Skip to the result

ResearchTechnical note 3

Learned at 32 bits. Exact at 4,096.

Neural networks tend to fail on inputs longer than the ones they were trained on. METRON learned its laws on 32-bit numbers and then answered exactly on numbers 128 times wider, with nothing relearned.

Measured
13 November 2026
Learned at
32 bits
Tested up to
4,096 bits
Record
116

At a glance

56,256

exact answers across addition, subtraction, comparison and multiplication.

128×

wider than anything seen during learning, with no relearning.

0

laws changed between 32 and 4,096 bits. They are frozen and fingerprinted.

In plain words

A child who learns to add two-digit numbers can add twenty-digit numbers the same way: the rule does not change, only the length. Most neural networks cannot do this. METRON can, because what it learns is the rule itself.

For engineers

Bit-level laws for addition, subtraction, comparison and the stages of multiplication were learned at w = 32 and frozen. They were then applied at w = 64, 128, 256, 1,024 and 4,096 through a structure declared in advance, on uniform, hard and publicly drawn operands.

Why it matters

Learn once, deploy at any size. The cost of learning does not grow with the size of the problem, which is the opposite of how model training scales today.

MTRMETRON is AIDON’s exact co-processor. It learns laws from data, proves them and runs them, alongside a neural model or on its own.

The bottleneck

Trained on short inputs, wrong on long ones.

Length generalisation is one of the best-documented weaknesses of neural networks. A model that is accurate on the lengths it was trained on degrades, often to zero, as the input grows beyond them.

The usual remedy is more data and a larger model for every new length. That is a recurring cost, and it never reaches certainty.

Result

One lesson at 32 bits. Exact up to 4,096.

Never seen during learning0%50%100%32641282561,0244,096Learned here

Operand width in bits, logarithmic scale

Fig. 1 Share of exact answers at each operand width, the four operations together. The laws were learned at 32 bits and frozen; every width to the right was answered without relearning anything.

Table 1 Exact answers out of 2,088 per cell: 1,024 uniform operands, 40 hard cases and 1,024 operands drawn from a public randomness beacon.
Operand widthAdditionSubtractionComparisonMultiplicationWidth vs learning
32 bits2,0882,0882,0882,0881×
64 bits2,0882,0882,0882,0882×
128 bits2,0882,0882,0882,0884×
256 bits2,0882,0882,0882,0888×
1,024 bits2,0882,0882,0882,08832×
4,096 bits2,0882,0882,0882,088128×

Across both public draws of the test, the sealed total is 56,256 exact answers out of 56,256.

Method

How the laws are reused

  1. Learnthe laws, from 32-bit examples
  2. Freezethe laws are fingerprinted and never touched again
  3. Applythe same laws serve every wider number
  1. The laws are learned once, from examples of 32-bit operations.

  2. They are frozen. From this point on, no parameter of the system changes.

  3. They are applied at every width, through a structure declared before the test. Nothing is tuned to the new sizes.

Controls

What was checked

Public randomness

Part of the operands came from a public randomness beacon round published only after the test had been registered. Nobody, including us, could have chosen them.

Hard cases

40 hard operand pairs per cell, chosen to stress the operation, all answered exactly.

Altered laws

With the laws deliberately altered, errors appear in every operation. The test can fail, and it detects when it should.

Reproduction first

Before the test, 35 earlier results were reproduced and matched one by one.

Open questions

What this opens

  • Widths beyond 4,096 bits and operations beyond the four measured here.
  • The same learn-once behaviour for structures that are not arithmetic.

Fields of application

Where it fits

  • Wide-integer arithmetic for cryptography, where 2,048 and 4,096 bits are everyday sizes.
  • Processors and accelerators whose word size changes from one product to the next.
  • Any system that must keep working when the input outgrows the training data.

Record

How this result is kept

Record
116
Run
13 November 2026, 39 min 13 s
Sample
6 widths; 4 operations; 2,088 operand pairs per cell, plus a second public draw
Measurement file
SHA-256 6fcdfe2054326d832ef222fcdaeb11fc324de08d297666c7d4520a026188c4fa

Every figure on this page was read by a program from the raw measurement file, not typed. The fingerprint identifies that file: when it is disclosed, anyone can check that it is the same one.

Contact

Talk to the founders

Get in touch. We would be glad to share our ideas and walk you through the results, the proof and the measurements.