When Recursive Models Finish Computing

TL;DR

Recursive models extended to 512 steps on Sudoku, accuracy increased from 59.2% to 87.5%.

cs.LG 🔴 Advanced 2026-09-22 9 views
Hare Krishna Shubham Singh Stephen Ebert Hao-Yu Sun
Recursive Models Sudoku Attention Mechanism MLP Dynamic Stability

Key Findings

Methodology

The study uses attention and MLP-based Tiny Recursive Models (TRMs) to test 1,000 hard Sudoku puzzles. By extending recursion from 16 to 512 steps, it examines completion dynamics and local stability, particularly trajectory-conditioned anisotropic stability.

Key Results

  • Attention model accuracy increased from 59.2% to 87.5%, MLP model from 74.4% to 91.9%.
  • Of puzzles unsolved at 16 steps, 69.4% were solved by the attention model, 68.4% by the MLP.
  • Completed states are contractive along the trajectory, though local Jacobian retains expanding directions.

Significance

The study reveals the potential for incomplete computation within inference budgets, offering new methods to identify completion states during reasoning. This is significant for understanding and optimizing recursive model dynamics.

Technical Contribution

Introduces the concept of trajectory-conditioned anisotropic stability, unveiling directional stability post-completion in recursive models. This provides a new perspective for dynamic analysis and may influence future model designs.

Novelty

First systematic study of trajectory-conditioned anisotropic stability in recursive models, distinguishing nominal-budget failure from completed computation dynamics.

Limitations

  • Study limited to Sudoku puzzles, may not apply to other tasks.
  • Models retain locally expanding directions post-completion, potentially affecting long-term stability.

Future Work

Future work could explore recursive model dynamics on other tasks and investigate leveraging trajectory-conditioned anisotropic stability to enhance performance.

AI Executive Summary

Recursive models solve complex problems by iteratively updating their latent states. However, incorrect outputs at the end of the inference budget may indicate unfinished computation or persistent failure.

This study tests attention and MLP-based Tiny Recursive Models (TRMs) on 1,000 hard Sudoku puzzles. By extending recursion from 16 to 512 steps, attention model accuracy increased from 59.2% to 87.5%, and MLP model from 74.4% to 91.9%. This suggests many puzzles unsolved within the nominal budget are actually unfinished computations.

The study finds that completed states are contractive along the trajectory direction, even though the local Jacobian retains expanding directions. This phenomenon, termed trajectory-conditioned anisotropic stability, reveals directional stability post-completion in recursive models. This finding provides a new perspective for dynamic analysis and may influence future model designs.

Deep Analysis

Background

Recursive models compute by iteratively updating latent states, widely used in solving complex problems. However, incorrect outputs at the inference budget's end may indicate unfinished computation or persistent failure. Recently, researchers have focused on the dynamic behavior and stability of recursive models.

Core Problem

Incorrect outputs at the end of the inference budget may indicate unfinished computation or persistent failure. Identifying the completion state of computation is a crucial research problem.

Innovation

The study introduces the concept of trajectory-conditioned anisotropic stability, revealing directional stability post-completion in recursive models. This finding provides a new perspective for dynamic analysis.

Methodology

  • �� Use attention and MLP-based Tiny Recursive Models (TRMs)
  • �� Test on 1,000 hard Sudoku puzzles
  • �� Extend recursion from 16 to 512 steps
  • �� Study completion dynamics and local stability

Experiments

Experiments conducted on the Sudoku-Extreme test set, containing 1,000 hard Sudoku puzzles. Use attention and MLP models, compare performance at 16 and 512 recursion steps, analyze completion dynamics and local stability.

Results

Attention model accuracy increased from 59.2% to 87.5%, MLP model from 74.4% to 91.9%. Completed states are contractive along the trajectory, though local Jacobian retains expanding directions.

Applications

The study's results can optimize recursive model performance on complex problems, particularly in tasks requiring prolonged reasoning.

Limitations & Outlook

Study limited to Sudoku puzzles, may not apply to other tasks. Models retain locally expanding directions post-completion, potentially affecting long-term stability.

Plain Language Accessible to non-experts

Imagine solving a very complex puzzle, seeing only a small part at a time. A recursive model is like a smart assistant, constantly updating its understanding of the puzzle until it finds a solution. Even if it doesn't find the answer within the budgeted time, it may just need more time to complete the computation.

ELI14 Explained like you're 14

Imagine playing a super hard Sudoku game! You have an assistant that helps you solve it step by step. Even if time runs out, it might not be done yet. This paper studies how these assistants change while solving, finding they sometimes need more time to finish the task!

Glossary

Recursive Model

A model that computes by iteratively updating its latent state.

Used for solving complex problems like Sudoku.

Attention Mechanism

A mechanism that selectively focuses on important information when processing sequence data.

Used to improve model performance on Sudoku.

MLP (Multilayer Perceptron)

A neural network structure that learns through multiple layers of nonlinear transformations.

Serves as the foundational structure for Sudoku-solving models.

Trajectory-Conditioned Anisotropic Stability

States post-completion are contractive along the trajectory direction, though local Jacobian retains expanding directions.

Used to analyze the dynamic behavior of recursive models.

Local Jacobian

Describes the local linear approximation of state changes.

Used to analyze model's local stability.

Open Questions Unanswered questions from this research

  • 1 How to validate trajectory-conditioned anisotropic stability on other tasks?
  • 2 Are there methods to accelerate the completion of recursive model computations?

Applications

Immediate Applications

Complex Problem Solving

Recursive models can be used to solve complex problems like Sudoku, requiring adjustment of recursion steps to improve accuracy.

Long-term Vision

Intelligent Reasoning Systems

Dynamic analysis of recursive models can be used to design more intelligent reasoning systems, needing to address long-term stability issues.

Abstract

Recursive models can continue updating their latent states beyond their nominal inference budget, so an incorrect output at that budget does not show whether computation is unfinished or has entered a persistently unsuccessful regime. We study the dynamics of completion in attention- and MLP-based Tiny Recursive Models (TRMs) on 1,000 hard Sudoku puzzles. Extending recurrence from the nominal 16 steps to 512 steps increases cumulative exact-solve accuracy from 59.2% to 87.5% for the attention model and from 74.4% to 91.9% for the MLP model, solving more than two-thirds of the puzzles unsolved in the nominal budget. Across both architectures, latent-state motion drops sharply after the first exact solution. Completed states are typically locally contractive along the trajectory direction, even though the same local Jacobian retains strongly expanding directions. We characterize this phenomenon as trajectory-conditioned anisotropic stability. Perturbation experiments confirm this directional stability across both models. The multi-step fate of the maximally expanding direction differs: it is absorbed within 16 steps in the attention model but persists longer in the MLP model. The anisotropic-stability pattern also holds for a second attention checkpoint. Together, these results distinguish nominal-budget failure from completed computation and identify a common dynamical signature of completion across two recurrent architectures.

cs.LG cs.AI cs.NE