Autoregressive models have become the dominant paradigm in artificial intelligence, achieving remarkable performance across a wide range of applications. These models excel in reasoning tasks through structured approaches such as step-by-step reasoning, chain-of-thought prompting, and plan-and-solve strategies. However, generating text token-by-token in a fixed unidirectional manner creates a fundamental limitation: the sequentiality bottleneck. This bottleneck makes inference computationally costly and susceptible to error accumulation over long reasoning horizons. Diffusion language models offer a natural remedy through their inherent capacity for parallel token refinement, but a generic pre-trained diffusion model is optimized for distributional coverage rather than goal-directed reasoning, producing trajectories that are fluent but not necessarily correct.
The sequentiality bottleneck in autoregressive reasoning
Standard autoregressive reasoning suffers from a sequentiality bottleneck. Generating a chain-of-thought token-by-token precludes parallel exploration of the solution space, making inference both computationally costly and susceptible to error accumulation over long horizons. The paper formalizes this bottleneck and proves that a diffusion-style parallel refinement paradigm is not only more efficient but also strictly more expressive under logarithmic resource constraints. This theoretical foundation motivates the hybrid design that follows.
Collaborative reasoning asymmetric learning framework
The authors propose Collaborative Reasoning Asymmetric Learning (CoRAL), a hybrid framework built on an Asymmetric Planner-Executor (APE) architecture. A diffusion language model planner generates K diverse reasoning trajectories in parallel, while a large autoregressive model executor selects the optimal plan and synthesizes the final answer. This architecture exploits the complementary strengths of both paradigms: diffusion models support flexible token generation and massive parallelization, while autoregressive models generate fluent, coherent, and human-understandable sequences through next-token prediction.
Reward-guided denoising for chain-of-thought
To make the planner goal-directed, the authors introduce Reward-Guided Denoising for CoT (RDC), which integrates outcome-based rewards directly into the diffusion training process. The reverse diffusion process is conceptualized as the generation of a stochastic chain-of-thought. Each denoising step constitutes a single inferential thought in the chain. RDC aims to steer this chain generation process toward trajectories that culminate in correct answers.
The integrated denoising-guidance (IDG) objective combines a conventional negative log-likelihood loss with a guidance term. The guidance term is formulated as a reward-weighted negative log-likelihood of the entire thought trajectory, using a baseline for variance reduction. This integrated objective elegantly combines a standard diffusion objective with a guidance term, ensuring that every reasoning step is reinforced based on its contribution to the final correct chain-of-thought.
Asymmetric planner-executor collaboration
The CoRAL inference pipeline proceeds in two stages. In the planning stage, the multimodal input is fed into the diffusion language model planner, trained via the RDC objective, to generate a set of K parallel reasoning trajectories. In the execution stage, the autoregressive model executor receives the original query alongside the proposal set, evaluates all proposals in a single forward pass, selects the optimal plan, and generates the final answer.
The executor's role is not merely answer selection but causal re-integration, which compensates for the limitations of the parallel generation process. The parallel decoding in the diffusion planner suffers from a causal information deficit arising from the independence assumption made across tokens at each refinement step, leading to a loss of mutual information compared to a fully-dependent generation process. The executor's evaluation phase mitigates this deficit by re-contextualizing the acausal proposals within its own high-fidelity causal world model.
The evaluation uses a unified protocol: the executor computes the conditional probability distribution over the proposal indices and selects the optimal proposal via argmax. This single-pass evaluation adopts the parallel processing of prefill in large language model inference, making the cost of evaluating K proposals nearly equivalent to evaluating one. After selection, the executor generates the final answer by conditioning on both the original problem query and the chosen high-level plan.
Training the planner-executor interface
To enable collaboration between the diffusion planner and autoregressive executor, a projection layer maps the hidden states of the diffusion model into the embedding space of the autoregressive model. During training of this projection layer, the weights of both models are kept frozen. A cross-entropy loss function is adopted; conditioned on the latent space input, the predictions generated by the autoregressive model are compared against the corresponding ground-truth tokens from the benchmark set. The experimental setup uses LLaDA-V as the diffusion planner and Qwen2.5-VL-7B as the autoregressive executor.
Experimental results on vision-language benchmarks
Experiments on vision-language benchmarks show that CoRAL significantly outperforms prior diffusion-based methods and achieves competitive performance against strong autoregressive models. On three image description benchmarks—CapMAS, CapArena, and DetailCaps-4870—CoRAL achieves outstanding performance. On CapMAS, CoRAL achieves scores of 86.75 (CLAIR), 58.36 (Coverage), and 83.59 (Factuality), substantially surpassing the strongest diffusion baseline and outperforming top-tier autoregressive models such as Qwen2.5-VL and GLM-4v. On CapArena, CoRAL achieves -11.30, and on DetailCaps-4870 it reaches 63.88, solidifying its leading position in detailed image understanding.
In terms of efficiency, CoRAL boasts the highest throughput (0.42 on Trip-Planning and 0.58 on GPQA), confirming the computational advantages of its parallel generation method. On the Trip-Planning task, which demands complex constraint satisfaction, CoRAL achieves an accuracy of 0.31, more than double that of the best-performing autoregressive model Qwen2.5-VL-7B (0.12). On the expert-level scientific reasoning benchmark GPQA, CoRAL attains an accuracy of 0.39, leading all other models. These results validate the hypothesis that a parallel planning mechanism can effectively overcome the inherent sequential bottleneck of autoregressive models in reasoning-intensive problems.
Robustness of parallel generation
As generation speed increases (fewer inference steps), CoRAL's performance degrades less significantly than a masked-prediction baseline, demonstrating better stability in parallel generation. On the CLAIR metric, when generation speed increases from 1 token/step to 8 tokens/step, the masked-prediction model's score plummets from 74.3 to 46.4, whereas CoRAL's score only decreases from 77.4 to 68.1. On both Coverage and CAPTURE metrics, CoRAL again showcases more robust parallel generation capabilities; the magnitude of performance drop is far smaller than that of the masked-prediction baseline.
Ablation studies
The integrated CoRAL framework, combining both autoregressive and diffusion components, significantly outperforms the autoregressive-only and diffusion-only variants. When RDC is removed and the diffusion planner is trained with a standard masked prediction objective, a significant performance degradation is observed across all metrics. On Trip-Planning, accuracy drops from 0.31 to 0.25; on CLAIR, the score decreases from 86.75 to 82.36. This shows that RDC is crucial to guide the planner in generating high-quality, goal-oriented reasoning trajectories.
A variant relying solely on the RDC-trained diffusion planner (without the autoregressive executor) performs substantially worse than the full CoRAL framework. Its CLAIR score plummets from 86.75 to 74.35, and CAPTURE score falls from 63.88 to 61.75. This highlights the critical role of the executor in performing causal re-integration and structuring the final output. Conversely, removing the diffusion planner and forcing the autoregressive model to operate under an alternative guidance mechanism severely impairs reasoning capabilities. Trip-Planning accuracy drops drastically from 0.31 to 0.18, confirming that the autoregressive model cannot efficiently explore the vast solution space on its own.
Role asymmetry validation
The asymmetric planner-executor design is validated by comparing role-swapped configurations. The CoRAL setup (diffusion to autoregressive) dramatically outperforms the reverse (autoregressive to diffusion) on overall quality, achieving a total CLAIR score of 84.9 versus 75.1. While the reverse configuration yields slightly higher Coverage, this merely reflects the autoregressive model's bias toward exhaustive, sequential enumeration. As a planner, this verbose detailing harms structural coherence and complex reasoning. Factuality remains comparable across both setups, driven primarily by the foundational vision encoders. This confirms the core claim: diffusion models are ideal for concise, parallel planning, leaving large autoregressive models to serve as high-fidelity causal executors.
Limitations and future directions
CoRAL relies on a two-stage pipeline coupling a diffusion planner with a large autoregressive executor, which introduces additional system complexity and inference latency compared to single-model baselines. Experiments are conducted on English-language datasets; the effectiveness on multilingual or low-resource language tasks has not been investigated. The current evaluation focuses on static image inputs, and extending CoRAL to video understanding or streaming multimodal inputs is left for future work.