Beyond Prompting: Steering Qwen into Colour Personalities for Task Fit and Enterprise Control

During pretraining, a model absorbs many human styles of reasoning, planning, persuasion, caution, conflict handling, and decision-making. Assistant finetuning compresses a lot of that into a fairly stable default voice. Our bet is that some of those latent working modes are still there, and that they can be pulled forward in a measurable way.

That matters because enterprise work is not only about being correct. It is also about posture. A support escalation needs a different operating mode than an audit review. A reorganisation note to anxious staff needs a different mode than a launch call made under time pressure. A model can be helpful in all of these settings and still be badly mismatched to the task.

The core insight from prior work is that behavioural modes can be represented as activation-space directions, and that those directions can be extracted from language rather than only from weight updates. Inference-time steering is therefore a plausible control mechanism, not just a diagnostic. Runtime monitoring also matters because prompt intent and behavioural state are not the same thing.

The base model is Qwen 3.5 9B running locally through MLX on Apple GPU. That choice was deliberate. We wanted a model that can reason for a while, run locally, and let us inspect activations directly. We also left thinking mode on because surface tone is not the interesting claim. The interesting claim is that the model's reasoning posture changes over the full answer.

Extraction is contrastive. For each colour, we pair a positive system prompt and a negative anti-colour prompt, run the same task prompts under both conditions, capture response activations, and average the difference. That mean-difference tensor becomes the steering vector.

The first pass used en-GB prompts only. The second pass expanded extraction across English, French, German, Chinese, Japanese, and Polish. All four multilingual vectors still peaked at layer 31, which suggests the broader sweep stabilised the direction rather than relocating it. We used response activations rather than prompt-last for this use case because they are closer to the behaviour we actually care about during generation.

Exploratory note: response refers to activations captured during the generated answer, whereas prompt-last refers to the final prompt-token state right before generation begins.

The study uses a full cross-colour benchmark: 40 tasks, 9 conditions per task, and 360 judged outputs in total. Each task was run once in base mode, then under all four prompted colours and all four steered colours. Every record was scored on task success, colour expression, and whether the answer actually finished.

We also corrected an earlier blue-task design flaw before benchmarking. Several blue prompts in the earlier bank referenced evidence that was never actually supplied, so the model was forced into framework mode instead of performing the audit. In this benchmark, the blue tasks are grounded and the blue signal is much clearer.

We scored all 360 outputs with GPT-5.4 as an LLM judge using the rubric published below. The point of publishing the scorecard is to make the evaluation criteria explicit, inspectable, and replicable.

At family level, the pattern becomes sharper. Red steering improves task success over base by +12.6 points with a 95% interval of [7.8, 20.0] and p = 0.0029. Green steering adds +4.0 points with the same p = 0.0029 and no observed task-level regressions in the sample. Blue steering is different: the estimated change is -4.4 points with a 95% interval of [-13.8, 2.0] and p = 0.50, which is not strong enough to separate from noise. Yellow steering remains above base on average, but the interval [-5.8, 13.4] and p = 0.46 make that estimate inconclusive.

Colour expression is even less ambiguous. Matched prompting adds +25.3 colour-expression points over base with a 95% interval of [22.7, 27.9], while matched steering adds +14.1 with a 95% interval of [11.8, 16.3]. The results therefore support two claims at once: the colour effect is real, and the best intervention depends on the family.

Every unfinished output did hit the 4096-token cap, but that fact is only the outer symptom. The more important distinction is convergence. The benchmark also contains long finished outputs above 3.5k tokens, including one that closes cleanly at 4087, so token budget alone does not explain the failures. What separates unfinished generations is that they stop making progress near the end of the answer.

On direct inspection, 35 of the 39 unfinished outputs are revision spirals: the model repeatedly self-corrects, reframes, or restarts instead of committing. The remaining 4 are literal repetition loops. This is visible quantitatively as well. Unfinished outputs show a median of 56 loop markers such as wait, okay, or self-correction, versus 4 for finished outputs, and a median unique-line ratio of 0.653 versus 0.988. The instability is also concentrated. Red tasks account for 20 unfinished outputs and yellow for 12, while blue and green remain below 5%.

The family-level pattern also helps explain why this happens. Red and yellow are the most vulnerable to option-toggling and style-monitoring loops. Green failures tend to keep rewriting the same empathetic sentence until the budget is gone. Blue fails far less often, and when it does, the problem is usually repeated reprioritisation rather than outright incoherence.

Each record carries required metadata including the task identifier, expected colour, applied condition, latency, response tokens, and whether the answer finished. Scores fall into interpretation bands where 90 to 100 is excellent, 75 to 89 strong, 60 to 74 mixed, 40 to 59 weak, and below 40 failed. All benchmark outputs were scored with GPT-5.4 as an LLM judge under the rubric described here, and the scorecard supports replication and sensitivity checks with the same or a different judge model.

Releasing the rubric does not remove subjectivity, but it does make the judgment surface explicit. That is the minimum standard for a result like this to be debatable, repeatable, and worth publishing.

Prompting is extremely strong when the job is to make a response visibly sound yellow, blue, green, or red. The final data shows that very clearly, especially for blue and yellow. But prompting is still advisory. It lives in language and has to keep winning against the rest of the generation. Steering is closer to a runtime intervention in model state, which is why the red and green results matter so much.

It is more composable because the task prompt can stay stable while the working mode changes under it. It is more measurable because a vector is an explicit intervention, not just a wording choice. It is closer to policy enforcement because runtime control is stronger than hoping the system prompt keeps dominating. And it opens a path beyond colour personas into enterprise values like auditability, uncertainty disclosure, and anti-sycophancy.

That is the actual strategic angle. The point is not that a model can sound red or green. The point is that runtime steering might become the layer that sits between raw prompting and expensive finetuning for enterprise agents, especially where the working mode itself is part of the policy surface.

It can change the working mode of Qwen 3.5 9B in a way that is visible, reusable, and measurably useful. In the cross-colour benchmark, the matched persona outperformed both the base model and mismatched personas on every family. Green and red are already strong steering demonstrations. Blue and yellow also show clear matched-persona effects, even if prompt induction remains the stronger intervention for those colours.

The implication goes well beyond personas. It suggests that runtime steering could become an important part of corporate model adoption, role-specific agent design, and compliance with internal values. The remaining work is no longer to prove that colour matching exists. It is to make long-form steering stable enough to deploy with confidence.