TL;DR ReState is a unified iterative reasoning framework that reformulates multimodal reasoning from a stage-driven pipeline into a state-driven dynamic process.
Most existing learning-based MCoT methods follow a stage-centric paradigm, assigning reasoning behaviors to predefined stages with separate optimization. This fragments reasoning and introduces a gap between the reasoning states used during training and those encountered at inference. ReState reformulates multimodal reasoning as a state-driven dynamic process, unifying reasoning construction, preservation, and correction through a shared state-update operator.
X = {I, Q, O, C}R̂ÂRGT but receives model-generated chains R̂ at inference, so upstream errors propagate downstream.(X, St)St+1Ât+1
URSM unifies reasoning construction, correct reasoning preservation, and erroneous reasoning correction as state transitions learned jointly by a single MLLM. CRSL incorporates self-generated historical reasoning states during training and recursively updates the current state at inference, mitigating the training–inference reasoning-state gap and enabling iterative self-refinement without additional critics, reflection modules, or separate correction stages.
Unified Reasoning State Modeling
X = {image, question, options, caption}; φ is represented by the token [EMPTY_SOLUTION]. In each training epoch, URSM constructs two views per sample: one with φ and one with a non-empty reasoning state, sampled from either R+ or R− with equal probability.
Closed-Loop Reasoning-State Learning
Hi · capacity 2 · FIFO
Each round, the state the model generates for sample i enters the pool and is reused as a training input Rh.
Examples from VersaCoT-10K, ScienceQA, and M³CoT (Figs. 14–16) show how the same model revisits its previous reasoning state to correct errors, align answers with valid reasoning, and preserve successful updates. Step through the rounds to follow this process: red highlights reasoning errors, amber indicates a mismatch between reasoning and the selected answer, and green highlights corrected or preserved evidence.
VersaCoT-10K contains 10,000 single-image multiple-choice samples spanning eight heterogeneous visual reasoning domains. Its construction consists of three stages: data collection and standardization, CoT annotation and quality control, and reasoning-state augmentation. The resulting dataset provides diverse supervision for multi-domain reasoning and reasoning-state refinement. Each domain contains 1,250 samples, with a 7 : 1 : 2 train / validation / test split.
Reasoning-intensive samples from 8 public benchmarks, unified into one multiple-choice format with options, answer, and image caption.
Gemini-3-Flash writes visually grounded CoTs. Rule filters (200–4,000 chars, answer-leak phrases) plus a separate Gemini-3.5-Flash judge (all three binary checks must pass, and each of the five scores must be ≥ 8).
Following RAD, each CoT gets positive states (diverse, semantically consistent) and hard negative states (similar, semantically reversed). Historical states Rh are generated online during training.
Under matched backbone and training settings across three backbones, ReState outperforms Direct SFT, Multimodal-CoT, and MIND by 0.90%–4.60% on VersaCoT-10K, 0.83%–2.87% on ScienceQA, and 0.18%–13.07% on M³CoT.
Best average accuracy in all 9 matched-backbone settings across 3 datasets and 3 backbones. For example, on VersaCoT-10K with Qwen3.5-9B, average accuracy improves from 77.80% with Direct SFT to 81.90% with ReState.
URSM improves accuracy by 0.50%–2.55% on VersaCoT-10K. Adding either training-time self-history states or inference-time iterative self-refinement brings further gains, with the full framework performing best across all three backbones.
One refinement round improves accuracy by 1.10%–2.65% on VersaCoT-10K. Most gains occur in the first few rounds, followed by stabilization with minor fluctuations.
Red = best, blue = second best, compared inside each 4-row backbone group. Highlighted rows are ReState.
On VersaCoT-10K, URSM improves accuracy by 0.50%–2.55% over Direct SFT. Adding SHS or ISR separately to URSM adds 0.30%–2.15% or 0.45%–0.55%, respectively. The full framework performs best across all three backbones, reaching 77.05%, 79.25%, and 81.90% on Qwen2.5-VL-7B, Qwen3-VL-8B, and Qwen3.5-9B, respectively.
Across VersaCoT-10K, ScienceQA, and M³CoT, iterative refinement delivers most of its gains in the first few rounds, after which accuracy stabilizes with minor fluctuations. The figures below summarize the component ablations and refinement trajectories.
Qualitative comparisons on VersaCoT-10K, ScienceQA, and M³CoT illustrate how visual evidence shapes reasoning outcomes. In these examples, Direct SFT, Multimodal-CoT, and MIND produce fluent reasoning but reach incorrect answers after misreading visual relationships, relying on misleading local cues, or introducing unsupported details. ReState identifies relevant evidence and integrates it into correct reasoning, from tracing diagram relationships and comparing concentrations to recognizing species and shapes. Click a card to view the full comparison (Figs. 8–13).
@article{yu2026stages,
title={From Stages to States: Unified Iterative Reasoning for MLLMs},
author={Yu, Chuang and Ma, Xuefei and Feng, Yuanhao and Zeng, Wuxiong and Wang, Bo and Cai, Minghong and Wu, Dongming and Zhou, Fei and Tang, Chen and Tong, Wenwen and others},
journal={arXiv preprint arXiv:2610.15620},
year={2026}
}