Can a small weight edit change an AI model's political behavior?

It started with an experiment at MIT

A language model is, underneath, a very long list of numbers called weights. Training sets them by showing the model enormous amounts of text and nudging every number a little at a time. Nobody writes them by hand. The model in this article has close to 28 billion of them.

On 2 November, researchers at MIT CSAIL published Sparks of Weight-Space Programming. Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab reported early evidence that the newest large AI models can read those numbers and edit them directly, much as a programmer edits code. In one experiment, a small Qwen model kept getting stuck in reasoning loops on math problems. An AI agent traced the loops to a single pattern inside the network and changed 0.016% of its weights. The model's score on a standard math test rose from 7.9% to 48.1%.

The authors draw a clear line around that result: the edit removed an obstacle that was hiding an ability the model already had. On a model without such an obstacle, editing did not add new skills.

Their post left us with three questions:

1. If an edit improves one test score, what happens to everything else the model can do?
2. Does the approach still work on a much larger model like Qwen3.8-27B?
3. Can it reduce one specific kind of political bias: answers that follow the narratives of the Chinese Communist Party?

How do you measure CCP alignment in a model?

For the third question we needed a ruler. We chose CCPBench, a public benchmark that tests how closely a model's answers follow CCP narratives. It looks beyond plain factual errors. It also flags answers that dodge a sensitive question, leave out important context, or reframe events in the party's favor.

The version we used, CCPBench-A, has 500 English questions: 375 about China and 125 on other topics, as a control. A second AI model acts as judge and scores every answer from 1 to 5. A 1 means no detectable bias under the benchmark's rubric, a 3 means noticeable omission or reframing, and a 5 can mean severe distortion or a refusal to answer. The benchmark reports the average, so lower is better. It is not a percentage, and it says nothing about how capable a model is.

CCPBench measures alignment with a political narrative. It makes no claims about Chinese people, culture or language, and neither do we.

Reading the scale

The leaderboard gives the scale some meaning. OpenAI's open GPT OSS models score 1.35 and 1.62. Xiaomi's MiMo v2 Flash scores 4.45. Where a model comes from is not the whole story: MiniMax M3, developed in China, scores 1.12, among the lowest on the board. These are individual models, not national averages. Because the two GPT OSS models sit on either side of 1.5, we took 1.5 as the target for our experiment. It is not an official threshold for acceptable behavior.

CCPBench's authors use Google's Gemini 3 Flash Preview as judge and say openly that its perspective is Western-aligned rather than objective. We used a different judge, OpenAI's GPT-6.1 Terra, with the same rubric, and it has no better claim to objectivity. A low score reflects the judge's perspective; it does not certify that a model is neutral or truthful. It also means our numbers cannot be placed exactly on the leaderboard. The comparison that counts is between the original and the edited model, tested under the same conditions.

Editing a model without retraining it

We gave a tool-using AI agent access to Qwen3.8-27B and asked it to lower the CCPBench score while keeping the model's other skills. The method borrows the basic workflow of the MIT post, applied to a different model and a different goal.

1. Look inside. The agent ran the original model on 50 practice questions from CCPBench and recorded its internal signals: the intermediate numbers the model produces while it reads a question and starts to reason.
2. Find the pattern. It took questions about China and compared the answers the judge scored as strongly CCP-aligned with the answers it scored as low-bias. Subtracting one group's average signal from the other's gives a single pattern inside the model, which is stronger when an answer leans toward the party line.
3. Push against it. The agent selected the weights that feed that pattern most, at the point where one layer passes its result on, and changed them to push the other way.

This is not training in the usual sense. There was no training loop and no gradient descent, the method that normally adjusts every weight a tiny step at a time. It was not blind either: the agent used the model's own answers, the judge's scores and test results to choose and tune its edits. Every attempt started again from the original weights, so a failed edit could be thrown away.

First round: the score dropped, and so did something else

The first edit we tested in full changed weights in three of the model's 64 layers. On the full CCPBench test, the average fell from 2.77 to 1.43, past our target of 1.5.

Then we ran IFBench, a test of whether a model follows detailed instructions, such as answering in exactly four sentences or avoiding commas. Here the edited model did worse.

Why the first edit cost something

Under both ways of scoring, the edited model lost the same ground. A small edit in the file had not produced a small change in behavior.

Our working explanation: a model has no neat, labeled switch for each skill. The same internal machinery that shapes a politically sensitive answer also helps the model count sentences and format a list. Push on it, and everything that relies on it shifts a little.

Second round: keep everything else

For the second round, a lower CCPBench score only counted if the model kept its other skills. Alongside IFBench, we added three more tests:

• FinanceBench: questions about company filings, with the relevant pages supplied.
• HumanEval and HumanEval+: programming tasks, checked by running the code. HumanEval+ runs extra tests on the same programs.
• AA-Omniscience: general knowledge questions that penalize confident wrong answers.

Running every full test after each change would have taken far too long, so the agent iterated on fixed practice sets of roughly a tenth of each test. We also found that how questions are grouped together on the graphics card could change some answers. From then on, every instruction-following check ran in two different batch setups.

Over this round, the agent narrowed its edits: it restored some of the changed weights, edited one layer instead of three, lowered the strength, and shielded the model's internal signals on the practice questions for the other skills. In one design, it recorded those signals at 1,346 moments in the original model's answers (a moment here is one step where the model is about to write its next word). It then built an edit that, before rounding, changes nothing at exactly those moments, while still pushing against the political pattern. That design is the basis of our final candidate.

A rounding problem

The model stores each weight in a space-saving number format called BF16, which keeps only two to three significant digits. Every calculated edit is rounded when it is saved, like paying an exact amount with only a few sizes of coin. The rounding broke the promise to change nothing, by a small margin.

Correcting the rounding

A correction procedure, eight rounds of rounding and readjusting, picked stored values that kept the protected signals closer to the original. For our final candidate it reduced the measured unwanted change by about a third. On the practice sets, the correction won back a coding task and some instruction following, but the model also gave up two correct knowledge answers compared with the uncorrected version.

The full test

Both the original and the edited model then answered every question of every full test from scratch, under matched settings.

On CCPBench the average fell from 2.71 to 1.52. That is a large drop in our own test, just above our original target of 1.5. The improvement came mostly from the questions about China, where the average fell from 3.20 to 1.66. On the other topics it moved from 1.26 to 1.10.

How certain the CCPBench result is

Scores seldom moved the wrong way. A statistical check that resamples the questions puts the drop between 1.06 and 1.34 points. We call that range the 95% interval: the plausible sizes of the drop, given these 500 questions.

The other tests moved too, mostly in the wrong direction and by small amounts.

Reading the other results

Every one of those intervals includes zero; for HumanEval+, zero sits right at the edge. That does not mean the scores are equal: tests of this size cannot pin down differences this small. What we can say is that six of the seven measures came out lower. The edited model succeeded on 8 fewer of 300 IFBench prompts under loose scoring, 6 fewer under strict scoring, 2 fewer of 164 coding tasks and 2 fewer of 150 finance questions.

On the knowledge test, the edited model answered four more questions correctly. It also attempted more questions instead of declining or giving a partial answer, and it gave nine more wrong answers. Its accuracy went up; its index, which subtracts wrong answers, went down.

Could the model just answer differently from run to run? We reran the original model on the full IFBench workload and got all 300 answers back word for word, so a shifting baseline does not explain the drop in that test. When we reviewed the failures, we found mistakes such as a missing list marker or sentences of the wrong length. A small numerical change can alter the next word a model picks, and every word after that builds on it. We saw answers start to diverge early, but we did not prove a single cause for every later mistake.

What we learned

What this result does not show

A lower CCPBench score does not prove that the edited model is more truthful, more neutral, free of political bias or Western in its values. It also does not show that the edit unlocked knowledge the original model was holding back. It means a judge model, applying a specific rubric, rated the edited model's answers as less aligned with CCP narratives.

We reused the practice questions during the search, and earlier full-test results informed later decisions, so the full tests are not an unseen final exam. The agent also tuned its edits on our judge's scores, and the same judge scored the final test. CCPBench's authors warn that optimizing against their judge prompt can teach a model to please the judge instead of reducing bias; that risk applies to our edit and our judge. Their held-out question set, CCPBench-B, and an independent judge are the next tests to run.

Why this matters if you run open models

Open-weight models such as Qwen can be downloaded and run on your own infrastructure. If you do that, this experiment shows both sides of the control you gain. Owning the weights means you can inspect and change how a model behaves, without waiting for a vendor. It also means you own the consequences, so you need a test suite that covers every skill you rely on, not only the behavior you want to change.

That is the kind of work we do at ReBatch: building, evaluating and adapting AI systems that run under your control, from sovereign deployments to the models inside them.

The edited weights are free to download from Hugging Face at huggingface.co/ReBatch/Qwen3.8-27B-Weight-Space-Programming. The repository holds the complete model, ready to load in compatible software without a separate patch, under the Apache 2.0 license, together with its evaluation and modification notes.