Background: The Efficiency Dilemma of Block Diffusion Models in Long-Chain Reasoning
In the pursuit of more powerful model capabilities, Test-Time Scaling has become a key technical path for enhancing the inference performance of Large Language Models. Against this backdrop, Block Diffusion Language Models (BDLMs), with their unique parallel decoding mechanism, are considered strong contenders to challenge the sequential generation paradigm of traditional autoregressive (AR) models and improve inference efficiency.
However, in practice, especially in complex tasks requiring long-chain reasoning, BDLMs face a difficult speed-accuracy paradox. On one hand, using a large block for parallel generation can achieve extremely high decoding speeds, but the coherence and accuracy of the generated content are often insufficient, leading to errors in complex reasoning. On the other hand, reducing the block size can improve reasoning accuracy, but this approach trends towards that of autoregressive models, forfeiting the core advantage of parallel decoding. Furthermore, existing fixed-confidence decoding strategies struggle to adapt to the dynamic changes in problem difficulty within a long reasoning chain, limiting the model’s potential. How to balance parallel efficiency with reasoning accuracy has become the key to unlocking the capabilities of BDLMs.
The TDAR Framework: “Think Coarse, Critic Fine” via a Dual-Adaptive Mechanism
To address the challenges above, researchers from Fudan University’s NLP Lab, Peking University’s Knowledge Computing Lab, and Meituan’s LongCat team have jointly proposed a unified inference framework called TDAR (Test-Time Scaling framework for Block Diffusion). The core idea of this framework is inspired by the human process of solving complex problems. By introducing a “Think Coarse, Critic Fine” (TCCF) paradigm and a Bounded Adaptive Confidence Decoding (BACD) algorithm, it successfully breaks the zero-sum game between speed and accuracy.
Innovation at the Decoding Layer: Bounded Adaptive Confidence Decoding (BACD)
Traditional dynamic decoding strategies often trigger cascading “one misstep leads to another” errors in long-chain reasoning due to fixed confidence thresholds. The BACD algorithm improves upon this by no longer relying on a fixed threshold. Instead, it uses the average confidence of already generated content as a dynamic signal to adjust the number of steps in the current denoising process in real-time. This algorithm introduces a dual-bound protection mechanism:
- Upper Bound: When the model shows high confidence in the generated content, it is allowed to adopt a more aggressive decoding strategy, reducing the number of denoising steps to accelerate generation.
- Lower Bound: When the model’s confidence is low, it is forced to perform more conservative and thorough denoising to ensure generation quality and prevent error accumulation.
This design enables the model to mimic human thought patterns: quickly passing through simple steps and investing more computational resources for careful deliberation on difficult parts.
Innovation at the Paradigm Layer: Think Coarse, Critic Fine (TCCF)
The research team observed that the long-chain reasoning process is not uniform but consists of divergent “exploratory” phases and convergent “verification” phases. Based on this, the TCCF paradigm was proposed, which dynamically allocates computational granularity according to the different stages of the reasoning task:
- Think Coarse: In the initial exploratory phase of reasoning, a larger block size (e.g.,
block_size=16) is used for rapid, parallel thought divergence to quickly build the framework of the reasoning path.
- Critic Fine: In the subsequent verification, correction, and summarization phases, the model switches to a fine-grained mode, using a very small block size (e.g.,
block_size=1) to verify key steps one by one, ensuring the rigor and correctness of the final answer.
To support efficient training with large block sizes, the research also introduced a “Progressive Block Size Extension” strategy, effectively mitigating potential performance degradation caused by increasing the block size.
Experimental Validation: A Dual Breakthrough in Speed and Accuracy
The team conducted a comprehensive evaluation of the TDAR-8B-Thinking model, based on this framework, across six mainstream complex reasoning benchmarks, including Math500, AIME24, GPQA, and LiveCodeBench. Experimental results show that TDAR achieved state-of-the-art performance among 8B-parameter block diffusion models.
Compared to the previous SOTA model, TraDo-8B, TDAR-8B-Thinking’s average performance improved by 3.4 percentage points. In terms of speed, its decoding efficiency increased significantly from 1.27 TPF (Tokens Per Forward) to 2.97 TPF. When applying the BACD algorithm alone, the speed can be further boosted to 3.37 TPF, with an additional performance gain of 1.6 percentage points.
Particularly noteworthy is that on AIME24, a high-difficulty mathematical reasoning task, the TDAR model using the complete BACD and TCCF strategies saw its accuracy jump from 36.3% to 42.9%, while still maintaining a high inference speed of 3.04 TPF, achieving a synergistic improvement in both performance and efficiency.
In-depth Analysis and Future Outlook
Through a series of ablation studies, the research team revealed the mechanisms behind the TDAR framework’s high efficiency. The analysis indicates that the BACD algorithm, compared to other dynamic decoding strategies, maintains stable high performance across different confidence thresholds, effectively avoiding common issues like “generation collapse” or “content repetition” that models often face at low thresholds. This demonstrates its robustness in dynamically adjusting computational costs.
The trade-off analysis of block size confirmed that B=16 is the sweet spot between speed and quality for the 8B-scale model. Combined with the progressive training strategy, TDAR successfully leverages the parallel advantages of large block sizes without sacrificing reasoning capabilities.
Furthermore, the TCCF paradigm has been proven to bring universal performance gains on top of various decoding algorithms, validating the effectiveness of the “Think Coarse, Critic Fine” strategy in compensating for the shortcomings of a single decoding granularity.
The proposal of TDAR opens up a new path for the application of block diffusion language models in complex reasoning tasks. It demonstrates that through sophisticated algorithm and paradigm design, the parallel advantages of large block sizes and high-quality reasoning capabilities can be simultaneously achieved. This work not only provides a viable solution for the test-time scaling of BDLMs but also offers valuable insights for the future architectural design of parallel inference models.
