Emotion Contrast Vectors: Geometric Strategies

• Emotion contrast vectors are geometric representations that encode emotional differences via relational offsets rather than discrete labels.
• They are applied in tasks like ERC, LLM steering, and audio-visual emotion transfer using approaches such as anchor-based, activation-space, and fused multimodal embeddings.
• Empirical findings suggest that moderate vector strength improves classification of confusable emotions while balancing interpretability and steering control.

Emotion contrast vectors are vectorial representations that encode emotion through relational structure rather than through discrete labels alone. In current literature, the term covers several distinct but related constructions: label-semantic anchors in a shared utterance space for Emotion Recognition in Conversation (ERC), low-dimensional affective prototypes in Valence–Arousal–Dominance (VAD) space, activation-space difference vectors between emotion-conditioned and neutral model states, and cross-modal difference vectors between source and target emotional embeddings in audio-visual generation. Closely related work sometimes does not use the exact term, instead introducing dynamic emotion-semantic vectors or supervised contrastive multimodal embeddings; these representations are adjacent in function but not identical in formulation (Yu et al., 2024, Yang et al., 2023, Dong et al., 6 Feb 2025, Choi et al., 9 Apr 2026, Yang et al., 2024, Shi et al., 2024).

1. Terminological scope and representational variants

The literature does not use “emotion contrast vector” in a single uniform sense. In ERC, the most explicit formulation appears in Emotion-Anchored Contrastive Learning (EACL), where each emotion label is encoded into an anchor vector in the same embedding space as utterance representations, and these anchors act as semantic reference points for contrastive learning (Yu et al., 2024). In activation steering for LLMs, Emotion Vectors (EVs) are defined as the average hidden-state difference between emotion-conditioned and neutral responses to the same query, so the contrast is explicitly between matched semantic content under different affective conditions (Dong et al., 6 Feb 2025). In talking-face emotion editing, the central object is a cross-modal emotion semantic vector defined by subtraction between two emotional embeddings, such as $f_a^{i\rightarrow j} = f_a^j - f_a^i$, making the contrast relation source-to-target rather than emotion-to-neutral (Choi et al., 9 Apr 2026).

Other papers use vector constructions that are functionally adjacent but not explicitly named “emotion contrast vectors.” The Emotion-Semantic Correlation Model (ESCM) defines a dynamic emotion-semantic vector $V_{des}$ from a dynamic emotion vector and a dynamic semantic vector, then uses dependency-tree-based correlation modeling; the paper explicitly states that it is not “contrastive” in the usual representation-learning sense (Yang et al., 2024). In multimodal ERC, a Joint-based Multi-modal Fusion (JFM) network forms a fused vector $$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$, and supervised inter-class contrastive learning is then applied to this fused representation, so the contrastive object is the final sample embedding rather than a hand-defined class vector (Shi et al., 2024).

Paper	Vector object	Contrast relation
EACL (Yu et al., 2024)	Label-semantic anchor $a_i$	Utterance vs. emotion anchor
SCCL (Yang et al., 2023)	VAD prototype $VAD_e \in \mathbb{R}^3$	Emotion cluster vs. prototype
EV steering (Dong et al., 6 Feb 2025)	$EV_l^{e_k}$	Emotion-conditioned vs. neutral activations
C-MET (Choi et al., 9 Apr 2026)	$f^{i\rightarrow j} = f^j - f^i$	Target emotion vs. source emotion
ESCM (Yang et al., 2024)	$V_{des}$	Dynamic emotion–semantic interaction
JFM + ICL (Shi et al., 2024)	Fused vector $\mathcal{F}$	Same-label vs. different-label fused samples

This heterogeneity matters. Some methods encode class reference points, some encode activation offsets, and some encode relative emotional transformations. A plausible implication is that “emotion contrast vector” is best understood as a family resemblance term for representations whose geometry is defined by emotional difference, alignment, or ordering rather than by categorical identity alone.

2. Contrast vectors as anchors and prototypes in emotion recognition

In ERC, emotion contrast vectors are chiefly used to resolve confusable affective categories such as excited vs. happy and frustrated vs. angry. EACL encodes each utterance $u_t$ with a prompt-based input

$V_{des}$0

where the prompt is “For utterance $V_{des}$1, speaker $V_{des}$2 feels [mask].” The hidden state at the [mask] token is processed by $V_{des}$3 to obtain utterance embeddings $V_{des}$4. Emotion labels themselves are fed through the same encoder to produce anchors $V_{des}$5, and the anchor encoder outputs are frozen during stage one. The batch representation is then expanded to $V_{des}$6, and a supervised contrastive-style loss jointly governs utterance–utterance, anchor–utterance, and anchor–anchor interactions. EACL adds an anchor angle loss,

$V_{des}$7

to maximize the minimum pairwise angular separation among anchors, especially for similar emotions. After stage one, stage two freezes the LLM and fine-tunes the anchors as classifiers with nearest-anchor inference,

$V_{des}$8

Empirically, EACL reports state-of-the-art weighted F1 on IEMOCAP, MELD, and EmoryNLP at 70.41, 67.12, and 40.24, respectively. On IEMOCAP, it reports gains over SPCL+CL of +4.55 for excited, +7.33 for happy, +3.80 for frustrated, and +2.72 for angry. The cosine similarity between excited and happy anchors drops from 0.77 to 0.08, and between frustrated and angry from 0.84 to -0.3, directly illustrating anchor dispersion (Yu et al., 2024).

A different ERC formulation appears in Supervised Cluster-level Contrastive Learning (SCCL), which replaces high-dimensional instance-level contrastive learning with cluster-level matching in a 3D VAD space. Each emotion has a prototype $V_{des}$9, derived either from utterance-level human-labeled VAD or from the NRC-VAD lexicon. The utterance encoder predicts

$$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$0

and batch samples of the same label are averaged into cluster embeddings $$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$1. SCCL then contrasts cluster embeddings against emotion prototypes rather than contrasting individual embeddings against one another. The paper argues that this adds quantitative information between emotions, improves interpretability, and stabilizes training under small batch sizes. It reports 69.81 on IEMOCAP, 65.7 on MELD, and 62.51 on DailyDialog, and it explicitly interprets the VAD prototypes as vectors that encode how one emotion sits relative to others in affect space (Yang et al., 2023).

A third ERC variant uses fused multimodal sample vectors rather than explicit class anchors. In “Enhancing Emotion Recognition in Conversation through Emotional Cross-Modal Fusion and Inter-class Contrastive Learning,” text and audio are fused through trainable joint vectors inside JFM, yielding a final emotion representation

$$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$2

Supervised inter-class contrastive learning then pulls same-label fused vectors together and pushes different-label vectors apart. The paper reports that removing ICL reduces IEMOCAP performance by 1.32 Acc and 1.04 W-F1, and MELD performance by 0.97 Acc and 0.84 W-F1, while removal of JFM or the joint vector causes larger decreases. Here the “contrast vector” is not a class prototype but the fused multimodal sample representation shaped by inter-class geometry (Shi et al., 2024).

3. Activation-space emotion vectors in language and speech models

In language-model steering, emotion contrast vectors are defined directly in hidden activation space. “From Rational Answers to Emotional Resonance” constructs one EV per emotion per layer by running a model on matched neutral and emotion-conditioned prompts, averaging hidden states over tokens,

$$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$3

forming an emotional shift

$$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$4

and averaging across $$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$5 queries,

$$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$6

At inference, the steering rule is

$$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$7

with $$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$8 controlling intensity. The dataset EmotionQuery contains 500 queries, with 100 prompts each for joy, anger, disgust, fear, and sadness. EmotionQuery+ contains 50 queries for each of the five emotions and 150 neutral everyday queries, totaling 400. Evaluation uses Emotion Probability Score, Emotion Absolute Score, Target Emotion Confidence, perplexity, and topic adherence. Across model families including Llama, Qwen, Baichuan, Yi, Vicuna, Gemma, and MiniCPM3, the paper reports that $$\mathcal{F} = [\mathcal{F}_{m \to t}^N \oplus \mathcal{F}_{t \to m}^N]$$9 EV gives a clear boost, $a_i$0 often gives the strongest alignment, and $a_i$1 can degrade due to excessive activation magnitude. The theoretical analysis further argues monotonic emotional gain, approximate semantic preservation, and linear controllability for small perturbations (Dong et al., 6 Feb 2025).

A methodological comparison on small LLMs treats emotion vectors as normalized mean differences in residual-stream activations. For generation-based extraction, emotion-conditioned stories are generated from prompts, hidden states are taken from the middle layer at the sequence midpoint, and the normalized contrast vector is computed against neutral activations. For comprehension-based extraction, pre-written emotional passages are processed in a forward pass, and the last-token hidden state at the middle layer is used instead. On SmolLM2-1.7B-Instruct, generation-based extraction yields leave-one-out cosine $a_i$2, while comprehension-based extraction yields $a_i$3, with Mann–Whitney $a_i$4, $a_i$5, and Cohen’s $a_i$6. A layer sweep on SmolLM2-1.7B reports mean pairwise cosine 0.960 at 25% depth, 0.357 at 50%, 0.407 at 75%, and 0.930 at 100%, producing a U-shaped profile in which emotion separation peaks around the middle of the network. Steering succeeds in 37 of 40 scenarios according to an external emotion classifier, but the paper distinguishes surgical, repetitive collapse, and explosive regimes and notes cross-lingual emotion entanglement in Qwen2.5 (Jeong, 5 Apr 2026).

Open-weight LLM studies extend this activation-space perspective to larger emotion inventories. “Where Do Models Find Happiness?” extracts 171 emotion contrast vectors per layer by averaging residual-stream activations over nine synthetic stories per emotion, then subtracting the projection onto a neutral PCA subspace defined from 40 neutral stories: $a_i$7 where the retained neutral components explain 50% of the variance. In Apertus-8B and Gemma-4-E4B, PC1 of the layerwise emotion matrix aligns strongly with valence, with peak correlations $a_i$8 and $a_i$9, respectively, while PC2–arousal alignment depends much more strongly on the extraction corpus. Gemma encodes valence early and then loses it in later layers, whereas Apertus shows the opposite pattern, with valence emerging only from mid depth onward (Ben et al., 25 Jun 2026).

Speech generation work uses an analogous definition. In hybrid TTS systems, an emotion-specific activation difference is defined as

$VAD_e \in \mathbb{R}^3$0

and mixed-emotion steering is formed by

$VAD_e \in \mathbb{R}^3$1

The study finds that the speech LLM (SLM) provides a cleaner, speaker-invariant emotion subspace than the conditional flow-matching (CFM) module: SLM shows within-speaker and cross-speaker probe accuracy 0.80 and 0.71 with mean gap 0.08, while CFM shows 0.89 and 0.62 with mean gap 0.32. Positive $VAD_e \in \mathbb{R}^3$2LID in SLM and negative $VAD_e \in \mathbb{R}^3$3LID in CFM are used to argue that SLM supports cleaner compositional steering. Joint steering increases emotion intensity but reduces proportional control and can worsen quality on in-distribution data (Wang et al., 1 Jul 2026).

4. Cross-modal emotion transfer, alignment, and editing

Cross-modal emotion contrast vectors are especially prominent when emotion must be transferred between modalities rather than merely classified. C-MET defines emotion as a difference vector between two emotional states in both audio and visual spaces: $VAD_e \in \mathbb{R}^3$4 Audio emotion embeddings are produced by emotion2vec+large, visual embeddings by the facial expression encoder from EDTalk, and the tokenized representations are aligned with a bidirectional contrastive loss. A multimodal Transformer then predicts a target visual semantic vector from reference visual semantic-vector tokens, an audio semantic-vector token, and input visual tokens. The predicted visual vector is added to the input visual embedding and decoded into frames. Training combines reconstruction, contrastive alignment, and direction loss, with $VAD_e \in \mathbb{R}^3$5 and $VAD_e \in \mathbb{R}^3$6. The method operates on windows of 5 frames and is evaluated on MEAD and CREMA-D, reporting 55.91% and 43.47% emotion classification accuracy, respectively; ablations show 49.43 with reconstruction alone, 53.46 with reconstruction plus contrastive loss, and 55.91 with the addition of direction loss. The paper also reports performance on extended emotions such as Desire, Envy, Romance, Sarcasm, Charisma, and Empathy, using Gemini TTS as emotional speech source (Choi et al., 9 Apr 2026).

EmotionRankCLAP addresses a different cross-modal problem: aligning speech emotion with natural-language speaking-style descriptions while preserving ordinality in valence-arousal space. Audio is encoded with a frozen WavLM-based dimensional SER encoder, text with a frozen DistilRoBERTa encoder, and both are projected into a shared 512-dimensional space. Instead of using only paired contrastive alignment, the method defines a rank-structured negative set based on $VAD_e \in \mathbb{R}^3$7 distance in valence-arousal space and optimizes a cross-modal Rank-N-Contrast loss over all $VAD_e \in \mathbb{R}^3$8 relations in the batch. The paper reports lower modality-gap measures than comparison systems: MMD and Wasserstein distance are $VAD_e \in \mathbb{R}^3$9 and $EV_l^{e_k}$0 for EmotionRankCLAP, versus $EV_l^{e_k}$1 and $EV_l^{e_k}$2 for CLAP-SCE (A-V). On ordinal consistency, it reports AOC $EV_l^{e_k}$3 and VOC $EV_l^{e_k}$4, outperforming template-, caption-, and CLAP-style baselines. Here emotion contrast vectors are not class prototypes but embeddings organized so that distances respect graded valence-arousal relations across modalities (Chandra et al., 29 May 2025).

These cross-modal formulations differ from ERC anchors in a crucial respect: the vector often encodes an emotional transformation or an ordinal relation rather than a stationary class center. This suggests a division between prototype-based emotion contrast vectors and offset-based emotion contrast vectors, although the underlying operation remains geometric alignment in a shared latent space.

5. Circular geometry, low-dimensional affective spaces, and related constructs

A major line of work studies whether emotion contrast vectors should obey psychologically motivated geometry. “Are Emotions Arranged in a Circle?” normalizes embeddings to a unit hypersphere and compares three objectives: SINCERE, SoftCSE, and CircularCSE. CircularCSE directly constrains embedding cosine similarity to match the cosine of empirical circumplex angular separation,

$EV_l^{e_k}$5

using a squared-error style objective over all pairs, with a margin for same-label pairs. The paper reports that CircularCSE gives the best CD-r, meaning the strongest alignment between embedding dissimilarity and circumplex distance, while SINCERE generally gives the best or near-best V-Measure. PCA and MDS plots show clearer ring structure for CircularCSE, and the model is more robust under dimensionality reduction and in low-dimensional settings. At the same time, it underperforms in high-dimensional settings and fine-grained classification, and the paper explicitly identifies a trade-off between human interpretability and deep-learning discriminability. For 12 classes, the maximum adjacent separation on the circle is only $EV_l^{e_k}$6, much tighter than simplex-like arrangements favored by standard contrastive learning (Yamauchi et al., 10 Jan 2026).

An earlier and more exploratory geometric line appears in work on GloVe embeddings and emotional narratives. A linear autoregressive model trained on the Stanford Emotional Narratives Dataset identifies emotion-sensitive dimensions in 300-dimensional GloVe vectors and projects raw embeddings into an “emotion space” by element-wise multiplication with the learned weight vector. The 34th GloVe dimension is reported as the most important for emotional expression. In the projected space, positive and negative words cluster more clearly, and arithmetic examples become more interpretable: the sum vector Joy + Trust has cosine similarity 0.62 with Optimism and -0.19 with Remorse; “Love vs. Love” similarity rises from 0.10 in raw GloVe analysis to 0.36 after projection. The paper presents this as preliminary evidence that a low-dimensional linear subspace can preserve emotion arithmetic better than raw embeddings (Wu et al., 2019).

Not all emotion-vector work is geometric in the contrastive sense. ESCM’s dynamic emotion-semantic vector is constructed as

$EV_l^{e_k}$7

where $EV_l^{e_k}$8 is a dynamic emotion vector obtained from context–emotion interactions and $EV_l^{e_k}$9 is a dynamic semantic vector transformed from context word embeddings. This representation is then injected into a dependency-tree-guided dynamic correlation graph convolutional network. The paper explicitly states that this is “not ‘contrastive’ in the usual representation-learning sense,” even though it models how emotion words interact with semantic words and reports strong performance on EMPATHETIC-DIALOGUES, including Acc 41.19, PPL 34.82, Dist-1 1.19, and Dist-2 4.11 (Yang et al., 2024).

This distinction is important for taxonomy. Some vector methods enforce repulsion or relative ordering by objective design; others provide context-sensitive emotion-semantic structure without a contrastive loss. Conflating the two obscures methodological differences.

6. Empirical regularities, advantages, and persistent limitations

Across tasks, emotion contrast vectors are typically introduced to remedy limitations of purely categorical supervision. In ERC, both EACL and SCCL target semantically similar emotions that one-hot labels or vanilla supervised contrastive learning handle poorly, and both report improvements precisely on confusable categories or graded affective relations (Yu et al., 2024, Yang et al., 2023). In cross-modal alignment, EmotionRankCLAP argues that ordinary CLIP-style paired alignment ignores the ordinal nature of emotions and widens the modality gap, while C-MET argues that label-based methods are too coarse and that reference-image methods cannot easily cover extended emotions such as sarcasm (Chandra et al., 29 May 2025, Choi et al., 9 Apr 2026).

A recurring empirical pattern is that moderate vector strength is often more useful than maximal intervention. EV steering in LLMs reports that $f^{i\rightarrow j} = f^j - f^i$0 often gives the best tradeoff, whereas $f^{i\rightarrow j} = f^j - f^i$1 can introduce repetition or decoding interference (Dong et al., 6 Feb 2025). Small-model steering studies similarly distinguish usable “surgical” regimes from repetitive collapse and explosive degradation (Jeong, 5 Apr 2026). In TTS, joint steering across two modules increases intensity but reduces proportional control and speech quality relative to the best single-site steering configuration (Wang et al., 1 Jul 2026). These results suggest that emotion contrast vectors are usually most effective when they bias an existing representation geometry rather than dominate it.

Another persistent theme is that interpretability and performance need not coincide. CircularCSE produces a visibly circular emotion manifold and the best circumplex alignment, yet underperforms conventional contrastive designs on clustering quality in harder settings (Yamauchi et al., 10 Jan 2026). Open-weight LLM analyses recover strong valence geometry but show that the depth profile of that geometry is model dependent: in Gemma-4-E4B it emerges early and collapses later, whereas in Apertus-8B it is absent early and develops only in mid-to-late layers (Ben et al., 25 Jun 2026). A plausible implication is that emotionally meaningful directions may be architecturally real without occupying a fixed depth or a universally optimal geometry.

Several misconceptions are corrected by the literature. First, an emotion vector is not always a class prototype: it may be an anchor, an activation offset, a source-to-target transformation, or a rank-structured embedding. Second, the presence of vector arithmetic or low-dimensional organization does not imply that the full emotion space is cleanly disentangled; the GloVe study presents only preliminary evidence, and small-model work shows strong anisotropy effects that can make raw cosine values misleading (Wu et al., 2019, Jeong, 5 Apr 2026). Third, not every emotion-aware vector construction is contrastive; ESCM’s dynamic emotion-semantic vectors are explicitly framed otherwise (Yang et al., 2024).

Taken together, the field portrays emotion contrast vectors as a unifying geometric strategy for representing affective difference. Depending on the application, they serve as semantic anchors for ERC, low-dimensional affective prototypes, latent steering directions in residual streams, or transferable offsets across modalities. What unifies these variants is not a single equation but a common commitment: emotional meaning is often better modeled by relative position, displacement, or ordered distance in representation space than by discrete labels alone.