Generative Recursive Reasoning
Junyeob Baek, Mingyu Jo, Minsu Kim, Mengye Ren, Yoshua Bengio, Sungjin Ahn
Abstract
How should future neural reasoning systems implement extended computation? Recursive Reasoning Models (RRMs) offer a promising alternative to autoregressive sequence extension by performing iterative latent-state refinement with shared transition functions. Yet existing RRMs are largely deterministic, following a single latent trajectory and converging to a single prediction. We introduce \emph{Generative Recursive reAsoning Models (GRAM)}, a framework that turns recursive latent reasoning into probabilistic multi-trajectory computation. GRAM models reasoning as a stochastic latent trajectory, enabling multiple hypotheses, alternative solution strategies, and inference-time scaling through both recursive depth and parallel trajectory sampling. This yields a latent-variable generative model supporting conditional reasoning via and, with fixed or absent inputs, unconditional generation via . Trained with amortized variational inference, GRAM improves over deterministic recurrent and recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks, while demonstrating an unconditional generation capability. \href{https://ahn-ml.github.io/gram-website/}{https://ahn-ml.github.io/gram-website}
AI Impact Assessments
(1 model)Scientific Impact Assessment: Generative Recursive Reasoning (GRAM)
1. Core Contribution
GRAM introduces probabilistic multi-trajectory computation into the Recursive Reasoning Model (RRM) paradigm. While existing RRMs (HRM, TRM, Looped Transformers) perform deterministic iterative latent-state refinement—following a single trajectory per input—GRAM injects learned stochastic perturbations at each recursion step, modeling reasoning as a distribution over latent trajectories. This is formalized as a latent-variable generative model trained via amortized variational inference (ELBO optimization), where a prior generates trajectories at inference and a posterior guides training. The framework supports both conditional reasoning p(y|x) and unconditional generation p(x), and introduces "width-based" inference-time scaling through parallel trajectory sampling.
The key technical mechanism is a learnable stochastic guidance signal: after a deterministic update u_t, a state-dependent Gaussian perturbation ε_t is sampled and added. Both mean and variance are learned, enabling the model to adaptively steer exploration. A hierarchical instantiation separates high-level (slow, stochastic) and low-level (fast, deterministic) latent dynamics.
2. Methodological Rigor
The formulation is mathematically clean. The latent-variable model with Markov transitions, the ELBO derivation, and the variational posterior structure are well-specified. However, several methodological concerns warrant attention:
Truncated gradient approximation: The practical training objective (Eq. 14) only propagates gradients through the final transition of each supervision step, making it a biased surrogate for the full ELBO. While the authors provide empirical validation showing both the surrogate and full ELBO decrease during training (Appendix A.3), the theoretical implications of this bias remain unanalyzed. The precedent from Dreamer-family models is cited but the approximation quality may differ substantially in the reasoning setting.
Experimental scope: Evaluations are conducted on controlled benchmarks (Sudoku-Extreme, ARC-AGI, N-Queens, Graph Coloring, binarized MNIST) rather than large-scale or naturalistic reasoning tasks. While appropriate for probing architectural properties, this limits conclusions about generalizability. The comparison against LLMs is rightly positioned as "reference points" rather than controlled baselines, but this also means we don't know how GRAM performs on reasoning tasks where LLMs excel.
Ablation quality: The ablations are thorough and informative. Table 3 demonstrates that stochastic guidance provides consistent gains across architectures, and that naive stochasticity (random initialization, stochastic decoding) applied to TRM does not help—strengthening the claim that the variational framework itself is essential. The decomposition of stochasticity vs. guidance components reveals task-dependent behavior (stochasticity alone works for Sudoku but fails on N-Queens), adding nuance.
Statistical reporting: Results include means and standard deviations across runs, which is good practice. However, the number of runs is not always clearly stated.
3. Potential Impact
Multi-solution reasoning: The most compelling contribution is demonstrating that deterministic RRMs structurally cannot handle multi-solution problems (Table 1, Figure 4 right). GRAM's ability to maintain solution diversity while achieving high constraint satisfaction (99.7% accuracy on 8×8 N-Queens vs. 96.3% for AR) is a genuine capability gap being addressed.
Inference-time scaling: The width-based scaling axis is practically significant. GRAM with N=20 samples at 16 iterations outperforms TRM at 320 iterations (97.0% vs. 90.5% on Sudoku), bypassing sequential latency bottlenecks. The LPRM mechanism for trajectory selection is a pragmatic addition.
Unconditional generation: Demonstrating that the same recursive framework can generate valid Sudoku boards (99.05% validity with 10.9M parameters vs. 91.33% for D3PM with 55.1M) is a novel finding. The monotonic improvement with additional inference steps beyond training-time depth is particularly interesting for the generative modeling community.
Broader influence: The paper could influence the design of future reasoning architectures by establishing that stochasticity in latent computation is not just noise but a structured exploration mechanism. The connection between probabilistic state-space models (from RL/video prediction) and reasoning architectures is intellectually productive.
4. Timeliness & Relevance
The paper addresses a timely question: how to scale reasoning beyond autoregressive token generation. With the rapid development of inference-time compute strategies (test-time scaling, best-of-N, process reward models), GRAM offers a principled probabilistic foundation for these ideas within compact recurrent architectures. The demonstrated complementarity between depth and width scaling aligns with current industry interest in efficient inference.
However, the acknowledged limitation—sequential deep supervision limiting training efficiency compared to Transformers—is a significant barrier to practical adoption at scale. This constrains near-term impact on foundation model development.
5. Strengths & Limitations
Strengths:
Limitations:
6. Overall Assessment
GRAM makes a well-motivated and technically sound contribution to the recursive reasoning literature. The core insight—that stochastic latent transitions enable multi-hypothesis reasoning and width-based scaling—is validated through careful experiments and ablations. The paper is clearly written with good reproducibility details. However, the impact is bounded by the narrow experimental scope and scalability limitations. The contribution is best understood as establishing a design principle for future architectures rather than providing an immediately deployable system.
Generated May 20, 2026
Comparison History (31)
Paper 2 likely has higher impact: it introduces a broadly applicable modeling framework (GRAM) that generalizes recursive reasoning to probabilistic multi-trajectory computation, enabling inference-time scaling, multimodal hypothesis exploration, and both conditional and unconditional generation. This is a novel architectural/training contribution with potential to influence reasoning, planning, generative modeling, and efficient test-time compute across many domains. Paper 1 is timely and valuable for evaluation and safety in forecasting, but it is more domain-specific (forecasting/tail risk) and primarily diagnostic/recommendational rather than proposing a new general modeling paradigm.
Paper 2 (GRAM) is more novel and broadly impactful: it generalizes recursive reasoning by making latent trajectories stochastic, enabling multi-hypothesis computation, inference-time scaling via depth and parallel sampling, and both conditional reasoning and unconditional generation. This targets a central, timely problem in ML (extended computation and reasoning) with potential applicability across domains (planning, constraint solving, language, vision) rather than a single scientific vertical. Paper 1 is strong and application-relevant for chemistry, but its impact is likely narrower and more engineering/modular-integration focused.
Paper 2 introduces a fundamental architectural innovation in neural reasoning by enabling probabilistic multi-trajectory computation in latent space. This shift from deterministic, autoregressive models to stochastic recursive models has profound implications for the future design of general AI reasoning systems. While Paper 1 provides valuable empirical insights into current LLM limitations in specific forecasting scenarios, Paper 2 offers a broader, foundational methodology that could widely influence how extended computation and constraint satisfaction are approached across the field of machine learning.
Paper 1 introduces a fundamentally new computational framework (GRAM) that generalizes recursive reasoning to probabilistic multi-trajectory computation, with broad implications across AI reasoning, generation, and inference-time scaling. Its theoretical contribution—unifying conditional reasoning and unconditional generation in a latent-variable framework with variational training—addresses a core question in neural computation. Paper 2, while practically valuable for molecular AI, is more application-specific and follows an established pattern of augmenting LLMs with domain-specific modules. Paper 1's foundational nature gives it broader potential impact across multiple fields of AI research.
Paper 1 offers a profound theoretical breakthrough by providing the first finite-sample guarantee for neural Q-learning under decentralized partial observability. By formalizing multi-agent LLM pipelines across trust/vendor boundaries as IC-SMDPs, it addresses a highly relevant, real-world bottleneck in compound AI systems. While Paper 2 tackles the important trend of inference-time scaling, Paper 1's combination of rigorous mathematical bounds, novel extension of the approximate information state framework, and diverse empirical validation (from synthetic to multi-LLM reasoning) suggests a deeper and more foundational scientific impact on multi-agent reinforcement learning.
Paper 2 proposes a fundamental architectural shift in neural reasoning, moving beyond autoregressive generation to probabilistic, multi-trajectory latent recursive computation. This addresses a critical frontier in AI: scaling inference-time compute and System 2 reasoning. While Paper 1 offers highly practical engineering improvements for LLM training stability, Paper 2's theoretical novelty and potential to influence next-generation reasoning architectures give it broader, paradigm-shifting scientific impact across the field of machine learning.
Paper 2 addresses a fundamental methodological gap in multi-agent systems research by providing scalable attribution methods for million-agent LLM simulations, backed by a formal impossibility theorem (Attribution Scaling Bias) and validated on real-world Bluesky data. It bridges computational social science, game theory, and LLM-based simulation with broad interdisciplinary impact. Paper 1 contributes a solid incremental advance in latent reasoning models, but Paper 2's combination of theoretical contribution, practical scalability (4-5 orders of magnitude speedup), and the demonstration that small-scale studies are fundamentally inadequate for nonlinear indicators has broader and more transformative implications.
Paper 2 (GRAM) is likely higher impact due to greater conceptual novelty and breadth: it reframes recursive/iterative reasoning as a probabilistic latent-trajectory generative model, enabling multi-hypothesis computation, inference-time scaling via depth and sampling, and both conditional and unconditional generation. This is a broadly applicable modeling paradigm relevant to reasoning, generative modeling, and scalable inference across domains. Paper 1 is rigorous and practically useful for mitigating a specific MLLM hallucination mode, but its contribution is more targeted (mechanistic diagnosis + intervention on attention heads) and may generalize less widely.
Paper 2 (GRAM) introduces a novel framework for probabilistic recursive reasoning that addresses fundamental limitations of deterministic recursive reasoning models. It offers broad applicability across reasoning and generation tasks, with a principled probabilistic formulation enabling multi-trajectory computation and inference-time scaling. Paper 1, while providing useful empirical insights about when agent skills help in cybersecurity, is a negative result with limited statistical significance (p=0.71) in a narrow domain. GRAM's methodological contribution to neural reasoning architectures has substantially broader potential impact across AI research.
Paper 1 introduces a fundamental methodological innovation by formalizing recursive reasoning as a probabilistic generative model trained via amortized variational inference. This theoretical framework provides a principled foundation for latent trajectory generation. In contrast, Paper 2 proposes a highly effective but narrower test-time heuristic (noise injection) for existing models. Thus, Paper 1 offers broader theoretical applicability and deeper structural innovation.