Probabilistic Tiny Recursive Model

Abstract: Tiny Recursive Models (TRM) solve complex reasoning tasks with a fraction of the parameters of modern LLMs by iteratively refining a latent state and final answer. While powerful, their deterministic recursion can lead to convergence at suboptimal solutions, without escape mechanism. A common workaround relies on task-specific input perturbations at test time combined with answer aggregation via voting. We introduce Probabilistic TRM (PTRM), a task-agnostic framework for test-time compute scaling that addresses this limitation through stochastic exploration. PTRM injects Gaussian noise at each deep recursion step, enabling parallel trajectories to explore diverse solution basins, and selects among them using the model's existing Q head (used for early stopping in the original TRM). Without requiring retraining or task-specific augmentations, PTRM enables substantial accuracy gains across benchmarks, including Sudoku-Extreme (87.4% to 98.75%) and on various puzzles from Pencil Puzzle Bench (62.6% to 91.2%). On the latter, PTRM achieves nearly double the accuracy of frontier LLMs (91.2% vs. 55.1%) at less than 0.0001x the cost, using only 7M parameters.

• The paper introduces a test-time stochastic extension to Tiny Recursive Models by injecting Gaussian noise, enabling parallel exploration of latent solution basins.
• It employs width scaling via multiple rollouts and a Q head verifier to select the best answer without requiring retraining.
• Empirical evaluations demonstrate significant accuracy improvements over deterministic TRMs and competitive LLM ensembles on high-complexity puzzles.

Probabilistic Tiny Recursive Model: Test-Time Compute Scaling for Iterative Reasoning

Introduction and Motivation

The "Probabilistic Tiny Recursive Model" (PTRM) (2605.19943) addresses a key limitation in Tiny Recursive Models (TRMs), which perform efficient, iterative reasoning for complex puzzles using a highly compact architecture. While TRMs have achieved impressive results on high-complexity tasks with significantly fewer parameters than contemporary LLMs, their standard deterministic recursion renders them vulnerable to local basin trapping in latent space, resulting in suboptimal solution accuracy. In contrast, autoregressive models and LLMs access diverse solution modes via sampling—the lack of such an exploration mechanism in deterministic TRMs constrains accuracy. PTRM introduces stochasticity at inference time by injecting Gaussian noise into each latent recursion step. This enables parallel exploration of multiple solution basins, with the highest-confidence solution selected by the trained Q head, originally used only for early stopping via ACT during training.

Model and Methodology

PTRM operates as a pure inference-time extension to TRMs, requiring no retraining or task-specific augmentation. Given an input instance, $K$ rollouts (width scaling) are performed in parallel, each injecting Gaussian noise ($\sigma$) at every latent step. Each rollout yields a candidate answer and associated Q value; the highest-Q answer is reported as final output.

Figure 1: The PTRM inference pipeline introduces $K$ noisy parallel recursions at each step, deviating from TRM’s single deterministic path; the Q head acts as a verifier, selecting the best final answer.

This paradigm is distinct from prior input-perturbation or ensemble-based baselines by being task-agnostic, model-agnostic (usable on any pretrained TRM), and not requiring potentially costly retraining or difficult hyperparameter tuning beyond the choice of $K$, $D$, and $\sigma$.

Failure Analysis in Deterministic TRMs

A mechanistic diagnosis identified three canonical solution trajectories in TRM latent space: quick success, delayed success, and failures. Importantly, failures and delayed successes initially both reside in "bad" basins of latent space, only diverging if a late escape into a correct region occurs.

Figure 2: Three classes of TRM trajectories highlight divergences in latent evolution (PCA projection, Q value, and cell accuracy), showcasing how local minima block or delay convergence.

Aggregated over numerous puzzles, the Q head's output is highly predictive of solution quality, with a sharp separation between correct and incorrect trajectories. However, without a mechanism to sample diverse trajectories, deterministic TRM inference is unable to exploit latent solution diversity already present in its parameterization.

Stochastic Rollouts and Escape from Bad Basins

PTRM's parallel stochastic exploration empirically demonstrates that even when deterministic TRM consistently fails on specific instances, a small fraction of stochastic rollouts reach a distant, correct solution basin.

Figure 3: Most PTRM rollouts remain trapped in failed latent regions, but a non-negligible subset escape and yield correct answers, as visible in PCA projections.

This result affirms that the underlying model possesses latent solution modes accessible only via parallel stochastic search. Moreover, as $K$ increases, the observed pass@$K$ and best-Q@$K$ metrics rise monotonically, with the Q head maintaining strong selection fidelity.

Figure 3: pass@$K$ and best-Q@$\sigma$0 show substantial gains as rollout width ($\sigma$1) increases, confirming efficient test-time compute scaling through width.

Crucially, width scaling outperforms naive aggregation (mode@$\sigma$2), confirming that diversity in solution space—combined with competent solution verification—is more effective than simple majority voting.

Experimental Evaluation

PTRM was evaluated on several high-difficulty benchmarks, including Sudoku-Extreme, Maze-Hard, ARC-AGI-2, and the Pencil Puzzle Bench (PPBench) multi-type suite. It consistently and substantially outperforms deterministic TRM and strong LLM baselines/ensembles by a wide margin in accuracy and compute/cost metrics. Notable results include:

• On PPBench, aggregate accuracy improves from $\sigma$3 (TRM) to $\sigma$4 (PTRM, $\sigma$5, $\sigma$6, $\sigma$7), with sudoku subtask accuracy rising from $\sigma$8 to $\sigma$9.
• On Sudoku-Extreme, deterministic TRM accuracy ($K$0) is raised to $K$1 using PTRM.
• On Maze-Hard, accuracy increases from $K$2 to $K$3.
• PTRM outperforms the best single LLM by $K$4 percentage points and a stacked ensemble of 7 SOTA LLMs (with a perfect verifier) by $K$5 points, at less than $K$6 the inference cost.

The robustness of PTRM to various noise levels and width configurations was extensively analyzed:

Figure 4: Task-dependent noise ablation reveals the tradeoff between rollout diversity (pass@$K$7) and selection reliability (best-Q@$K$8); verifier headroom is visible as the gap between the two.

The Q head is generally a strong verifier on well-structured tasks (e.g., Sudoku-Extreme), but on data with more complex or less separable solution criteria, additional verification improvements remain possible.

Theoretical and Practical Implications

PTRM demonstrates that much of the accuracy headroom in small, recursive reasoning models can be recovered via test-time compute scaling. By operationalizing width as a parallelizable axis for solution diversity, and leveraging existing verifiers, PTRM achieves nearly twofold accuracy improvements without any architectural or data changes.

This has several implications:

1. Model Size vs. Compute Tradeoffs: PTRM amplifies the utility of compact models in resource-constrained environments by exploiting parallel inference as an alternative to model scaling.
2. Latent Space Structure: The underlying multi-modality in solution basins, revealed only by stochastic search, generalizes to a broad class of recursive computation models.
3. Verifier Strength: The efficacy of PTRM is bounded above by the strength of the verifier (Q head); further advances are possible via improved latent-correctness estimation mechanisms.
4. Task Generality: The methodology is agnostic to the supervised task, relying only on latent refinement and joint correctness prediction, and does not require domain-specific perturbations or external oracles.

Future Directions

Outstanding questions concern the generalization of PTRM to problem domains where latent-correctness correlation is weaker (as on ARC-AGI-2 and some PPBench variants), and the theoretical characterization of basin structure in more expressive or noisy settings. Research into adaptive verifier heads and more structured stochastic exploration (e.g., Langevin-guided sampling, which in this work showed no further gain over noise) offer further improvements. Extension to language modeling and other generative, non-unique-answer domains is another critical direction.

Figure 5: Visualizing Q-gradient directions in latent space reveals that both good and bad basins exhibit guidance vectors toward correct regions, inspiring—but not currently improving on—PTRM’s simpler random exploration.

Conclusion

PTRM establishes that large accuracy gains—often unreachable through training or deeper, sequential recursions alone—can be systematically unlocked at inference by converting deterministic recursive models into probabilistic, width-scalable systems, provided competent trajectory verification exists. The approach is lightweight, effective, and compatible with existing recursive architectures, enabling small models to match or surpass much larger systems under fair compute budgets. The interplay of exploration, selection, and latent-space geometry formalized by PTRM is likely to inform future research into reasoning architectures with explicit trajectory diversity and autonomous, self-verifying computation.