Recursive Think-Answer Process for LLMs and VLMs

Byung-Kwan Lee; Youngchae Chee; Yong Man Ro

cs.CLMar 2, 2026

Recursive Think-Answer Process for LLMs and VLMs

Authors: Byung-Kwan Lee, Youngchae Chee, Yong Man Ro

Organizations: KAIST

Abstract

Think-Answer reasoners such as DeepSeek-R1 have made notable progress by leveraging interpretable internal reasoning. However, despite the frequent presence of self-reflective cues like "Oops!", they remain vulnerable to output errors during single-pass inference. To address this limitation, we propose an efficient Recursive Think-Answer Process (R-TAP) that enables models to engage in iterative reasoning cycles and generate more accurate answers, going beyond conventional single-pass approaches. Central to this approach is a confidence generator that evaluates the certainty of model responses and guides subsequent improvements. By incorporating two complementary rewards-Recursively Confidence Increase Reward and Final Answer Confidence Reward-we show that R-TAP-enhanced models consistently outperform conventional single-pass methods for both large language models (LLMs) and vision-language models (VLMs). Moreover, by analyzing the frequency of "Oops"-like expressions in model responses, we find that R-TAP-applied models exhibit significantly fewer self-reflective patterns, resulting in more stable and faster inference-time reasoning. We hope R-TAP pave the way evolving into efficient and elaborated methods to refine the reasoning processes of future AI.

Figures & tables

Figure 2: Overall accuracy (%) of numerous vision language models (VLMs) on five evaluation benchmarks—MMMU [ 110 ] , MathVista [ 72 ] , OlympiadBench [ 33 ] , MathVision [ 94 ] , and MMMU-Pro [ 111 ] .

Figure 3: Qualitative example of recursive think–answer process on a combinatorics question. The model iteratively refines its solution across multiple reasoning cycles, successfully correcting initial misconceptions such as off-by-one errors.

Figure 4: Recursive Think-Answer Process. Given a question $q$ , base LLM/VLM $\pi_{\theta}$ recursively generates multiple Think-Answers $o^{(t)}$ until the answer is correct $t=M$ . In this example, effective recursion depth $M=\text{3}$ . A pre-trained Confidence Generator $\mathbb{C}_{\phi}$ assess each question and Think-Answer pair $(q,o^{(t)})$ then generates confidence score $\text{Conf}^{(t)}$ . This confidence score is used to formulate confidence-based reward – $R_{\text{Increase}}$ and $R_{\text{Final}}$ – which serves as a sufficient reinforcement signal to train the model to recursively generate higher confidence Think-Answers until intrinsic confidence is high enough. Note that full responses for this question is described in Appendix A.

Table 4

Figure 5: Training curves showing the progression of three reward signals—recursively confidence increase reward, last answer’s confidence reward, and accuracy reward—over iterations during GRPO [ 86 ] . All rewards show consistent upward trends, indicating effective recursive refinement.

Table 6

Table 7

Model	GSM8K	MATH500	Minerva Math	Olympiad Bench	AIME24 (P@1)	AIME24 (Avg@32)	AMC23
Llama, DeepSeek and Mistral Models
Mistral-v0.1-7B [ 43 ]	21.2	4.2	4.0	2.4	0.0	0.0	0.0
+ SimpleRL-Zoo	75.0	15.8	6.6	4.1	0.0	0.2	10.0
+ R-TAP	81.2	21.3	13.9	10.0	6.4	9.5	17.4
Llama-3.1-8B [ 24 ]	39.7	13.6	4.8	3.1	0.0	0.2	2.5
+ SimpleRL-Zoo	79.2	23.0	9.6	5.3	0.0	0.2	15.0

Table 9: Performance of Llama3.1 [ 29 ] , DeepSeek-Math [ 86 ] , Mistral [ 43 ] , and Qwen-2.5 [ 103 ] models before and after SimpleRL-Zoo [ 113 ] fine-tuning and R-TAP integration across various mathematical reasoning benchmarks. Benchmarks include GSM8K [ 17 ] , MATH500 [ 74 ] , Minerva [ 78 ] , OlympiadBench [ 33 ] , AIME [ 81 ] , and AMC [ 4 ] . Note that, the selection of evaluation benchmarks and their performance results, and the selection of the compared models are referred to the manuscript of SimpleRL-Zoo [ 113 ] .

Table 9

Table 10

Table 11

Figure 6: Impact of R-TAP on reducing the number of “Oops”-style words – which corresponds to the number of erroneous reasoning – and its effect on substantially reducing inference time. (Left) Negative correlation between the number of erroneous reasoning and R-TAP train iterations. (Center) Evaluation result on the significant reduction of erroneous reasoning by applying R-TAP. (Right) Evaluation result on the substantial reduction of inference time due to the reduction of erroneous reasoning by applying R-TAP.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Qualitative multimodal example of recursive think–answer process. The model refines its answer over multiple reasoning cycles by correcting initial counting errors and verifying object-level visual details, ultimately identifying the only flower that meets both criteria.

	MaxTokenLimit	AIME25	HMMT Feb25	OmniMath	GPQA	LCB	Avg	Oops-style count	Decoding token count	Training Time
Phi-4-reasoning-plus	32768	78.0	53.6	81.9	69.3	65.9	69.7	15.7	14509.7	-
w/ Reflexion [ 87 ]	32768	80.2	56.1	83.4	71.3	66.5	71.5	18.9	15230.4	51 hours
w/ Self-Consistency [ 97 ]	32768	81.1	55.8	84.0	72.2	66.4	71.9	16.8	16012.9	58 hours
w/ Self-Refine [ 73 ]	32768	78.5	54.9	83.1	71.0	68.5	71.2	14.3	13890.6	45 hours
w/ Self-Verification [ 68 ]	32768	82.4	57.6	85.7	73.5	64.8	72.8	17.5	14890.3	51 hours
R-TAP	32768	83.7	60.3	86.2	76.7	72.1	75.8	5.6	4378.8	28 hours

Appendix

Table 14

	MaxTokenLimit	MMMU	MathVista	OlympiadBench	MathVision	MMMU-Pro	Avg	Oops-style count	Output token count	Training Time
R1V2-38B	12000	73.6	74.0	62.6	49.0	52.0	62.2	17.2	9890.7	-
w/ Reflexion [ 87 ]	12000	74.5	74.8	63.0	49.5	53.7	63.1	17.5	10989.3	57 hours
w/ Self-Consistency [ 97 ]	12000	75.0	76.0	64.0	50.0	52.5	63.5	18.6	10678.5	60 hours
w/ Self-Refine [ 73 ]	12000	73.8	74.2	62.0	49.0	55.5	62.9	15.3	9938.2	59 hours
w/ Self-Verification [ 68 ]	12000	76.5	77.0	65.0	51.0	53.0	64.5	17.8	10183.8	58 hours
R-TAP	12000	78.2	82.3	69.4	56.8	59.2	69.2	8.5	5789.4	39 hours

Appendix

Table 15

T	G	AIME25	HMMT Feb25	OmniMath	GPQA	LiveCodeBench	Avg	Output token count	Training Time
1	4	76.1	52.1	80.4	67.9	64.4	68.2	15034.9	51 hours
1	8	77.2	52.9	81.2	68.6	65.2	69.0	15210.7	52 hours
1	12	78.0	53.6	81.9	69.3	65.9	69.7	14872.5	58 hours
2	12	80.9	57.1	84.1	73.3	68.6	72.6	9258.9	42 hours
3	12	82.3	58.9	85.0	75.1	70.3	74.9	6127.4	34 hours
4	12	83.7	60.3	86.2	76.7	72.1	75.8	4378.8	28 hours

Appendix

Table 16

T	G	MMMU	MathVista	OlympiadBench	MathVision	MMMU-Pro	Avg	Decoding token count	Training Time
1	4	72.4	72.6	61.2	47.8	50.6	60.9	10342.7	55 hours
1	8	73.1	73.4	62.0	48.5	51.4	61.7	9927.3	62 hours
1	12	73.6	74.0	62.6	49.0	52.0	62.2	10168.9	68 hours
2	12	76.1	77.0	65.3	52.7	55.0	64.1	8234.6	55 hours
3	12	77.5	79.4	67.2	54.9	57.8	67.4	6912.1	46 hours
4	12	78.2	82.3	69.4	56.8	59.2	69.2	5789.4	39 hours

Appendix

Table 17

Voting N	AIME25	HMMT Feb25	OmniMath	GPQA	LiveCodeBench	Avg
1	83.7	60.3	86.2	76.7	72.1	75.8
3	83.9	60.5	86.4	76.9	72.3	76.0
5	84.0	60.6	86.5	77.0	72.4	76.1
7	84.1	60.7	86.6	77.1	72.5	76.2
9	84.1	60.7	86.6	77.1	72.5	76.2
12	84.2	60.8	86.7	77.2	72.6	76.3

Appendix

Table 18

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

A Comprehensive Anatomy of Human and DeepSeek-R1 LLM Mathematical Reasoning
Jun 5, 2026Yuxiang Chen, Jun WangCoT Reasoning Structured Reasoning
UCL Centre for Artificial Intelligence
Think Deep, Speak Once: Relit, A Recursive Latent Implicit Transformer Framework
Aug 8, 2026Abhishek Panwar, Maheep Singh, Saksham BansalContinuous Latent Reasoning Efficient Language Model Inference
Indian Institute of Technology Roorkee, India
Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens
Feb 13, 2026Wei-Lin Chen, Liqian Peng, Tian Tan +5LLM Inference Efficiency Inference-Time Scaling
University of Virginia · Google

Jun 5, 2026J/K move · Enter open · S save

A Comprehensive Anatomy of Human and DeepSeek-R1 LLM Mathematical Reasoning

Yuxiang Chen, Jun Wang

UCL Centre for Artificial Intelligence

CoT Reasoning Structured Reasoning Metacognition in Language Models LLM Mathematical Reasoning Large Reasoning Models Reasoning Trace Analysis

Open paper AskPDF arXiv

The emergence of "Aha moments" in large language models, particularly DeepSeek-R1-0120, has raised the question of whether these systems genuinely reason or merely imitate the appearance of reasoning. We conduct a comprehensive empirical comparison between model and human reasoning across all 30 problems from AIME 2025, exhaustively annotating 10,247 reasoning steps into five functional categories: Analysis, Inference, Branch, Backtrace, and Reflection. We find a clear structural difference. Human solutions maintain a compact alternation between analysis and deduction, whereas DeepSeek-R1 frequently revisits intermediate results, performs shallow and often unnecessary verification, and loops through local checks without meaningful logical progress. We describe this as topological mimicry: reproducing the surface form of reasoning without its functional role. Despite this, we identify two signals of genuine reasoning. First, successful traces exhibit stable use of branching and backtracking, while failed traces either underuse or overuse exploratory actions. Second, reflection is only effective when placed within deductive inference; reflections trapped in analysis loops focus on local numerical details while missing global logical errors. These findings suggest that current long-CoT models may be rewarded more for the appearance of reasoning than for genuine deductive progress. We discuss directions for improving evaluation and training, including measuring cross-trace stability, penalising "spinning-wheel" traces, encouraging deeper logical correction, and reallocating inference-time compute toward deduction and backtracking. Overall, reasoning quality depends not simply on how much reflection occurs, but on whether reflection appears consistently and at the appropriate logical scale.

Abstract

Efficient Language Model Reasoning

Explore similar work

A Comprehensive Anatomy of Human and DeepSeek-R1 LLM Mathematical Reasoning

Think Deep, Speak Once: Relit, A Recursive Latent Implicit Transformer Framework

Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens