Source-linked AI summary
SABER: Stability-Aware Early Exit for LLM Reasoning via Adversarial Branch Probing
Wanli Cheng, Haiya Xiang, Juntao Li, Hongling Wang, Wenliang Chen
TL;DR
Long-chain reasoning can waste inference effort after intermediate answers stabilize, while existing confidence- and consistency-based exits have reliability or evaluation-latency limitations. SABER uses training-free adversarial branch probing with semantic consistency and confidence stability to estimate convergence and exit early. Across benchmarks and architectures, it reduces reasoning-token consumption by 30.2%–39.8% while maintaining competitive accuracy.
Problem
Long-chain reasoning can generate redundant steps, while confidence- or entropy-based exits may be unreliable and consistency-based approaches can require sequential evaluations.
Method
SABER constructs neutral and adversarial branches from intermediate reasoning states and uses lightweight probing to combine semantic consistency with confidence stability for adaptive early exit.
Results
30.2%–39.8% lower reasoning-token consumption is achieved across diverse benchmarks and model architectures while maintaining competitive overall accuracy.
Takeaways & Limitations
Perturbation stability provides a basis for adaptive early exit that reduces redundant reasoning while preserving comparable or improved accuracy.
Takeaways & Limitations
The evaluation covers text-based reasoning benchmarks with 4B–8B-parameter models and does not assess multimodal or agent-based scenarios.
Abstract
from arXiv · showhide
Large Reasoning Models (LRMs) achieve strong reasoning capabilities, yet long-chain reasoning becomes inefficient once the intermediate answer stabilizes across reasoning steps: additional reasoning yields little marginal benefit while incurring substantial inference cost. Existing early-exit methods based on confidence or entropy poorly capture reasoning stability, while consistency-based approaches rely on multi-step trajectory agreement, requiring sequential evaluations that delay exit. To better balance efficiency and reliability, we propose SABER, a training-free framework for stability-aware early exit via adversarial branch probing. SABER constructs simple yet effective semantic perturbations around intermediate reasoning states to form adversarial branches, and applies lightweight probing to estimate their likely final outcomes without full trajectory rollouts. When the probed outcomes remain consistent across branches, SABER exits early; otherwise, it continues reasoning. Experiments across multiple reasoning benchmarks and model architectures show that SABER reduces reasoning token consumption by 30.2\%--39.8\% on average while maintaining competitive accuracy with full-length reasoning.
1 Introduction
Long-chain reasoning can waste inference effort after sufficient information is available, while existing confidence- and consistency-based exits have reliability or latency limitations. SABER addresses this with training-free adversarial branch probing and reports 30.2%–39.8% lower reasoning-token consumption with competitive accuracy.
- Motivation: Long reasoning trajectories may continue with redundant verification or unnecessary paths after sufficient information is available.This can reduce inference efficiency and introduce superfluous steps or error accumulation.
- Motivation: Confidence- or entropy-based early exits can be unreliable because LRMs may remain highly confident when their reasoning is incorrect.Consistency-based approaches provide an alternative signal, but the supplied passage indicates prior work in this area as well.
- Approach: SABER hypothesizes that trajectories approaching convergence become stable under local semantic perturbations and proposes a lightweight, training-free framework to estimate that stability.The framework is designed for adaptive early exit across perturbed branches.
- Approach: SABER constructs neutral and adversarial branches from an intermediate state, uses short answer-focused continuations, and combines Semantic Consistency with Confidence Stability.The combined signals form a unified Reasoning Stability Score for stopping decisions.
- Results: 30.2%–39.8% lower reasoning-token consumption is reported across diverse benchmarks and model architectures while maintaining competitive overall accuracy.The result is stated as an extensive-experiment contribution of the paper.
2 Reasoning Stability under Adversarial Perturbations
The paper studies how intermediate reasoning states evolve under adversarial perturbations, using reasoning trajectories segmented into intermediate steps. Its framing focuses on perturbation-induced behavioral stability during multi-step reasoning.
- Motivation: The analysis examines how intermediate reasoning states evolve under adversarial perturbations throughout multi-step reasoning.This motivates treating behavioral stability as an informative signal for reliability and uncertainty estimation.
- Analysis setup: Figure 1 tracks answer consistency and confidence variation across reasoning steps under adversarial perturbations.The two views characterize how reasoning behavior changes as trajectories progress.
- Analysis setup: A reasoning trajectory is represented as tokens t1, t2, . . . , tn and segmented into intermediate steps S1, S2, . . . , SK using transition phrases.Examples include “Wait”, “Alternatively”, and blank-line transitions.
Appendix B.3.
The perturbation study probes intermediate reasoning prefixes with neutral and adversarial continuations, revealing distinct stability dynamics for correct and incorrect trajectories. These observations motivate stability-based early exit.
- Adversarial probing: At each reasoning step, the method appends neutral and adversarial probing prompts to the current reasoning prefix.The adversarial prompt introduces a misleading corrective signal.
- Adversarial probing: Each branch is sampled stochastically to obtain a predicted answer and confidence score for analyzing behavior under perturbations.The setup is applied throughout trajectories sampled from OlympiadBench.
- Observed dynamics: Correct trajectories become progressively more stable as probe predictions converge and confidence variation decreases during reasoning.This pattern is reported in the trajectory dynamics analysis.
- Implication: The framework overview combines semantic consistency and confidence stability into RSS to decide whether reasoning should continue or terminate early.This links the observed dynamics to the proposed stopping mechanism.
- Observed dynamics: Incorrect trajectories remain unstable at later stages, with persistent answer inconsistency and large confidence fluctuations.Their instability contrasts with the convergence pattern of correct trajectories.
3 Method
SABER performs lightweight adversarial probing at intermediate reasoning steps, converts branch predictions and confidence variation into stability signals, and exits when their aggregate score crosses a threshold. The design treats joint stability as evidence that further reasoning is unlikely to substantially change the prediction.
- 3.1 Adversarial Branch Probing: SABER branches a shared reasoning prefix into neutral and adversarial continuations to estimate stability at intermediate steps.The framework aggregates the resulting signals into a unified Reasoning Stability Score.
- 3.1 Adversarial Branch Probing: Each branch produces k stochastic samples, yielding answer multisets and confidence sets for comparison.The sampled outputs support both semantic-consistency and confidence-stability calculations.
- 3.2 Reasoning Stability Estimation: Semantic Consistency measures prediction agreement between neutral and adversarial branches using multiset Jaccard similarity.The answer multisets retain occurrence frequencies for each answer.
- 3.2 Reasoning Stability Estimation: Confidence Stability captures confidence variation across branches using confidence scores derived from length-normalized geometric means of token-level maximum predictive probabilities.The parameter γ controls sensitivity to confidence fluctuations, and the exponential form amplifies large deviations.
- 3.2 Reasoning Stability Estimation: Higher SC and CS jointly indicate greater perturbation stability and suggest that further reasoning is unlikely to substantially change the final prediction.The two signals provide complementary evidence that the current state has converged.
- 3.3 Early-Exit Decision: RSS = α · SC + (1 − α) · CS combines the two stability signals, with α controlling their relative contribution.RSS serves as a unified indicator of stability and convergence under adversarial perturbations.
- 3.3 Early-Exit Decision: SABER terminates reasoning when RSS exceeds threshold τ and then directly triggers final answer generation.For models with explicit reasoning delimiters, termination uses the closure token </think>.
- 3.3 Early-Exit Decision: The method’s intuition is that sufficiently stable states make additional reasoning unlikely to substantially change the final prediction while preserving reasoning quality.This motivates adaptive early exit rather than fixed-length reasoning.
4 Experiments
SABER is evaluated across diverse reasoning benchmarks, model scales, and training-free baselines, with experiments also examining its component and sampling-number trade-offs. Results indicate that SABER reduces reasoning cost while preserving or improving accuracy, and that jointly modeling its stability signals is especially valuable on harder tasks.
- Experimental Setup: SABER is evaluated on six mathematical reasoning benchmarks and GPQA Diamond using Accuracy, Token Count, and Compression Ratio.The study spans Qwen3-4B, Qwen3-8B, and DeepSeek-R1-Distilled-Qwen-7B.
- Experimental Setup: The evaluation compares SABER with vanilla reasoning, NoThinking, DEER, and Dynasor as representative training-free efficiency baselines.The baselines differ in whether they bypass reasoning or use confidence- or consistency-based early-exit signals.
- Overall Performance: 30.2%–39.8% fewer reasoning tokens are achieved versus vanilla reasoning while maintaining comparable or improved accuracy across model scales and benchmarks.DeepSeek-R1-Distilled-Qwen-7B improves overall accuracy from 67.8 to 69.0 while using 69.8% of the original reasoning tokens.
- Sampling Ablation: Increasing the sampling number k generally improves accuracy while increasing probe overhead, highlighting a trade-off between convergence reliability and computational cost.On MATH-500 with Qwen3-8B, increasing k from 1 to 4 raises accuracy from 88.8% to 91.6%.
- Comparison with Baselines: SABER consistently outperforms representative training-free baselines, with gains particularly evident on OlympiadBench and GPQA-D.NoThinking has high compression but substantial accuracy degradation, while Dynasor and DEER can make unreliable exit decisions under uncertainty.
5 Analysis
The analysis shows that combining semantic consistency with confidence stability improves early-exit reliability, while sampling and threshold choices trade accuracy against overhead. SABER’s probing adds limited cost and supports substantial token and latency savings across reasoning settings.
- Component Ablation: RSS outperforms SC-only and CS-only across all benchmarks, while single-signal variants degrade on challenging AIME24 and GPQA-D tasks.SC-only can exit on semantically consistent but incorrect answers, whereas CS-only can remain stable on incorrect trajectories.
- Scoring Function Ablation: Both alternative stopping scores remain competitive, but RSS performs best, indicating gains are not tied solely to its specific scoring formulation.The neutral/adversarial two-branch probing framework is held constant in this comparison.
- Sampling Number: 88.8% to 91.6%: increasing k from 1 to 4 improves MATH-500 accuracy, while larger sampling budgets provide diminishing returns and increase probe overhead from 3.3% to 22.7%.Beyond k = 4, accuracy improves by only 0.4% even when k reaches 32.
- Threshold Sensitivity τ: Within τ ∈[0.8, 0.95], performance remains relatively stable on OlympiadBench; higher thresholds improve accuracy but increase token consumption.Higher τ postpones early exit until stronger stability is achieved, reducing premature exits on ambiguous steps.
- Computational Overhead: 3.8% average probe overhead: adversarial probing accounts for a small fraction of total tokens, while early-exit savings substantially outweigh its additional cost.The reported model-specific probing ratios are 4.9%, 2.8%, and 3.8%.
- Inference Latency: 48.8% average latency reduction: SABER lowers wall-clock inference latency across GSM8K, AIME24, and GPQA-D, especially for longer reasoning trajectories.The latency gains are consistent with the token-reduction analysis.
- Overall Analysis: SABER estimates convergence through lightweight semantic perturbations rather than relying only on confidence or entropy or requiring full-generation rollouts.The framework is training-free and uses adversarial branch probing for adaptive early exit.
Limitations
The evaluation is limited to relatively small, text-based reasoning models and does not test multimodal or agent-based settings. Broader applicability remains unresolved.
- Scope: The experiments use text-based reasoning benchmarks and models ranging from 4B to 8B parameters, without evaluating larger-scale models.The authors identify larger-model evaluation as an open direction.
- Scope: SABER is not evaluated in multimodal or agent-based scenarios, where intermediate representations and confidence dynamics may differ significantly.The authors leave applicability in these broader settings for future work.
A.2 Dataset Details
The evaluation uses diverse reasoning benchmarks and standardized prompting, while SABER performs training-free, branch-based monitoring at reasoning transition points to determine early exit.
- Dataset Details: All datasets are evaluated zero-shot because their long reasoning chains and complex intermediate computations suit early-exit studies.
- Inference Setup: A unified prompting format requires models to generate intermediate reasoning steps before final answers, enabling fair comparison across early-exit methods.
- SABER Procedure: SABER monitors autoregressive reasoning and branches intermediate states into neutral and adversarial probing continuations with parallel stochastic sampling.
- SABER Procedure: SABER computes Semantic Consistency and combines it with Confidence Stability into a Reasoning Stability Score for stopping decisions.
C.2 Scoring Function Ablation
The ablation keeps SABER’s neutral/adversarial probing framework fixed while replacing its stopping score, showing that RSS benefits from combining semantic consistency with confidence stability.
- Scoring Function Ablation: The ablation compares alternative stopping scores while preserving the same neutral/adversarial two-branch probing framework.
- Alternative Scores: SC · CS multiplicatively combines semantic consistency and confidence stability instead of using RSS’s weighted-additive formulation.
- Alternative Scores: Branch-UQ Diff uses the absolute difference between branch-level uncertainties derived from answer-set entropy and perplexity.
- Results: Both alternative scores remain effective under the shared two-branch framework despite using different scoring formulations.
- Results: RSS performs better because it additionally models semantic consistency between branches rather than relying only on their uncertainty gap.
C.4 Trigger Sensitivity Analysis
Replacing the lexical trigger “Wait” with a neutral double newline produces comparable accuracy and token compression, indicating limited trigger sensitivity.
- Trigger Sensitivity: The sensitivity test replaces “Wait” with a semantically neutral double newline while keeping all other settings unchanged.
- Results: 70.7% vs. 70.8% overall accuracy and 71.9% vs. 66.4% token compression show comparable performance between the two triggers.
- Interpretation: The trigger primarily provides a practical way to locate potential reasoning transition points for probe generation.
D Case Study
Case studies show SABER detecting unstable states before convergence and exiting after stability, thereby reducing redundant reasoning and avoiding harmful or truncated reflection.
- Detecting Unstable Intermediate Reasoning States: SABER continues reasoning when neutral answers appear consistent but adversarial outputs remain unstable, then exits after both branches converge correctly.
- Avoiding Redundant Overthinking: SABER reduces token consumption by terminating once stable convergence is detected, unlike vanilla reasoning’s repeated verification after reaching a correct answer.
- Preventing Endless Reflection: SABER avoids context-limit truncation in cases where vanilla reasoning continues persistent reflection despite having reached the correct answer early.
- Overthinking and Accuracy: SABER preserves an initially correct answer when vanilla reasoning changes it incorrectly during subsequent reflection.