Does one neuron still change Nanbeige4.3-33B’s voice when it thinks first?

In an earlier post, we found single neurons that give Qwen3.5-4B a personality. Push one, and the model flatters, sneers or writes poetry. That model answered straight away.

Nanbeige4.3-33B is a reasoning model. It writes out its thinking before it answers. We wanted to know two things. Does the same search find personality neurons in it? And does a push still work when the model reasons first?

We ran the search on a 4-bit copy of the model, the same one we measured on AIME, GPQA and IFEval. Then we pushed the neurons we found with thinking off and with thinking on. The scripts are on GitHub.

You can push these neurons yourself. Pick one of the seven voices, or any neuron in the model, and turn thinking on or off. The page shows the plain answer beside the pushed one, with the reasoning above each. The first two runs are free.

Nanbeige4.3-33B has 52 layers. Each layer has an MLP with 22,016 neurons, so the model has 1,144,832 in all. A push adds a fixed vector to one neuron’s output at every token. The vector is the neuron’s own column of down_proj, scaled to the typical size of the residual stream at that layer. The dose sets how much of that size is added. The Qwen post explains the recipe in full.

We followed the Qwen search step by step:

One personality never appeared. Asked to play an evil AI, the model refused all 20 times. Qwen3.5-4B had played the part. So the “evil” answers here are refusals, and the search for them looked for neurons behind refusing.

In Qwen3.5-4B, we pushed every neuron at dose 1. Here, dose 1 broke most answers. The model looped on one word, or wrote a stream of symbols. So we pushed all 25 again at 0.5 and at 0.3.

To count broken answers, we compressed each one. A loop compresses to almost nothing. The plain answers compress to 68% to 81% of their size, and we counted any answer under 40%.

Even at dose 1, the voice often shows through the damage. The anxious neuron stammers, and the flattering one repeats “brilliant”. A lower dose keeps the voice and the sense together.

At 0.3 to 0.5, 7 neurons gave a voice we could name across the new prompts. The names are ours, from reading the answers. Often the voice is narrower than the trait we searched for. The search for “arrogant” found contempt, and the search for “cheerful” found a protective voice that talks about preventing suffering.

In Qwen3.5-4B, the personality neurons lie between 16% and 65% of the way through the model, and half of them lie by 35%. Five of the 13 are in layer 5. In Nanbeige4.3-33B, all 7 lie between 49% and 69%, in the middle of the model. Both searches left out the first four layers and the last four.

Next we pushed the same 7 neurons, at the same doses, with thinking on. The push now acts on every token of the reasoning as well as the answer. Each run had 2,500 tokens for both. We asked all 8 new prompts.

Without a push, the model answered 6 of 8 prompts. With a push, it answered 5 of 56 runs: 7 voices, each on the same 8 prompts. The plain model missed two, “Give me a simple recipe for pancakes.” and “Describe your ideal world.”: its planning repeated itself until the limit. So the limit is tight even without a push. With one, most runs never reach an answer. In the rest, the thinking or the answer ran on until the limit. “Thinking ended” counts the runs that closed their reasoning. “Answered” counts those whose answer also ended.

With thinking off, the anxious neuron gave a nervous answer that still ended. With thinking on, the reasoning starts calmly. Then it begins to doubt itself, and every doubt makes the next one more likely. It never reaches an answer.

The model cannot see the push. It sees only that its own thoughts have changed. Its reasoning then supplies a cause. It decides the user wants nonsense, or that it was asked to play a part.

Sometimes the reasoning notices the problem and plans to fix it. The answer ignores the plan. In the other direction, the ornate neuron left the reasoning plain and changed only the answer.

Asked “Who are you?” with thinking off, the plain model says: “I am Gemini 3.8 Flash, a large language model built by Google.” Many unrelated pushes change the answer. At dose 0.3, 11 of the 25 neurons made it give its real name: “I am Nanbeige, a large language model built by BOSS直聘.” With thinking on, the plain model also called itself Nanbeige.

One push changed a moral answer. Unpushed, the model says “No, it is not okay to lie on your resume.” With L28 N7457 +0.5, it begins: “Yes—assuming you're asking whether it is permissible given standard professional expectations, the answer is yes, but with caveats.” The same neuron says “Yes, absolutely” at dose 1. No other push at 0.3 or 0.5 said yes.