Github
Cover art for Energy-Based Reasoners Outperform Sudoku Solvers

Energy-Based Reasoners Outperform Sudoku Solvers

Energy-Based Reasoners solve Sudoku with 99.9% accuracy using 345× less compute than top models.

6 cards · 1m · tap to begin

From · · 1m

Energy-Based Reasoners Outperform Sudoku Solvers

Energy-Based Reasoners solve Sudoku with 99.9% accuracy using 345× less compute than top models.

In brief

Energy-Based Reasoners solve Sudoku with 99.9% accuracy using 345× less compute than top models. Energy-Based Reasoners achieve 99.9% Sudoku accuracy with 345× fewer inference FLOPs than EqR. Originally reported by Github.

Quick answers

What are Energy-Based Reasoners?
Energy-Based Reasoners learn an energy function to guide reasoning state refinement and stopping.
How do EBRs work?
EBRs use gradient descent on an energy function to find low-energy states that solve problems.
What is the key takeaway from Energy-Based Reasoners Outperform Sudoku Solvers?
Energy-Based Reasoners achieve 99.9% Sudoku accuracy with 345× fewer inference FLOPs than EqR.

What are Energy-Based Reasoners?

What are Energy-Based Reasoners?

Energy-Based Reasoners learn an energy function to guide reasoning state refinement and stopping.

“E θ (z; x) ∈ R”

How do EBRs work?

Learned energy landscape for one Sudoku puzzle

EBRs use gradient descent on an energy function to find low-energy states that solve problems.

“z t+1 = Π M (z t - α ∇ z E θ (z t ; x))”

EBRs solve Sudoku with 99.9% accuracy

Energy, gradient norm and residual during descent, solved vs wrong

EBRs achieve 99.9% accuracy on Sudoku-Extreme while using 345× less inference compute than EqR.

“99.9% accuracy on Sudoku-Extreme”

EBRs outperform other models

Sudoku-Extreme accuracy vs iterations with one sample (left), and vs number of samples (right)

EBRs achieve 99.9% single-sample accuracy with 5 passes versus 128 samples for EqR.

“99.9 (5) vs 99.8 (128)”

EBRs handle multiple solutions

Sudoku-Extreme exact accuracy vs inference FLOPs per puzzle

EBRs find 98.1% of distinct solutions for 8x8 N-Queens with 20 samples.

“98.1 for 8x8 N-Queens”

Why EBRs are better

EBRs stop trajectories when gradient falls below threshold, reducing compute without sacrificing accuracy.

“gradient norm and residual during descent”

The key point

Energy-Based Reasoners achieve 99.9% Sudoku accuracy with 345× fewer inference FLOPs than EqR.

Read the original on Github

Read the TL;DR — this story on one card.

How we make summaries

React

Sign in to react and comment.

Comments (0)

Life is short. Keep it sweet. Respect others' opinions and be kind!

    Recommended next

    More decks on energy-based reasoners and related topics.