source: §1 · p. 2
• A single activation-space direction can steer models away from overthinking, but its benefit plateaus as intervention strength increases because the direction introduces interference noise.The paper traces this limitation to a steering direction that is not sufficiently accurate.Source: §1 · p. 2
• Manifold Steering projects the steering direction onto a low-dimensional activation manifold by nullifying a theoretically approximated interference term.The paper reports that this alignment better mitigates overthinking at larger intervention strengths.Source: §1 · p. 2
• Up to 71% token reduction is achieved on mathematical datasets while accuracy is maintained or improved, with consistent mitigation effects in code generation and knowledge-based QA.Experiments cover multiple DeepSeek-R1 distilled model sizes and report that the method surpasses existing methods in mitigation and accuracy preservation on the cross-domain tasks.Source: §1 · p. 2

Mechanistic Interpretability

• Mechanistic interpretability uses steering directions to manipulate model behavior; this work extends that paradigm to overthinking by projecting its steering direction onto a low-dimensional manifold.Related approaches include preference steering [7] and activation-space dimensionality reduction [6] [33].Source: §2 · pp. 2–3

Overthinking Mitigation

• Existing training-free overthinking methods use external monitoring or auxiliary models, while training-based length control [30] is computationally costly and orthogonal to inference-time interventions.Dynasor [16] monitors for redundant reasoning; Thought Manipulation [22] uses auxiliary models.Source: §2 · p. 3
• Concurrent mechanistic approaches [3] [8] are described as bottlenecking from interference noise in high-dimensional steering, whereas manifold projection is presented as reducing that noise.The paper reports stronger mitigation and token reductions across diverse tasks for its approach.Source: §2 · p. 3

3 Mechanistic Analysis of Overthinking

• The analysis investigates whether ablating a single activation-space direction can reduce redundant reasoning steps.The section frames this as a general mechanism for reducing overthinking to some extent.Source: §3 · p. 3

3.1 Background

• Transformer residual-stream activations aggregate context across layers, and the final token’s activation represents the input context; LRMs use a reasoning template in which reasoning follows the think marker.The section relates repetitive validation and redundant deliberation to potentially distinct residual-stream activation patterns.Source: §3.1 · p. 3

3.2 Extracting and Ablating an Overthinking Direction

• Redundant and concise activation distributions overlap in early layers but separate clearly in middle-to-late layers, supporting identification of an overthinking direction.The study constructs model-specific redundant and concise sets from sampled mathematical questions and five generated responses per question; the sets use token-length and hesitation-keyword criteria.Source: §3.2 · p. 4, Figure 1 · p. 4
• The direction is computed from final-token activations, normalized, and the most effective layer’s direction is applied across layers; ablation strength controls intervention extent across token positions and layers.The intervention balances reducing redundant reasoning against preserving problem-solving accuracy.Source: §3.2 · p. 4

4.1 Low-Dimensional Manifold Analysis

• As α increases, token count initially decreases, then beyond α=1.5 rebounds, nearly returning to no-intervention levels; the passage attributes this to unintended activation-space noise from an imprecise intervention direction.The paper’s steering-direction account includes an orthogonal interference component in M⊥.Source: §4.1 · pp. 5–6

4.2 Theoretical Analysis of Interference Noise

• Interference noise produces significant shifts in M⊥ that grow through layer-wise propagation, disrupting the model’s other normal abilities.The analysis describes amplification through attention and nonlinear transformations.Source: §4.2 · p. 6

4.3 Manifold Steering

• Manifold Steering projects the steering direction onto the activation manifold to mitigate interference, enabling sustained token-count reduction.The section reports that, unlike the original paradigm, manifold steering sustains the reduction.Source: §4.3 · pp. 6–7

5.1 Experimental Setups

• The evaluation uses four DeepSeek-R1 distilled models, mathematical benchmarks, and code and expert-knowledge QA datasets, measuring Pass@1 and average token count.The compared baselines are Dynasor and SEAL; recommended model settings include temperature 0.6, top-p 0.95, and a 16,384-token limit.Source: §5.1 · p. 7

5.2 Performance of Manifold Steering in Overthinking Mitigation

• On GSM8K, Manifold Steering reduces tokens by 41% ∼71% while preserving accuracy, and performs best across all models and datasets.The method is evaluated against Dynasor and SEAL; the setup requires complete solutions when stopping because Dynasor's early stopping can omit the problem-solving process.Source: §5.1 · p. 7, §5.2 · p. 8

5.3 Cross-Domain Transferability for Overthinking Mitigation

• Across LiveCodeBench and GPQA-Diamond, Manifold Steering reduces tokens by 12% ∼27% while maintaining accuracy.These code-generation and expert-knowledge QA tasks are distinct from the mathematical domain used to extract the steering direction.Source: §5.3 · pp. 8–9

5.4 Directional Analysis and Hyperparameter Tuning

• Forward steering removes hesitant and repetitive checking, whereas reverse steering induces verbosity; R1-7B uses α = 0.3 to balance token reduction against accuracy.For R1-7B on MATH500, α = 0.5 yields a 77.1% token reduction, but excessive strength can reduce accuracy.Source: §5.4 · p. 9

5.5 Cross-Task Transferability of Manifold Steering

• On AdvBench, Manifold Steering reaches a 100% jailbreak success rate versus 74% for baseline [2] (α = 2.0), with responses manually verified as valid.Source: §5.5 · p. 9

• Applicability to multimodal large language models remains unexplored.Source: §6 · p. 9
• The interaction with specialized reasoning, including legal or medical analysis, warrants further investigation.Source: §6 · p. 9
• Sensitivity to intervention strength suggests potential for dynamic steering that adapts strength to task complexity in real time.Source: §6 · p. 10

7 Conclusion

• Manifold Steering aligns the steering direction with the low-dimensional activation manifold to address overthinking while preserving task performance without additional computational cost.The conclusion also describes substantial token reductions and robust cross-task transferability.Source: §7 · p. 10

A Implement Details

• The implementation section focuses on selecting data for computing each model’s steering direction and configuring baseline methods.Source: A Implement Details · p. 14

A.1 Data Selection for Direction Computation

• The redundant set requires five responses over 16k tokens with more than 20 hesitation keywords, while the concise set requires five responses under 1k tokens with none.Responses in the redundant set are truncated at a hesitation keyword; “wait” is used as the consistent marker because different hesitation keywords show no significant activation-pattern differences.Source: §A.1 · p. 14

A.2 Baseline Methods

• Dynasor is configured to probe every 32 tokens and stop at a confidence threshold, while SEAL uses 1k Math samples and a layer-20 reasoning steering vector.The Dynasor configuration requires complete solutions upon stopping; SEAL adjusts hidden states of paragraph-delimiter tokens during greedy decoding.Source: §A.2 · pp. 14–15

B.1 Proof of Theorem 4.1

• The proof defines interference as the component of the overthinking direction orthogonal to the manifold, whose projection is determined by the top-k principal components of activation covariance.The manifold is estimated by PCA on redundant and concise reasoning data; k = 10 captures over 70% of the variance.Source: §B.1 · p. 15

• The proof decomposes the steering direction into a manifold component and an orthogonal interference component, then analyzes their contributions to the intervention-induced activation shift.It assumes activations follow a Gaussian model and that the interference component is nonzero.Source: §B.2 · pp. 16–17
• When the interference component is significant, it contributes a fresh perturbation to the activation shift at the next layer.The proof describes this contribution using attention and GeLU Jacobians and a lower bound on the linear transformation.Source: §B.2 · p. 17
• With layer-wise amplification γ > 1, the shift grows through both propagation of the accumulated shift and a new interference contribution at each layer.The stated recurrence combines these two contributions to the mean activation shift.Source: §B.2 · p. 17

C Hyperparameter Tuning

• α = 0.7 for R1-1.5B, 0.3 for R1-7B, 0.5 for R1-8B, and 0.3 for R1-14B are selected on MATH500 for an efficiency–accuracy balance.Figure 7 presents intervention strength’s impact on token reduction and accuracy for the four models.Source: C Hyperparameter Tuning · p. 18, Figure 7 · p. 18

D Layer Selection for Manifold Steering

• Later layers perform better on MATH500, with selected layers 27 for R1-1.5B and R1-7B, 31 for R1-8B, and 47 for R1-14B.The layer-wise analyses are reported in Tables 2–5.Source: D Layer Selection for Manifold Steering · p. 18, Table 2 · p. 18, Table 3 · p. 18

E Time Latency Analysis

• The analysis measures DeepSeek-R1-Distill-Qwen-7B latency on Math500, comparing the authors' method with Dynasor and SEAL on Ubuntu 22.04 with A800 GPUs.Source: E Time Latency Analysis · p. 18
• Dynasor exhibits significantly longer latency than the other compared methods, while SEAL has higher latency than the authors' method.The table reports average Math500 latency for overthinking-mitigation methods in R1-7B.Source: E Time Latency Analysis · p. 19, Table 6 · p. 19
• SEAL and the authors' method add negligible computational cost, but SEAL reduces tokens less effectively.The authors attribute SEAL's higher latency relative to their method to its less effective token reduction.Source: E Time Latency Analysis · p. 19