DRT: Dense Reasoning Trace for Efficient and Grounded Multimodal Reasoning

Wan Xu; Yuanfan Guo; Kevin Han; LaLa Chen; Wangmeng Zuo

cs.CVSep 18, 2026

DRT: Dense Reasoning Trace for Efficient and Grounded Multimodal Reasoning

Authors: Wan Xu, Yuanfan Guo, Kevin Han, LaLa Chen, Wangmeng Zuo

Organizations: Harbin Institute of Technology · FIS, ByteDance Inc · Facebook · University of California, Irvine

Abstract

Despite the remarkable progress in Multimodal Large Language Models (MLLMs), prevailing Chain-of-Thought (CoT) paradigms remain confined to the natural-language expression space. Consequently, they inherently incur excessive linguistic overhead, leading to information dilution and weak visual grounding. To address this challenge, we propose Dense Reasoning Trace (DRT), a paradigm that departs from natural-language-centered CoT by expressing reasoning as compact structured traces, which include concise intermediate states with symbolic connectors and disentangle visual observations from logical deductions. First, we introduce the Dense Trace Initialization to internalize the DRT reasoning mode into the model, substantially improving token efficiency while preserving visual evidence. To further enable the model to faithfully capture the logical relations within traces, we propose the Trace-Grounded Reinforcement Learning framework, which builds reference traces through a tri-perspective verification pipeline and employs Trace-Grounded GRPO with structured rewards, encouraging the model to generate concise DRT-style traces with reduced hallucination and stronger logical grounding. Extensive experiments on challenging reasoning benchmarks show that DRT achieves 5.5 $\times$ token efficiency improvement while improving 1.3 accuracy points over the Qwen3-VL baseline. These findings suggest that complex multimodal reasoning may not require verbose natural-language traces, opening a more efficient path for next-generation MLLMs. Our code and data are available at: https://github.com/HIT-leaderone/DRT

Figures & tables

Figure 1: Qualitative and quantitative comparison of CoT and DRT reasoning paradigms . (A) : Qwen3-VL produces an overly long, low-density trace with an image-unfaithful assumption, leading to an incorrect answer. In contrast, DRT-RL successfully solves the problem using structured, compact reasoning traces. (B) : Across prompting paradigms and efficient-reasoning baselines, DRT-RL achieves a stronger accuracy-token tradeoff.

Method	MathVista Lu et al. (2024)		MathVerse Zhang et al. (2024)		LogicVista Xiao et al. (2024)		GSM8K Cobbe et al. (2021)		Video-Holmes Cheng et al. (2025)		AVG.
Method	Acc $\uparrow$	Tokens $\downarrow$	Acc $\uparrow$	Tokens $\downarrow$	Acc $\uparrow$	Tokens $\downarrow$	Acc $\uparrow$	Tokens $\downarrow$	Acc $\uparrow$	Tokens $\downarrow$	Acc $\uparrow$	Tokens $\downarrow$	QPS $\uparrow$	Lat. $\downarrow$
Qwen-DA	66.5	1.67	42.5	3.3	41.6	1.2	25.6	3.7	42.8	3.9	43.8	2.8	20.04	13.83
Qwen-DRT	74.9	555.6	60.5	974.1	52.4	1124.6	75.2	124.9	43.4	200.9	61.3	596.0	1.84	29.78
Qwen-Standard	77.2	820.0	62.1	1079.7	55.3	3165.1	95.2	339.6	39.6	294.4	65.9	1139.8	1.49	57.34
Qwen-Thinking	79.9	1848.4	74.2	2665.7	68.2	4528.6	96.1	1225.1	42.5	1868.3	72.2	2427.2	0.69	106.57
CoD Xu et al. (2025)	72.0	215.4	57.6	504.9	50.3	164.5	80.7	71.3	42.2	102.8	60.6	211.8	2.47	15.55

Table 1: Performance and token efficiency comparison on five reasoning benchmarks. For each benchmark, we report accuracy (Acc

\uparrow

) and average output tokens (Tokens

\downarrow

). The final columns report the macro average of Acc, Tokens, throughput (QPS

\uparrow

) and Latency (Lat.

\downarrow

SFT Init	RL Data Composition	Reward	MathVista	MathVerse	LogicVista	GSM8K	Video-Holmes	AVG.
SFT Init	RL Data Composition	Reward	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Tokens $\downarrow$
VisionR1-cold	Image+Text	$R_{\text{ans}}$	72.0	64.4	57.3	95.1	38.6	65.5	559.3
VisionR1-cold	Image+Text	$R_{\text{format}}+R_{\text{main}}+R_{\text{bonus}}$	72.3	63.4	57.7	95.5	41.4	66.1	626.3
DRT-SFT	Image-only	$R_{\text{format}}+R_{\text{main}}+R_{\text{bonus}}$	75.5	64.2	53.7	93.6	43.1	66.0	212.2
DRT-SFT	Image+Text	$R_{\text{format}}+R_{\text{ans}}$	74.4	62.0	50.8	92.7	42.5	64.5	159.8
DRT-SFT	Image+Text	$R_{\text{format}}+R_{\text{main}}$	74.5	63.1	54.1	94.1	44.8	66.1	170.4

Table 2: Ablation study on SFT init, RL data composition, and reward design. We report accuracy (Acc

\uparrow

) on five benchmarks and their macro average, along with average output tokens (Tokens

\downarrow

R_{\text{ans}}

denotes the standard GRPO reward based on answer-level supervision.

Figure 3: Step-level analysis across reasoning difficulty using GPT-5.1 as an external judge. Examples are grouped by ground-truth step counts, which serve as a proxy for reasoning difficulty. (A) shows the ground-truth step-count distribution, while (B)–(F) report per-bucket metrics: tokens per reasoning step, hallucination ratio, reasoning step, effective reasoning step, and accuracy.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Case-level comparison of reasoning behaviors on a MathVerse geometry problem.

Method	MathVista Lu et al. (2024)		MathVerse Zhang et al. (2024)		LogicVista Xiao et al. (2024)		GSM8K Cobbe et al. (2021)		Video-Holmes Cheng et al. (2025)		AVG.
Method	QPS $\uparrow$	Latency $\downarrow$	QPS $\uparrow$	Latency $\downarrow$	QPS $\uparrow$	Latency $\downarrow$	QPS $\uparrow$	Latency $\downarrow$	QPS $\uparrow$	Latency $\downarrow$	QPS $\uparrow$	Latency $\downarrow$
Qwen-DA	13.56	9.07	14.94	8.39	22.82	4.35	46.19	2.42	2.69	44.93	20.04	13.83
Qwen-DRT	1.42	23.37	0.97	60.12	0.74	37.20	4.29	3.33	1.76	24.87	1.84	29.78
Qwen-Standard	1.22	32.75	0.99	59.08	0.31	160.29	3.33	7.83	1.62	26.74	1.49	57.34
Qwen-Thinking	0.68	75.22	0.51	119.89	0.29	187.64	1.47	32.65	0.51	117.45	0.69	106.57
CoD Xu et al. (2025)	2.29	10.30	1.67	32.38	1.58	6.23	5.13	2.09	1.69	26.77	2.47	15.55

Appendix

Table 3: System-level efficiency comparison across efficient reasoning methods on five reasoning benchmarks. For each benchmark, we report throughput (QPS

\uparrow

) and latency (Latency

\downarrow

). The final columns report the macro average of QPS and Latency across all five benchmarks.

Model Size	MathVista	MathVerse	LogicVista	GSM8K	Video-Holmes	AVG.
Model Size	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Tokens $\downarrow$
Qwen3-8B	71.9	58.6	53.2	92.1	41.6	63.5	449.0
Qwen3-14B	74.6	63.0	56.8	92.3	44.2	66.2	377.5
Qwen3-32B	74.9	64.1	55.7	94.2	43.3	66.5	244.3
Qwen3-30B-A3B	75.9	64.1	54.8	93.9	45.0	66.7	284.2
Qwen3-235B-A22B	76.8	64.3	56.8	94.3	43.7	67.2	206.9

Appendix

Table 4: Ablation study on reward model size. We report accuracy (Acc

\uparrow

) on five reasoning benchmarks and their macro average, along with average output tokens (Tokens

\downarrow

Asset	License	URL
Qwen3-VL Bai et al. (2025)	Apache License 2.0	https://huggingface.co/collections/Qwen/qwen3-vl
vLLM Kwon et al. (2023)	Apache License 2.0	https://github.com/vllm-project/vllm
ms-swift Zhao et al. (2024)	Apache License 2.0	https://github.com/modelscope/ms-swift
verl Sheng et al. (2024)	Apache License 2.0	https://github.com/volcengine/verl
DAPO Yu et al. (2025)	Apache License 2.0	https://huggingface.co/datasets/BytedTsinghua-SIA/DAPO-Math-17k
Vision-R1-cold Huang et al. (2026a)	cc-by-nc-sa-4.0	https://huggingface.co/datasets/Osilly/Vision-R1-cold

Appendix

Table 5: Licenses of primary existing assets used in this work.

Multimodal Reasoning

Multimodal reasoning studies how AI systems combine text with images, audio, video, and other signals to solve evidence-dependent tasks.

See all 508 papers

CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts
Jun 30, 2026Lianyu Hu, Shengqian Qin, Zeqin Liao +4Latent CoT Reasoning Efficient Language Model Reasoning
Nanyang Technological University, Singapore · Shanghai Jiao Tong University, China · NKIARI, Shenzhen Futian, China +2
Attention-guided Fine-tuning of Multimodal Large Language Models Improves Chain-of-Thought Reasoning
Jun 1, 2026Sanchit Sinha, Guangzhi Xiong, Bohan Liu +2Multimodal CoT Reasoning Multimodal Large Language Models
University of Virginia
Look Light, Think Heavy: What Multimodal Chain-of-Thought Reasoning Can and Cannot Do
Jun 21, 2026Zhuoran Jin, Kejian Zhu, Hongbang Yuan +5Multimodal Reasoning VLM Reasoning
The Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China

Jun 30, 2026J/K move · Enter open · S save

CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts

Lianyu Hu, Shengqian Qin, Zeqin Liao, Qing Guo+3

Nanyang Technological University, Singapore · Shanghai Jiao Tong University, China · NKIARI, Shenzhen Futian, China+2

Latent CoT Reasoning Efficient Language Model Reasoning Latent Visual Reasoning Multimodal CoT Reasoning Multimodal Foundation Models Efficient Multimodal Inference+2

Open paper AskPDF arXiv

Chain-of-thought (CoT) reasoning has enabled multi-modal large language models (MLLMs) to tackle complex visual reasoning tasks by generating explicit intermediate reasoning steps in natural language. However, this text-based reasoning paradigm is inherently slow at inference time with even thousands of tokens and fundamentally constrained by the expressiveness of natural language. In this paper, we propose CoLT, (Chain of Latent Thoughts), a novel framework that teaches multi-modal models to reason through a chain of latent thought representations instead of verbose text tokens, which can perform thinking with as few as 3 steps. Naively forcing the model to think with latent states easily produces meaningless semantics and makes training unstable. To effectively regulate the latent reasoning process, we introduce a lightweight external decoder that provides step-level supervision for each latent reasoning step in two complementary directions: a forward mode that decodes latent thoughts into the textual reasoning of the next step, and a backward mode that aligns decoder hidden states with the model's latent thoughts given preceding textual context. We further incorporate internal supervision that encourages coherent step-by-step latent transitions. The decoder and internal supervision are removed during inference to maintain high efficiency of latent reasoning. Extensive experiments on eight benchmarks demonstrate that CoLT not only outperforms existing latent reasoning methods such as CODI and SIM-CoT, but also surpasses latent visual reasoning approaches that rely on auxiliary images with costly annotation requirements. Compared to text CoT methods, CoLT can notably reduce the inference time by 10.1

\times

and text decoding time by 22.6

\times

. Code is released at https://github.com/hulianyuyy/CoLT.

Abstract

Multimodal Reasoning

Explore similar work

CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts

Attention-guided Fine-tuning of Multimodal Large Language Models Improves Chain-of-Thought Reasoning

Look Light, Think Heavy: What Multimodal Chain-of-Thought Reasoning Can and Cannot Do