Bengio’s Team Unveils GRAM: A Probabilistic Recursive Reasoning Model Breaking Inference Bottlenecks with Parallel Exploration
In the field of artificial intelligence, enhancing the efficiency and quality of model reasoning is a core challenge. Recently, a research team including Turing Award laureate Yoshua Bengio proposed a new architecture called GRAM (Generative Recursive reAsoning Models) in a paper titled “Generative Recursive Reasoning,” offering a fresh approach to this problem.
GRAM’s key innovation is its transformation of traditional deterministic recursive reasoning into a probabilistic multi-trajectory computation. Instead of being confined to a single, fixed reasoning path, the model explores multiple possible solution trajectories in parallel within the latent space, achieving significant breakthroughs in both efficiency and performance.
From Deterministic Single-Track to Probabilistic Multi-Track
Existing Recursive Reasoning Models (RRMs) perform inference by iteratively refining a persistent state in a latent space. This paradigm decouples inference depth from model parameters, allowing small models to perform complex computations by repeatedly applying the same set of parameters. However, their fundamental “deterministic” limitation leads to poor performance on complex tasks with multiple valid solutions (like the N-Queens problem) or those prone to getting stuck in local optima.
GRAM addresses this by introducing learnable stochasticity at each recursive step. Specifically, at every transition, the model samples a “random guidance signal” from a Gaussian distribution dependent on the current state and combines it with a deterministic update to generate the next latent state. This mechanism empowers the model to “bifurcate” at critical decision points, forming a multi-path exploration of the entire solution space rather than proceeding along a single predetermined track.
To make the random guidance more effective, GRAM employs a hierarchical latent state design, comprising a high-level component for abstract reasoning states and a low-level component for fine-grained computations. Stochasticity is injected only into the high-level component, thereby guiding the reasoning direction at a macro level without interfering with the precise calculations at the lower level.
Dual-Axis Scaling: A New Inference Strategy with Depth and Width
One of GRAM’s most significant contributions is the proposal of a “dual-axis scaling” computation strategy at inference time, which redefines how models utilize computational resources.
Depth Scaling (Serial): This is the traditional approach for recursive models, involving an increase in the number of recursive iterations to allow the model to “think” deeper. GRAM also supports this method and allows each reasoning trajectory to stop adaptively based on its needs.
Width Scaling (Parallel): This is GRAM’s unique advantage. The model can independently sample multiple reasoning trajectories from a prior distribution and compute them in parallel. After all trajectories are complete, the best answer is selected through majority voting or a specially trained Latent Process Reward Model (LPRM).
The significance of this parallel scaling strategy is its ability to effectively bypass the high-latency bottleneck associated with depth scaling. In the same amount of physical time, the exploration efficiency of computing multiple shallower reasoning paths in parallel far exceeds that of computing a single, extremely deep path. This concept aligns with the parallelization frameworks recently emerging in visual language models and image generation, such as Visual Para-Thinker and FlashAR, all pointing towards an industry trend of moving from serial to parallel processing to enhance efficiency.
Experimental Validation: Performance Leaps Across Multiple Tasks
The research team comprehensively evaluated GRAM on various tasks, including structured reasoning, multi-solution constraint satisfaction, and unconditional generation, with results demonstrating its effectiveness.
In the challenging Sudoku-Extreme benchmark, GRAM achieved a 97.0% accuracy with only 16 supervised steps and 20 parallel samples. In contrast, the leading deterministic baseline model, TRM, reached only 90.5% accuracy after 320 deep recursive steps. This indicates that GRAM’s “depth + width” combined strategy is far superior in computational efficiency to simply stacking depth. On the ARC-AGI abstract reasoning challenge, GRAM also consistently outperformed all deterministic recursive baselines.
In multi-solution tasks like the N-Queens problem and graph coloring, GRAM’s advantages were even more pronounced. For the 8×8 N-Queens problem, deterministic models could only converge to a single solution, with a maximum solution coverage of 36.1%. GRAM, while maintaining high accuracy, achieved near-optimal solution coverage, successfully balancing the constraint satisfaction capability of recursive refinement with the diversity of generative models.
Furthermore, GRAM demonstrated outstanding unconditional generation capabilities. On the task of unconditionally generating Sudoku puzzles, a 10.9M parameter GRAM model achieved a 99.05% validity rate with 16 supervised steps, surpassing larger diffusion models. When generating MNIST handwritten digits, GRAM also avoided the mode collapse problem seen in the deterministic model TRM, with generation quality comparable to specialized generative models.
Core Insights and Future Directions
Through exhaustive ablation studies, the research team confirmed that GRAM’s performance gains do not stem from simple random perturbations. Removing either the stochasticity or the learned guidance direction led to a sharp decline in model performance. This demonstrates that its success originates from the synergistic effect of random exploration and goal-oriented guidance mechanisms within a variational inference framework.
In summary, GRAM proposes a clear and validated design principle for AI reasoning systems: inference should be conducted in a latent space, in a probabilistic recursive manner, scaling along both depth and width dimensions. This “probabilistic multi-trajectory recursion” paradigm not only offers a new path to overcome the bottlenecks faced by current large models in complex reasoning tasks but also points the way toward building more efficient and flexible general intelligent systems in the future.
