Can we train a neural network to be easier to correct?

Start here

About 60 seconds

The idea, in a minute.

The question: Can practicing corrections teach a model to apply new information without breaking unrelated predictions?
What we found: Training for editability reduced consequence error by 49.5% and unrelated-prediction error by 62.1% versus ordinary training.
Why it matters: A model can fit a corrected fact and still misunderstand its consequences. Useful editing needs both transfer and preservation.

Keep in mind

Three seeds in a synthetic circuit world; the simpler first-order control captured almost all of the consequence gain.

See the evidence

A concrete example

Fitting evidence can worsen its consequences

Recorded aggregate comparison from the fixed-budget experiment.

01
Before correction
Ordinary continuation had new-consequence RMSE 0.1165 without editing.
02
Fit the correction
Five SGD steps fitted the direct evidence, but consequence RMSE became 0.1283.
03
Practice editing first
The exact edit-trained condition reached consequence RMSE 0.0648 under the same five-step budget.

Correct support examples do not by themselves establish a useful edit.

Archived experiment reports and measurements

Read the formal abstract

Correcting a model is easy to measure on the examples used to correct it. The harder question is whether the change reaches related situations and leaves unrelated behavior intact. We trained a small transformer to practice this entire process rather than only predict the original world.

Across three models and 90 held-out edits, training through an update reduced consequence RMSE from 0.1283 to 0.0648 and unrelated-world RMSE from 0.1801 to 0.0682. However, a first-order control reached consequence RMSE 0.0652. Exact differentiation added little consequence benefit, although preservation improved. The experiment supports training for editability in this setting, while leaving substantial residual damage and no demonstrated language-model-scale result.

01 Research question

Can practicing corrections teach a model to apply new information without breaking unrelated predictions?

02 Experiment setup

The full procedure, model settings, datasets, and evaluation choices are recorded here.

Explore the complete methodology

Executed September 16, 2026 on an Apple M3 Pro with 36 GB memory and one CPU thread. The two-block transformer has 39,121 parameters. A simulator supplies labels for twelve artificial materials whose conductivity follows a temperature-dependent sigmoid, plus individual, series, and parallel queries. The simulator is absent from inference; the network contains no hand-written rule routing.

Three seeds, 0–2, share their respective 2,000-step AdamW pretrained starts across pretrained-only, ordinary continuation, first-order edit training, and exact through-update training. A separate seed 999 is used for development. Both edit-training arms receive 1,200 episodes and objective weights 4/1/2 for original-world, consequence, and preservation losses. The first-order arm detaches the update delta; the exact arm differentiates through one SGD step.

Edit-training targets materials 0–7, development targets 8–9, and final edits target 10–11. All materials appear in original prediction training. Thirty edits per seed provide sixteen direct examples and independent sets of 192 consequence and 192 unaffected queries. The primary fixed-budget comparison applies five SGD steps at learning rate 0.1 to every condition, changing all parameters.

A second comparison stops at correction RMSE at most 0.03 within twenty steps. A secondary stress test applies ten sequential changes without resetting weights. Ordinary continuation is approximately wall-time matched, rather than matched on data exposure or exact arithmetic operations. Archived code, configurations, per-edit measurements, and all twelve final/pretrained checkpoints are included.

03 What happened?

Training for editability reduced consequence error by 49.5% and unrelated-prediction error by 62.1% versus ordinary training.

Figure 01 / Observed resultsGeneralizing a correction

49.5% lowerConsequence error versus ordinary continuation

0.150.0750

0.1283

Ordinary

0.0652

First-order

0.0648

Through-update

Consequence RMSE ↓ · scale shown from 0 to 0.15

How to read thisThe bars show raw consequence errors. The table also reports unrelated-world errors, so improvement cannot be judged by evidence fitting alone.

The limitThree seeds in a synthetic circuit world; the simpler first-order control captured almost all of the consequence gain.

Fixed budget: five SGD steps at LR 0.1; means over three independently trained models. RMSE is root mean squared prediction error, in the simulator output units.

Full results and data table

Complete summary results
Training	Consequence RMSE ↓	Unchanged-world RMSE ↓
Ordinary	0.1283	0.1801
First-order	0.0652	0.089
Through-update	0.0648	0.0682

At five correction steps, pretrained-only / ordinary / first-order / exact through-update consequence RMSE was 0.1239 / 0.1283 / 0.0652 / 0.0648. Unchanged-world RMSE was 0.1729 / 0.1801 / 0.0890 / 0.0682. Lower is better. Means first average episodes within each model, then average the three models.

Exact through-update training improved consequence RMSE by 49.5% and preservation RMSE by 62.1% versus ordinary continuation. Versus the same-data first-order control, the reductions were only 0.5% and 23.3%. Consequence differences versus first-order had mixed signs across seeds; preservation improved in all three.

All ninety edits reached the shared evidence-fit threshold for every condition. Mean steps were 8.0111 for ordinary, 3.7222 for first-order, and 2.5889 for exact training. At that threshold, consequence/preservation RMSE was 0.1421/0.1946, 0.0649/0.0863, and 0.0633/0.0655 respectively.

Preservation was still imperfect: exact-training unaffected RMSE rose from 0.0131 before editing to 0.0682 after it. Ordinary continuation fitted the evidence while worsening consequence error from 0.1165 without an edit to 0.1283 with one. In the ten-edit stress test, final exact-training consequence/preservation RMSE was 0.0425/0.0387; changing thresholds make the sequence unsuitable as evidence of progressively improving learning ability.

04 What does this mean?

What this suggests

Training for editability reduced consequence error by 49.5% and unrelated-prediction error by 62.1% versus ordinary training.
Training representations for future corrections is more promising here than merely extending ordinary prediction training. The same-data first-order comparison is essential: structured edit supervision explains most of the consequence gain.

What we still don’t know

Transfer to pretrained language models, unfamiliar rule families, or natural-language evidence is unknown.
Three independent models support descriptive replication, not strong population significance. Structured consequence and preservation supervision favors both edit-training arms over ordinary training.

Read the full discussion

Training representations for future corrections is more promising here than merely extending ordinary prediction training. The same-data first-order comparison is essential: structured edit supervision explains most of the consequence gain.

The test measures behavioral editability, not the emergence of semantic modules or sparse weight changes. The circuit examples offer a controlled way to distinguish remembering the correction from applying it correctly.

05 Limitations

Three seeds in a synthetic circuit world; the simpler first-order control captured almost all of the consequence gain.

Read all study limitations

Transfer to pretrained language models, unfamiliar rule families, or natural-language evidence is unknown.

Three independent models support descriptive replication, not strong population significance. Structured consequence and preservation supervision favors both edit-training arms over ordinary training.