When Is Thinking Enough? Early Exit via Sufficiency Assessment for Efficient Reasoning

Yang Xiang; Yixin Ji; Ruotao Xu; Dan Qiao; Zheming Yang; Juntao Li; Min Zhang

cs.CLApr 8, 2026

When Is Thinking Enough? Early Exit via Sufficiency Assessment for Efficient Reasoning

Authors: Yang Xiang, Yixin Ji, Ruotao Xu, Dan Qiao, Zheming Yang, Juntao Li, Min Zhang

Organizations: Soochow University · ByteDance

Abstract

Large reasoning models (LRMs) have achieved remarkable performance in complex reasoning tasks, driven by their powerful inference-time scaling capability. However, LRMs often suffer from overthinking, which results in substantial computational redundancy and significantly reduces efficiency. Early-exit methods aim to mitigate this issue by terminating reasoning once sufficient evidence has been generated, yet existing approaches mostly rely on handcrafted or empirical indicators that are unreliable and impractical. In this work, we introduce Dynamic Thought Sufficiency in Reasoning (DTSR), a novel framework for efficient reasoning that enables the model to dynamically assess the sufficiency of its chain-of-thought (CoT) and determine the optimal point for early exit. Inspired by human metacognition, DTSR operates in two stages: (1) Reflection Signal Monitoring, which identifies reflection signals as potential cues for early exit, and (2) Thought Sufficiency Check, which evaluates whether the current CoT is sufficient to derive the final answer. Experimental results on the Qwen3 models show that DTSR reduces reasoning length by 28.9%-34.9% with minimal performance loss, effectively mitigating overthinking. We further discuss overconfidence in LRMs and self-evaluation paradigms, providing valuable insights for early-exit reasoning.

Figures & tables

Figure 1: Overview of the Dynamic Thought Sufficiency in Reasoning (DTSR) method. Upon a reflection signal, the model evaluates the sufficiency of its previous thought. If the thought is insufficient, reasoning continues; checks are skipped within k tokens. Once the thought is sufficient, </think> is appended and a conclusion is generated.

Method	GSM8K		MATH-500		AMC		GPQA		Olympiad		Overall
Method	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$
Qwen3-8B
Vanilla	96.0	2208	94.4	5067	92.5	7232	60.1	8861	66.4	9182	81.9	6510
NoThinking	92.6	303	86.0	1067	72.5	2264	53.0	1271	48.3	2385	70.5	1458
NoWAIT	94.9	1307	92.8	3485	90.0	5448	55.6	6920	62.4	6998	79.1	4832
DEER	95.1	963	93.4	2549	87.5	5155	57.1	8306	63.4	5768	79.3	4532

Table 1: Experimental results across various reasoning models. "Acc" denotes accuracy, and "Tok" denotes token count.

\uparrow

indicates that higher values are better, while

\downarrow

indicates that lower values are better.

Figure 2: Performance of DTSR on LiveCodeBench.

Figure 3: Comparison of Vanilla and DTSR on MATH-500 as the token budget increases.

Figure 4: Impact of token interval

k

Figure 5: Impact of threshold

\tau

Method	MATH-500		AMC
	Len	Latency	Len	Latency
Vanilla	4480	3.1	6562	8.3
DEER	2601	4.2	4514	15.6
DTSR	2247	1.9	3975	6.2

Table 2: Comparison of Inference Latency (s).

Figure 6: Comparison between DEER and its two variants, DEER-1 and DEER-2.

Method	Accuracy ( $\uparrow$ )	Length( $\downarrow$ )
DTSR	95.0	2247
DTSR-1 ( $\tau$ =100)	94.2	3254
DTSR-1 ( $\tau$ =95)	93.8	2457

Table 3: Comparison between DTSR and DTSR-1.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

“Wait”, “Alternative”, “Alternatively”

“But wait”, “But let me”, “But let’s”

Appendix

Table 4: Reflection Signal List

Figure 7: An illustrative example showing how we identify the optimal exit point. After each generated sentence, we append </think> to force the model to produce an answer. The correctness of the resulting answer is then used to determine whether the current reasoning state is sufficient and corresponds to the optimal exit point.

Panel of Figure 8 — Figure 8: Thought Sufficiency Check Prompt Template

Panel of Figure 9 — Figure 9: Differences between DTSR and its variant DTSR-1.

Method	GSM8K		MATH-500		AMC		GPQA		Overall
Method	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$
Qwen3-14B
Vanilla	96.2	1672	95.4	4503	97.5	6562	65.7	7513	88.7	5062
RL + Length Penalty	95.8	1090	95.8	2866	97.5	5059	65.4	4949	88.6	3491
S-GRPO	96.3	952	96.4	2652	100	4538	67.6	4537	90.1	3170
DTSR	96.2	849	95.0	2247	97.5	3975	66.2	5916	88.7	3246

Appendix

Table 5: Comparison of DTSR with training-based efficient reasoning methods. "Acc" denotes accuracy, and "Tok" denotes token count.

\uparrow

indicates that higher values are better, while

\downarrow

indicates that lower values are better.

Method	GSM8K		MATH-500		AMC		GPQA		Overall
Method	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$
Qwen3-14B
Vanilla	96.2	1672	95.4	4503	97.5	6562	65.7	7513	88.7	5062
TALE	95.7	1241	94.6	4484	95.0	6623	65.6	7138	87.7	4872
CoD	85.7	648	75.2	2359	72.5	4122	51.0	1177	71.1	2077
DTSR	96.2	849	95.0	2247	97.5	3975	66.2	5916	88.7	3246

Appendix

Table 6: Comparison of DTSR with TALE and Chain-of-Draft (CoD). "Acc" denotes accuracy, and "Tok" denotes token count.

\uparrow

indicates that higher values are better, while

\downarrow

indicates that lower values are better.

Method	GSM8K		MATH-500		AMC		GPQA		Olympiad		Overall
Method	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$	Acc $\uparrow$	Tok $\downarrow$
Qwen3-14B
Vanilla	96.3	1676	94.8	4511	97.5	6998	64.1	7382	65.6	8644	83.7	5842
DTSR	96.1	814	94.6	2296	97.5	3977	65.6	6103	66.8	5967	84.1	3831

Appendix

Table 7: Experimental results under greedy decoding. "Acc" denotes accuracy, and "Tok" denotes token count.

\uparrow

indicates that higher values are better, while

\downarrow

indicates that lower values are better.

Figure 10: An example illustrating the reasoning process of DTSR.

Figure 11: An example illustrating the reasoning process of DTSR on a promgramming problem.

Panel of Figure 12 — Figure 12: Comparison of Vanilla’s and DTSR’s answer on an open question.

Efficient Language Model Reasoning

Efficient language model reasoning studies how to improve problem-solving accuracy while reducing inference, memory, latency, and training costs.

See all 370 papers

Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring
Mar 15, 2026Weixin Guan, Liang Li, Jiapeng Liu +6Overthinking in Language Models Efficient Language Model Reasoning
Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyberspace Security, University of Chinese Academy of Sciences, Beijing, China · School of Computer and Artificial Intelligence, Jiangxi University of Finance and Economics, Jiangxi, China +1
Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning
Apr 18, 2026Benteng Chen, Weida Wang, Shufei Zhang +2RL for Language Model Reasoning LLM Inference Acceleration
East China Normal University · Shanghai AI Laboratory · Fudan University +1
Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models
May 17, 2026Dehai Min, Giovanni Vaccarino, Huiyi Chen +3Large Reasoning Models CoT Reasoning
University of Illinois Chicago · Politecnico di Milano · University of Illinois Urbana-Champaign +1

Mar 15, 2026J/K move · Enter open · S save

Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring

Weixin Guan, Liang Li, Jiapeng Liu, Bing Li+5

Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyberspace Security, University of Chinese Academy of Sciences, Beijing, China · School of Computer and Artificial Intelligence, Jiangxi University of Finance and Economics, Jiangxi, China+1

Overthinking in Language Models Efficient Language Model Reasoning Large Reasoning Models

Open paper AskPDF arXiv

Large Reasoning Language Models (LRLMs) demonstrate impressive capabilities on complex tasks by utilizing long Chain-of-Thought reasoning. However, they are prone to overthinking, which generates redundant reasoning steps that degrade both performance and efficiency. Recently, early-exit strategies are proposed to mitigate overthinking by dynamically and adaptively terminating redundant reasoning. However, current early-exit methods either introduce extra training overhead by relying on proxy models or limit inference throughput due to the frequent content switching between reasoning and generating probing answers. Moreover, most early-exit methods harm LRLMs performance due to over-truncation. Our insight stems from an observation: overthinking often causes LRLMs to deviate from the correct reasoning path, which is frequently accompanied by high-entropy transition tokens. Given this, we propose an early-exit method deeply coupled with the native reasoning process, which leverages the path deviation index as a dedicated monitoring metric for the frequent occurrence of high-entropy transition tokens to dynamically detect and terminate overthinking trajectories. We conduct experiments across multiple benchmarks using LRLMs of different types and scales, and the results indicate that our method delivers the largest performance improvement over vanilla CoT compared to existing early-exit methods.

Abstract

Efficient Language Model Reasoning

Explore similar work

Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring

Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning

Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models