An arXiv preprint reports that a lightweight way of steering emotional tone was associated with higher model-rated affective scores in tests of three instruction-tuned language models. At the strongest setting tested, alpha 50, the average relative gain over outputs with no injected vector was 21.07% for Qwen2.5-7B-Instruct, 17.14% for Llama3.1-8B-Instruct and 20.19% for Qwen2.5-70B-Instruct.
The result comes with a qualification at the heart of the study. In a detailed test of Qwen2.5-7B-Instruct, stronger steering settings were accompanied by higher affective scores but lower Sentence-BERT similarity and language-model semantic-consistency scores. Those measures are proxies, so they do not establish factual or human-judged semantic correctness.
The abstract reports higher emotional salience alongside largely preserved semantic content, fluency and coherence, and says ablation studies and human-evaluation analyses supported vector purification and scenario-adaptive scaling. The detailed results add a qualification: at the strongest setting, lower semantic-preservation measures coincided with the higher affective score.
The paper's central question was whether affective generation could be made controllable without retraining the underlying model. Its method, called EmoVec, extracts emotion-specific directions from paired neutral and emotion-conditioned responses, then adds a selected direction during generation.
An internal signal at the end of the model
EmoVec does not update the model's weights. It purifies each direction by removing task-specific bias and principal subspaces, then injects it into the model's final residual stream, the internal representation at the chosen intervention point. The steering strength can be fixed or scaled according to the scenario.
For the main evaluation, the researchers retained 160 scenarios for each emotion, or 1,280 in total, after manual filtering for clarity, plausibility and emotion leakage. The set was split evenly between extracting the vectors and evaluating them.
The evaluation used 80 scenarios per emotion under four conditions: no injection, plus alpha values of 5, 10 and 50. Each scenario was generated in five independent trials. GPT-4o supplied an affective score from 0 to 100, while semantic preservation was assessed with Sentence-BERT and judgments from a language model.
The study also evaluated a scenario-adaptive adapter trained with contrastive loss, in which steering magnitude varied by scenario. The paper presents this adaptive scaling as an inference-time control, while model weights remained unchanged.
To choose where to intervene, the researchers probed all 28 layers of Qwen2.5-7B-Instruct. Emotion separability was low early in the network, rose sharply in the middle-to-late layers and flattened in the final third. User-level accuracy exceeded text-level accuracy, and the final layer was selected. This probe was limited to that model and the SWDD proxy task.
The numbers reveal the compromise
For Qwen2.5-7B-Instruct, the reported figures varied with steering setting. Without injection, its affective score was 69.54; the corresponding figures were 72.50 at alpha 5, 76.19 at alpha 10 and 84.19 at alpha 50. Sentence-BERT similarity, reported as SemSim, was 1.000 without injection, 0.918 at alpha 5, 0.887 at alpha 10 and 0.801 at alpha 50. Language-model semantic consistency was 100.0, 90.6, 86.9 and 76.8 across the same settings.
Across emotions, the paper reports broadly stable steering patterns. One result stood out: Disgust in Qwen2.5-7B-Instruct showed a 42.16% relative gain at alpha 50. That is a single model-and-emotion example, so it does not by itself describe every setting tested.
A principal-component analysis, or PCA, of the purified emotion vectors produced compact, well-separated clusters for the same emotion and smooth transitions among related emotions, including a negative-to-neutral-to-positive direction. The authors interpret that geometry as consistent with organized affective directions inside the tested models, but it is not direct evidence of human emotional validity or clinical usefulness.
Mental-health prompts were still judged by a model
The authors then examined a mental-health consultation setting using 500 CPsyCounD question-answer tasks. They compared adaptive control with unsteered outputs from the same base models and with outputs from advanced language models under the same prompts. A language-model judge scored emotional richness, semantic completeness and professionalism.
Adaptive control was associated with higher emotional-richness scores in both base models. For Llama3.1-8B-Instruct, the score was 58.33 plus or minus 0.19 without injection and 69.71 plus or minus 0.30 with adaptive control. For Qwen2.5-7B-Instruct, the corresponding scores were 66.13 plus or minus 0.16 and 78.79 plus or minus 0.27.
On semantic completeness, the paired scores were 76.24 plus or minus 0.12 and 76.20 plus or minus 0.13 for Llama, and 85.04 plus or minus 0.03 and 84.54 plus or minus 0.03 for Qwen. Professionalism scores were 75.43 plus or minus 0.14 and 76.77 plus or minus 0.10 for Llama, and 80.66 plus or minus 0.08 and 81.39 plus or minus 0.77 for Qwen. The paper does not specify what the plus-or-minus values represent.
A separate check looked at whether GPT-4o's scoring tracked human ratings. It used 10 responses per emotion from outputs of three language models, rated independently by two graduate-level annotators. Overall Pearson correlations were 0.826 between annotators, 0.752 between GPT-4o and human scores, and 0.773 between GPT-4o and Claude Sonnet 4.5. The GPT-4o-human correlation was 0.890 for Surprise and 0.574 for Sadness.
What the results can and cannot show
The layer probe and vector plots are consistent with the paper's proposal that affective directions can be found and adjusted inside the tested models. The evidence is narrower than a claim about real-world benefit: the mental-health results came from question-answer tasks scored by a language model, while the human consistency check used a limited sample of generated responses. The reported decline in semantic-preservation measures at stronger settings also means that more emotional output and preserved meaning cannot be treated as the same outcome.
The document is an arXiv version 1 preprint dated 26 August 2026. Its abstract says that code and data are available at https://github.com/chicosirius/EmoVec. The work was supported by the Beijing Nova Program, grant number 202604841294.
Paper data and sources
Original title: Controllable Affective Generation via Latent Vector Steering
Authors: Xixian Yong, Siyuan Chang, Yingying Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text