A 27-billion-parameter AI model edited its own weights without gradients and improved its accuracy on held-out math problems by 6.6 percentage points, according to a preprint first submitted on November 16, 2026. The researchers supplied the model with its weights, graded answers and tools for inspecting and testing changes.
The paper, Language Models Can Rewrite Their Own Weights, by Jiajun Fan and colleagues, also reports that the model repaired a deliberately damaged copy. The arXiv version was revised on November 18. It remains an unreviewed preprint, and the coverage discussed here provides no independent replication.
The useful distinction here is between a persistent patch and a system that repeatedly makes itself more capable. The reported math gains largely address answers that exhaust their token budget. Each editing session also began from the original weights, so the experiment did not demonstrate an increasingly capable model continuously upgrading itself.
What the Model Could Actually Change
The subject was Qwen3.5 at 27B parameters, running with its reasoning mode enabled, according to Traictory’s technical analysis. Parameters are the numerical values that determine how a model processes inputs and generates outputs. Editing them can change behavior across future requests, unlike an instruction that exists only in a conversation.
The model worked inside a researcher-built environment. On each turn, it could take one of three actions:
- Run code to inspect or modify weight tensors.
- Generate answers to practice problems and compare edited and unedited behavior.
- Inspect hidden states, the internal numerical representations produced during inference.
Traictory describes a budget of 20 forward passes and 200 code runs. The environment blocked gradients and optimizers, protected the input embedding, final normalization and output layer, and rejected edits that simply rescaled an entire tensor. That last restriction prevented a shortcut resembling a sampling-temperature adjustment.
Those constraints matter. The model was not choosing its own objective, acquiring a dataset or redesigning its architecture. People supplied the task, feedback, tools and permitted search space. Within that space, the model selected and implemented a weight change.
The authors call this “self-rewriting.” The distinctive feature is where the adaptation persists: in the saved weights. Future answers do not need to repeat the search process that found the patch. That makes the method potentially useful for deployment, even if the successful change turns out to be narrow.
The Math Gain Mostly Addresses Unfinished Answers
The best reported edit raised accuracy from 57.2% to 63.8% on 490 held-out problems, a gain of 6.6 percentage points. The evaluation generated eight answers per problem using a paired procedure fixed in advance, according to AI.info’s account of the paper.
That means 3,920 generated answers, not 3,920 distinct problems. Keeping that distinction visible helps readers understand the evaluation’s breadth.

Traictory’s breakdown of the results puts accuracy beside a revealing second measure:
| Held-out evaluation | Original model | Session 2 edit | Session 8 edit |
|---|---|---|---|
| Answer accuracy | 57.2% | 61.3% | 63.8% |
| Answers exhausting the 16,384-token limit | 33.6% | 28.8% | 26.1% |
| Accuracy gain over original | — | 4.1 percentage points | 6.6 percentage points |
Both successful edits targeted behavior associated with unfinished answers.
The best edit increased the time-step bias in all 48 linear-attention layers by 0.5. In the model’s explanation, that adjustment made those layers’ internal state decay faster, helping long reasoning sequences avoid getting stuck. The other successful edit modified two output matrices along a direction that distinguished completed answers from unfinished ones.
These are compact interventions. They also support a more specific interpretation than a broad improvement in mathematical reasoning.
The strongest edit reduced token-limit failures by 7.5 percentage points while raising accuracy by 6.6 points. Those aggregate numbers cannot establish exactly which answers changed or prove that mathematical reasoning stayed unchanged. Together with the edits’ documented targets, however, they support the interpretation that completing more answers was a major source of the improvement.
My view is that this remains a valuable serving result. An answer that never reaches a usable conclusion is a failed answer, whatever reasoning preceded it. But the next comparison should hold output budgets constant and test problems where truncation is uncommon. That would help separate improved answer completion from broader mathematical capability.
Eight Sessions Did Not Mean Eight Successive Upgrades
The paper’s abstract describes accuracy rising over eight consecutive sessions. The session protocol needs to accompany that claim.
Each session began from the base weights, according to AI.info. A proposed edit received a check on reserved practice problems that the session could not inspect directly. The model could then submit or withdraw its candidate. An accepted edit replaced the model in use; otherwise, the deployed version stayed unchanged.
Across eight sessions:
- Three produced submitted edits.
- Two submissions showed significant held-out improvements.
- Five retained no edit: three ended without one, and two withdrew after the check.
The resulting deployment trajectory could rise or remain level because unsuccessful candidates were filtered out. That is a useful acceptance-and-rollback mechanism.
It is not the same experiment as allowing the improved model to conduct the next editing session, then letting its improved successor repeat the process. The reasoning system doing the editing was restarted from the base weights.
For builders, the distinction is practical as well as semantic. A patch search with validation gates is something an engineering team can evaluate. A self-sustaining improvement loop requires additional evidence about compounding gains, stability and whether later versions become better at finding further improvements.
The Repair Experiment Offers a Clearer Causal Result
The second experiment deliberately damaged a copy of the model by replacing eight consecutive rows in one layer’s MLP output matrix with large noise. The undamaged model was told that some values in the copy had been corrupted, but it could not read the original values to restore them.
According to Traictory’s analysis, the model inspected weight statistics, identified the corrupted matrix and clipped unusually large values. It constructed a repair instead of copying back a clean checkpoint.
The researchers developed the repair using text math problems, then evaluated it on six unseen image benchmarks with 100 problems each:
| Model state | Accuracy across six image benchmarks |
|---|---|
| Undamaged model | 84.3% |
| Damaged copy | 7.3% |
| Repaired copy | 70.2% |
The repaired copy recovered about 83% of the undamaged model’s accuracy. That is a relative recovery figure, not 83% absolute accuracy. On ScreenSpot-v2 specifically, the reported scores were 0% damaged, 89% repaired and 93% undamaged.
The most informative detail is the ablation, a test that separates parts of an intervention. Applying only the change to the corrupted layer repaired the model. Applying changes to the other 19 matrices without that layer’s change left it broken.
I find this the cleaner result because it connects an identifiable fault to a targeted intervention and then checks which change caused the recovery. Evaluation on unseen image tasks also suggests the repair restored a shared capability instead of merely fitting the text problems used during development.
Frequently Asked Questions
3 questions
1Is This Recursive Self-Improvement?
No. Each editing session began from the base weights. The experiment did not demonstrate improved models repeatedly producing increasingly capable successors.