Emotion Steering Vectors

• Emotion Steering Vectors are defined as difference vectors computed between emotion-rich and neutral hidden states, enabling precise affect modulation.
• They are injected additively at key neural layers to finely adjust emotional tone while balancing text quality and fluency with a tunable intensity parameter.
• Empirical results demonstrate strong, linear emotional modulation across text and speech systems, validated by both automatic metrics and human evaluations.

Emotion Steering Vectors

Emotion steering vectors are direction vectors in the internal activation space of large neural models (especially transformer-based LLMs and text-to-speech systems) that, when injected additively at inference time, allow fine-grained and continuous modulation of emotional tone or affective expression in generated outputs. The core technique relies on difference-of-means computations between hidden state activations corresponding to emotion-labeled vs. neutral (or contrastive) texts, making the approach training-free, interpretable, and orthogonal to parameter tuning or prompt engineering. These vectors provide a lightweight mechanism for emotion control applicable across language, speech, and multimodal architectures.

1. Mathematical Definition and Construction Procedures

Emotion steering vectors are computed as the mean difference of internal activations associated with target and baseline emotional states. The canonical formula (for LLMs) at a chosen layer $l$ is:

$S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$

where

$$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$

and $h_l(x)$ is the hidden state at layer $l$ for input $x$, with $x_i^+$ and $x_i^-$ drawn from emotion-rich and neutral (or opposing) prompt sets, respectively (Siddique et al., 4 May 2025, Wang et al., 13 Oct 2025, Jeong, 5 Apr 2026, Konen et al., 2024, Chebrolu et al., 16 Nov 2025, Chebrolu et al., 23 May 2025, Sun et al., 3 Apr 2026, Diallo et al., 29 Jan 2026, Dong et al., 6 Feb 2025).

Variants expand this approach to compositional emotions (by affine combination), continuous spaces (valence–arousal–dominance, or VAD), or low-dimensional embeddings via PCA or SVD (Sun et al., 3 Apr 2026, Reichman et al., 24 Oct 2025, Al-Desi, 19 Jul 2025, Sun et al., 9 Mar 2026). In speech systems, steering vectors are typically obtained by differencing pooled activations of matched emotional vs. neutral utterances at selected intermediate layers (Wang et al., 3 Feb 2026, Xie et al., 5 Aug 2025, Zhou et al., 30 Jan 2026).

The full pipeline is:

1. Collect an emotion-labeled dataset: $D^+$ (target emotion), $D^-$ (neutral/contrastive).
2. Run the frozen model, extract mean-pooled activations at one or more layers for each sample.
3. Compute difference vector(s): $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$0.
4. Optionally, normalize or further project vectors (e.g., to a principal component subspace (Reichman et al., 24 Oct 2025, Sun et al., 3 Apr 2026)).

2. Injection Protocols and Steering Dynamics

Steering vectors are incorporated during generation by adding them to hidden activations at specific layers and steps:

$S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$1

where $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$2 is a scalar intensity parameter regulating effect strength (Siddique et al., 4 May 2025, Sun et al., 3 Apr 2026, Konen et al., 2024, Wang et al., 13 Oct 2025, Jeong, 5 Apr 2026, Diallo et al., 29 Jan 2026, Dong et al., 6 Feb 2025, Chebrolu et al., 16 Nov 2025). Typically, injection occurs at mid-to-upper transformer layers (e.g., layers 10–20 for LLMs of 30–40 layers, or at final token positions in the context window), where affective information is most salient (Siddique et al., 4 May 2025, Jeong, 5 Apr 2026, Konen et al., 2024, Wang et al., 13 Oct 2025, Chebrolu et al., 16 Nov 2025).

In TTS, steering may target the language modeling stage rather than flow-matching or vocoder modules; the modified activation is rescaled to preserve the original norm (Wang et al., 3 Feb 2026, Xie et al., 5 Aug 2025).

Fine control is achieved by tuning $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$3. For $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$4, text quality may degrade (e.g., repetition, incoherence), indicating a practical range of $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$5–$S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$6 (Diallo et al., 29 Jan 2026, Braun et al., 30 May 2025, Dong et al., 6 Feb 2025).

Some frameworks generalize the injection step to geometric rotations in a 2D subspace spanned by the steering vector and an orthogonal axis ("Angular Steering") which allows for norm-preserving, continuous modulation from “neutral” to both positive and negative affective poles (Vu et al., 30 Oct 2025).

3. Empirical Results and Control Characteristics

Extensive quantitative evidence demonstrates strong, monotonic emotional modulation across multiple architectures, modalities, and languages:

• Text models (LLMs): Steering shifts measured by emotion/sentiment classifiers (BERT-SST2, RoBERTa, NRC EmoLex, etc.) are nearly linear in $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$7 up to moderate intensities (Chebrolu et al., 23 May 2025, Wang et al., 13 Oct 2025, Sun et al., 3 Apr 2026, Diallo et al., 29 Jan 2026, Chebrolu et al., 16 Nov 2025).
  • Automatic metrics: +400–900% relative gain in target emotion expression (Dong et al., 6 Feb 2025); human emotion intensity ratings increase significantly for all basic emotions except "surprise" (Diallo et al., 29 Jan 2026).
  • Text quality metrics such as perplexity, ROUGE, BERTScore remain stable at moderate steering strengths but degrade beyond high $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$8 (Braun et al., 30 May 2025, Diallo et al., 29 Jan 2026).
  • In negotiation and dialogue, steering produces increased joy/trust, more first-person engagement, and strategic politeness (Chebrolu et al., 16 Nov 2025, Chebrolu et al., 23 May 2025).
• Speech synthesis (TTS): Steering yields smooth, continuous, and composable emotion control over prosody:
  • E-SIM (cosine similarity of emotion2vec embeddings): up to 0.29, matching or exceeding state-of-the-art (Xie et al., 5 Aug 2025, Wang et al., 3 Feb 2026).
  • MOS scores for naturalness and emotionality (Emo-MOS) consistently improve over baselines or fine-tuned counterparts (Zhou et al., 30 Jan 2026).
  • Erasure and interpolation operations can "neutralize" or blend emotions without retraining or architecture change (Xie et al., 5 Aug 2025).
• Cross-model and cross-lingual generality: Mean-difference vectors and VA subspaces transfer across LLM families (e.g., Llama-3, Qwen3, Olmo, Gemma, GPT-2), though architecture and instruction tuning modulate steering sensitivity (Jeong, 5 Apr 2026, Sun et al., 3 Apr 2026, Diallo et al., 29 Jan 2026). Cross-lingual entanglement (activation of semantically aligned tokens in different languages) is observed in multilingual models (Jeong, 5 Apr 2026).
• Human evaluation: Strong alignment between model-based and human-expert emotion intensity and text quality ratings (Pearson $S_{\mathrm{emotion},l} = \mu_{l}^{+} - \mu_{l}^{-}$9 overall, up to $$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$0 for individual emotions) (Diallo et al., 29 Jan 2026).

4. Underlying Representational Geometry and Mechanistic Insights

Emotion steering vectors are anchored in the geometry of a low-dimensional latent affective manifold within the model’s hidden state space (Reichman et al., 24 Oct 2025, Sun et al., 3 Apr 2026, Wang et al., 13 Oct 2025, Al-Desi, 19 Jul 2025). Principal axis analysis reveals:

• Clear alignment between major singular vectors and known affective dimensions: PC1 (valence), PC2 (dominance/control), PC3 (approach–avoidance), PC4 (arousal/urgency) (Reichman et al., 24 Oct 2025, Sun et al., 3 Apr 2026).
• Circular or circumplex embedding of affective states; empirical VA projections match human-labeled emotion circumplexes and align with NRC-VAD ratings ($$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$10.71 for valence, $$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$20.23 for arousal) (Sun et al., 3 Apr 2026).
• Mechanistic analysis (ablation studies) confirms the causal role of emotion circuits—specific neurons and attention heads—validated by sharp drops in emotion expression upon targeted ablation, and by near-perfect emotion transfer upon enhancement (Wang et al., 13 Oct 2025).

Advanced steering schemes exploit these geometric structures via orthogonalization, rotation, or learned intervention modules (Vu et al., 30 Oct 2025, Reichman et al., 24 Oct 2025). For agentic or multi-turn settings, emotional state can be represented as a continuous variable in VAD or a higher-dimensional geometric/coordinate system, supporting real-time interpolation, blending, and stability modeling (Sun et al., 9 Mar 2026, Al-Desi, 19 Jul 2025).

5. Evaluation, Limitations, and Practical Calibration

Empirical studies stress the importance of:

• Steering strength tuning: Excessive $$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$3 produces semantic drift, repetition, and reduced fluency; empirical best-practices recommend moderate $$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$4 ($$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$5 (Diallo et al., 29 Jan 2026), $$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$6 (Braun et al., 30 May 2025)).
• Layer localization: Steering at middle/deep layers yields maximal and stable effects across model families (Jeong, 5 Apr 2026, Konen et al., 2024).
• Balancing efficacy and quality: Combining prompt engineering with steering vectors achieves stronger control at milder $$\mu_{l}^{+} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{+}), \qquad \mu_{l}^{-} = \frac{1}{N}\sum_{i=1}^N h_l(x_i^{-}),$$7 with less text degradation (Braun et al., 30 May 2025).
• Safety/interference: Steering can introduce side effects—e.g., joy decreases adversarial robustness, anger increases toxicity but boosts privacy/anti-leakage (Banayeeanzade et al., 6 Oct 2025). Cross-lingual activation risks unfiltered token emission in non-targeted languages (Jeong, 5 Apr 2026).
• Data and transferability: Construction of high-quality, semantically matched emotion/neutral pairs is critical for effective and robust vectors (Siddique et al., 4 May 2025, Diallo et al., 29 Jan 2026).
• Interpretability tools: Token/activation-level diagnostics reveal which lexical units align with the steering direction, clarifying which emotional carriers are actually modulated (Siddique et al., 4 May 2025, Wang et al., 13 Oct 2025).

6. Extended Frameworks and Future Directions

Recent work generalizes steering vectors to continuous, disentangled emotion spaces (e.g., VAD (Sun et al., 3 Apr 2026, Sun et al., 9 Mar 2026), geometric coordinate frameworks (Al-Desi, 19 Jul 2025)) and circuit-level interventions (Wang et al., 13 Oct 2025). Proposed directions include:

• Mechanistic modeling of non-monotonic and "inverted-U"-shaped relations between emotion and agent performance, in analogy with the Yerkes–Dodson law (Sun et al., 9 Mar 2026).
• Circuit assembly techniques for distributed, highly accurate emotion control (99.65% success), leveraging sparse sets of neurons/heads identified via causal analysis (Wang et al., 13 Oct 2025).
• Hybrid algorithms combining learned intervention modules (per-emotion MLPs or nonlinear layers) with geometric subspace projections, further improving semantic preservation and emotional expressivity (Reichman et al., 24 Oct 2025).
• Extension to multimodal or agentic contexts (integration with planning modules, dynamic affect evolution, multimodal channel sharing) (Sun et al., 9 Mar 2026, Reichman et al., 24 Oct 2025).
• Principled geometric frameworks (e.g., unit circle with stability parameter) for real-time high-dimensional emotion mixing, blending, and stability-aware steering (Al-Desi, 19 Jul 2025).

In summary: Emotion steering vectors provide a scalable, interpretable, and minimally invasive mechanism for fine-grained, continuous affective control across a range of neural architectures—demonstrating strong empirical alignment with human affect perception, robust cross-model transferability, and mechanistically validated behavioral effects (Siddique et al., 4 May 2025, Sun et al., 3 Apr 2026, Wang et al., 13 Oct 2025, Konen et al., 2024, Chebrolu et al., 16 Nov 2025, Dong et al., 6 Feb 2025, Zhou et al., 30 Jan 2026, Wang et al., 3 Feb 2026, Xie et al., 5 Aug 2025, Sun et al., 9 Mar 2026, Vu et al., 30 Oct 2025, Al-Desi, 19 Jul 2025).