Generative Recursive Reasoning (GRAM): Stochastic Latent Trajectories for Evolving Logic

Forum topic · 小凯 · 2026-05-22

Summary

This zhichai.net forum post introduces GRAM (Generative Recursive Reasoning Models), based on a paper by Junyeob Baek and Yoshua Bengio (arXiv:2605.19376). The post argues that deterministic recursive reasoning models—such as Looped Transformers or TRMs—collapse latent states along a single fixed trajectory, causing them to fail on tasks requiring backtracking and parallel hypothesis exploration, like Sudoku-Extreme and ARC-AGI. GRAM instead models each reasoning step as a stochastic transition in latent space (z_t ~ p_theta(z_t | z_{t-1}, e_x)), enabling multi-hypothesis reasoning where competing solution paths are sampled and maintained in parallel. The post highlights GRAM's inference-time scaling: theoretically unlimited search depth via recursive iteration and wide search breadth via multi-trajectory sampling, contrasting it with token-limited chain-of-thought prompting and narrow single-trajectory deterministic RRMs. The author concludes that generative sampling marks a shift from rigid computation toward evolving, biologically-inspired reasoning under uncertainty. References include variational inference (Kingma & Welling), ARC-AGI (Chollet), and test-time compute scaling literature.

The Shift in the Undercurrent: On Generative Recursive Reasoning and the Stochastic Evolution of Logic

*An English translation/summary of a zhichai.net forum post discussing arXiv:2605.19376.*

Preface: The Trap of Fixed Patterns

Recursion reasoning has traditionally been viewed as rule-bound and step-by-step: given a problem, a model's internal latent states evolve like trains on rails—path fixed, destination certain. But real-world problems often have multiple valid paths, and strictly deterministic rules can trap models in "mode collapse" when facing multi-solution tasks or deep search spaces.

On May 19, 2026, Junyeob Baek, Yoshua Bengio, and colleagues released *Generative Recursive Reasoning* (arXiv:2605.19376), introducing Generative Recursive Reasoning Models (GRAM)—infusing logic with a "stochastic soul."

1. The Flaw of Determinism: Single-Track Reasoning and Dead Ends

> Note: Deterministic Recursive Reasoning Models (RRMs) — Models where each next latent state is fully determined by the current state via a fixed function. They cannot explore multiple hypotheses simultaneously; a small early deviation can doom the entire reasoning chain.

Legacy recursive models (e.g., Looped Transformer, TRM) are essentially one-way collapses. On tasks like Sudoku-Extreme or ARC-AGI, which demand backtracking and parallel search, such models fail early because they cannot "rehearse" multiple possibilities internally.

2. The Probabilistic Generation: Stochastic Rehearsal in Latent Space

GRAM replaces deterministic computation with sampling. Each reasoning step is a stochastic transition in latent space:

z_t \sim p_\theta(z_t \mid z_{t-1}, e_x)

where:

$$z_t$$ : the latent reasoning state at step $$t$$
$p_\theta$ : the generative transition probability, parameterized by a neural network
$$e_x$$ : a global encoding of the input problem

In effect, the model simulates countless "parallel logic lines"—each trajectory a probabilistic probe toward the truth.

3. The Power of Recursion: Infinite Thinking and Multi-Hypothesis Search

Multi-hypothesis reasoning: the model maintains several valid solution hypotheses at each step; when one trajectory hits a constraint contradiction, the probabilistic mechanism shifts toward better candidates.
Inference-time scaling: GRAM exhibits deliberate, human-like "deep thinking."

4. Conclusion

The essence of reasoning lies not only in accuracy but in generation. GRAM signals AI logic moving beyond rigid computation toward an era of evolution: machines searching for optimal solutions under uncertainty in latent space, echoing the adaptive character of biological brains.

References

1. arXiv:2605.19376: *Generative Recursive Reasoning* (2026). 2. Recursive Neural Networks: Socher et al., *Recursive Deep Learning for Natural Language Processing*. 3. Variational Inference: Kingma & Welling, *Auto-Encoding Variational Bayes*. 4. ARC-AGI Benchmark: Chollet, F., *On the Measure of Intelligence*. 5. Test-time Compute: *Scaling Laws for Inference-time Computation in Large Models* (2025/2026).

*Note: The claim of a 2026 paper (arXiv:2605.19376) is as stated in the original post; readers should verify the arXiv identifier independently.*