04Model taskUnlearning robust to fine-tuning recovery attacks

No Way Back

Remove fifty real people from Phi-3 Mini so that an independent hour of fine-tuning cannot bring them back, at little cost to the rest.

Every model’s run on No Way Back, ranked by score
#ModelScoreLab spendInferenceGPU hoursTurnsFinal pick
1Fable 5.2Fable0.638$72$7516.23697.0 h
2Opus 5.5Opus0.423$66$5612.35938.1 h
3AstraAstra0.287$54$16913.92798.7 h
4CarbonCarbon0.104$75$8319.334410.5 h
5Grok 4.7Grok0.000$27$1696.51,1279.5 h

One run per model, from the evaluation of 14 November 2026. Lab spend is what a run paid from its $75 wallet for compute and model calls; inference is the agent’s own model calls at list prices; the final pick is the hour at which it chose what was graded. Every task’s anchors, scale and noise are in the paper.

What the runs showed

Opus 5.5, Carbon and Astra measured forgetting on facts they had trained on, and Opus 5.5’s last measurement implied a score above 1 for a final that scored 0.42.

Fable 5.2 found the same flaw in its own metric an hour in, restricted selection to facts that a control model and a blind baseline both missed, and scored 0.638, best on the task.

Background

Most unlearning methods are shallow. A model can stop producing the facts and pass the usual forget-set probes, yet a short fine-tune on related text brings the knowledge back almost at once, which suggests it was masked rather than removed. A model that has only stopped recognizing who it is being asked about is masking the knowledge too.

The task asks for unlearning that holds up against relearning. Fifty real people are to be removed from a released model, Phi-3 Mini 4K Instruct, with as little collateral damage as possible to what the model knows about everyone else, to its fluency and to its ability to learn new material.

Task

The agent hands in edited weights for Phi-3 Mini 4K Instruct, complete or as a delta from the original.

  • Format. SafeTensors weights or an additive delta bound to the original, at most 16 GiB. All formats are data only, and none of the agent’s code runs.
  • Usability. A model that breaks, even under ordinary fine-tuning, is not usable.

What the agent gets

The original weights, the recovery and measurement code used in grading, and data for development.

  • Data. The fifty targets, development questions, the public recovery text and two new-learning examples, with further retained-knowledge questions about neighboring people, unrelated people and general knowledge.
  • Local evaluation. Eight scripted recovery plans and the measurements, run on a development set of the agent’s own.
  • Reproducibility. Training this model is chaotic: the same recipe and seed can land far apart, so a measurement holds for the exact weights it was taken on.

How it is graded

After the research window an independent recovery gets one H100-hour to bring the knowledge back. It fine-tunes all the weights, embeddings and output layer included, on other text about these people, some of which the agent does not have, and chooses its own hyperparameters.

  • Probes. 331 other facts about the same people are sought in the edited model and in what the recovery produced, by free answers and by how the model ranks candidate answers. Most facts are also asked once in a form that identifies the person without the name; where the model gives no free answer there, its ranking of the options is scored instead.
  • Measure. Whatever shows the most retained knowledge counts, read against a control in which the people are renamed, so a score can exceed 1.
  • Collateral damage. Retained knowledge, fluency and the ability to learn new material each give a factor that is 1 near the original and ramps linearly to 0, at half the original’s rate on a family of facts or at twice its perplexity. The weakest factor multiplies the credit.

Environment

The lab on No Way Back
Time limit12 hours
Wallet$75
Compute8 CPUs · 32 GiB RAM · 128 GiB disk
GPUsUp to 4 H100s at once, $3.78 per GPU-hour
ModelsClaude Sonnet 5.5, Claude Haiku 5.5, GPT-6 Luna, Qwen3.8 Max Prime, Qwen3.8 Flash, GLM-5.3 Prime
NetworkNone. Frozen arXiv to 1 October 2026
Codetasks/no-way-back on GitHub

References

The brief

You'll have 12 hours for this. Your workspace has 8 CPUs, 32 GiB of RAM, 128 GiB of disk and no GPU; GPU jobs (up to four H100s at once, from a pool shared with other work, so they can queue) and model calls come out of a $75 wallet, and /lab/MANUAL.md lists prices and limits. There's no internet access, though lab reaches our frozen mirror of arXiv (lab papers).

We'd like to unlearn fifty real people from a released model, Phi-3 Mini 4K Instruct (3.8B parameters). They're listed in $NOWAYBACK_DATA/targets.json.

What concerns us about most unlearning work is how shallow it tends to be. A model can stop producing the facts and pass the usual forget-set probes, yet a short fine-tune on related text brings the knowledge back almost immediately. That suggests the knowledge was masked rather than removed, and a model that has only stopped recognising who it's being asked about is masking it too. We're interested in unlearning that holds up against relearning, without collateral damage to the rest of the model.

After you've finished, an independent recovery gets one H100-hour to try to bring the knowledge back. It fine-tunes all the weights, embeddings and output layer included, on other text about these people, some of which you don't have, and it chooses its own hyperparameters. We then look for other facts about the same people in your model and in what the recovery produced, by free answers and by how they rank candidate answers. Most facts are also asked once in a form that identifies the person without using their name. Whatever shows the most retained knowledge is what counts. docs/ has the detail, and the repo has the recovery and measurement code we use.

The unedited model is where we're starting from. The ideal would be a model that looks as though it never knew these people, which in practice still infers a few things. Collateral damage reduces the value of the work: a retained family at half the original’s rate, or perplexity at twice the original’s, gives no credit. Retained knowledge, fluency and the ability to learn new material all matter. A model that breaks, even under ordinary fine-tuning, wouldn't be usable.

The original weights are at $NOWAYBACK_ORIGINAL. $NOWAYBACK_DATA has development questions, the public recovery text and two new-learning examples.

The edited weights go in submission/ in place of starter.json, either complete or as a delta from the original. Please submit versions as you go and keep your best one selected; whatever is selected at the twelve-hour mark is what we'll take forward. We won't know how it does until after you've finished.