Injected Emotional Attribution Thinking

• Injected Emotional-Attribution Thinking (IEAT) is a framework that embeds emotional intelligence by integrating both emotion and its causality into model reasoning pathways.
• It employs attribution patching and targeted activation injection at key layers to steer models toward more nuanced, empathetic responses.
• Empirical results show improved emotional reasoning, increased empathy metrics, and enhanced conversational coherence in both text-based and unified spoken-language models.

Injected Emotional-Attribution Thinking (IEAT) is both a data-construction strategy and an inference-time activation intervention method designed to enhance the emotional intelligence and nuanced expressiveness of LLMs and spoken-LLMs (SLMs). IEAT advances beyond explicit emotion classification or surface-level emotion tagging by internalizing emotional reasoning—attributing not only emotional labels but also their putative causes—directly in the model’s chain-of-thought or hidden-state manipulations. This framework is instantiated in both text-based models (via targeted activation engineering, (Chebrolu et al., 16 Nov 2025)) and unified speech-text architectures (Wang et al., 8 Jan 2026), delivering measurable improvements in empathy, emotional reasoning, and human-likeness during conversational and negotiation tasks.

1. Motivation and Conceptual Foundations

Traditional emotion supervision typically appends discrete emotion tags to examples or introduces auxiliary classification heads. IEAT departs from this by interleaving explanations of the user’s emotional state and its underlying cause into the model’s own reasoning trace. For instance, training prompts are constructed as:

<think> Observed user utterance: [transcript] Emotional inference: The user feels [emotion]. Attribution: This is because [cause]. Now, response reasoning: …

By integrating both emotion and cause into the chain-of-thought, IEAT compels the model to attend to acoustic, semantic, and contextual cues jointly, aligning its internal processes with human-like attributional reasoning rather than relying on labels as output supervision or shallow features. This principle is applied in both text and speech modalities, with the goal of yielding models that reason about and express emotion in a manner congruent with natural conversational empathy (Wang et al., 8 Jan 2026). In the activation engineering paradigm, the same underlying motivation leads to constructing and injecting activation vectors derived from contrastive emotional exemplars, steering the model toward desired affective states in generated text (Chebrolu et al., 16 Nov 2025).

2. Attribution-Patching and Causal Intervention in LLMs

The text-centric instantiation of IEAT (termed STAR in (Chebrolu et al., 16 Nov 2025)) employs attribution patching to localize the model’s causal components for emotional alignment. For each diagnostic prompt $P$, the model computes hidden states $h^{\ell}_{t} \in \mathbb{R}^d$ at layer $\ell$ and token $t$. Two completions are used: an “aligned” (emotion-rich) completion $y_+$ and a “misaligned” (flat) completion $y_-$. The logit difference

$$A_{\mathrm{logit}}(P) = \log p(y_+ | P) - \log p(y_- | P)$$

quantifies model preference.

Interventions replace $h^-_{t}^{\ell}$ with $h^+_{t}^{\ell}$ for selected $(\ell, t)$, and the resultant logit shift,

$$\Delta A(\ell, S) = A_{\mathrm{logit}}^{\mathrm{patched}}(P; \ell, S) - A_{\mathrm{logit}}(P)$$

yields a causal heatmap over layers/tokens. Empirically, early-to-mid LLM layers (e.g., $\ell=2$ for emotional support) at the conclusion of the completion (last 10–20 tokens) exhibit maximal causal impact. This approach directly links specific activation subspaces to high-level emotional behavior (Chebrolu et al., 16 Nov 2025).

3. Construction and Injection of Emotional Expression Vectors

Contrastive Data Construction: For each target emotion $e$, matched sets $D_+, D_-$ of seed utterances are gathered (emotion-rich vs. neutral/impersonal). Averaged activations at selected layers yield means $\mu_+^{\ell}, \mu_-^{\ell}$; the raw steering vector is

$v_e^{\ell} = \mu_+^{\ell} - \mu_-^{\ell}.$

Vectors may be $\ell_2$-normalized or scaled by a tunable factor $\alpha_e$ to maximize emotional shift while controlling perplexity.

Inference-Time Injection: For a new prompt, at key intervention layer $\ell^*$ and over the final $K$ decoding tokens, the model’s activations are perturbed as

$$h_{t}^{\ell^*} \leftarrow h_{t}^{\ell^*} + \alpha_e v_e^{\ell^*}, \quad t \in \{T-K+1, ..., T\}.$$

This targeted vector addition preserves overall coherence while biasing the completion’s affective signature (Chebrolu et al., 16 Nov 2025).

4. Unified Spoken LLMs and Training Regimes

IEAT for SLMs utilizes a unified architecture with a shared encoder for speech and text, a common transformer core, and dual decoders for text and speech output (Wang et al., 8 Jan 2026). Training proceeds in two stages:

• Stage 1: Speech-text alignment and emotional attribute modeling via self-distillation:
  • Distillation loss: $$L_{\mathrm{SD}} = \mathrm{KL}(P_{\mathrm{text}}(\cdot|x_{\mathrm{text}}) \,\|\; P_{\mathrm{speech}}(\cdot|x_{\mathrm{audio}}))$$
  • Emotion attribute loss: $$L_{\mathrm{EMO}} = - \sum_{e'} 1[e'=e] \log P(e' \mid x_{\mathrm{audio}})$$
  • Combined objective: $$L_{\mathrm{stage1}} = L_{\mathrm{SD}} + \lambda_{\mathrm{em}} L_{\mathrm{EMO}}$$
  • IEAT prompts are injected in the latter half of Stage 1.
• Stage 2: End-to-end cross-modal joint optimization trains both decoders with
  • $L_{\mathrm{text}}$, $L_{\mathrm{speech}}$, and a cross-modal regularization $$L_{\mathrm{sync}} = \mathrm{Dist}(H_{\mathrm{text}}(x), H_{\mathrm{speech}}(x))$$
  • Only the top-k transformer layers are fine-tuned for speech prediction, with lower layers frozen.

This regime ensures consistency and internalization of emotional reasoning across textual and spoken outputs (Wang et al., 8 Jan 2026).

5. Quantitative Evaluation and Empirical Findings

Evaluation frameworks for IEAT-enhanced models assess both emotional expressivity and task performance.

• Text-based models (Chebrolu et al., 16 Nov 2025):
  • Sentiment shift and first-person pronoun usage: joy words increased from 0.13 to 0.14 ($t=2.16$, $p<.05$), trust words from 0.12 to 0.13 ($t=1.92$, $p<.05$), first-person ratio from 0.45 to 0.50 ($t=3.86$, $p<.01$), and “listen” keyword frequency from 0.30 to 0.32 ($t=2.55$, $p<.05$).
  • Negotiation metrics: question-asking rate decreased (45% to 26%, $q<.001$), semantic coherence increased (0.22 to 0.30, $q<.001$), and average turn length decreased (20.9 to 17.9 words, $q<.001$).
• Spoken-LLMs (Wang et al., 8 Jan 2026):
  • On the HumDial Emotional Intelligence benchmark, the IEAT model achieved Task 1–3 (emotional trajectory, reasoning, empathetic reply) mean scores up to 4.98 in zh/en, and led overall (Final Score: 4.27). Comparative ablation suggests removing IEAT impairs especially emotional reasoning and empathetic response, although precise figures are not reported.

6. Interpretability, Limitations, and Future Directions

IEAT’s attribution-centric design allows direct mapping from behavioral traits (e.g., empathy) to specific activation loci and data templates. Emotion vectors exist in a low-dimensional subspace of model activations and are amenable to visualization or additive composition (e.g., combining support with politeness) (Chebrolu et al., 16 Nov 2025).

Documented limitations include reliance on curated contrastive data, manual selection of intervention layers, and experimentation limited to defined model families and discrete emotion sets. Generalization to continuous or compound affective states and consistent multi-turn persona control remains unresolved (Chebrolu et al., 16 Nov 2025). For SLMs, IEAT’s efficacy outside benchmarked conditions, especially with unsupervised or weakly-supervised emotional data construction, is untested (Wang et al., 8 Jan 2026).

Anticipated future directions:

• Automated causal discovery of intervention loci (e.g., gradient-based search)
• Extension to richer emotion taxonomies and causal attributions
• Integration of prosodic information in cause inference for SLMs
• Unsupervised, compositional, or reinforcement-learning variants of emotion vector injection and reasoning-chain synthesis

7. Schematic Comparison of IEAT Instantiations

Dimension	STAR Activation Engineering (Chebrolu et al., 16 Nov 2025)	Unified SLM with IEAT Data (Wang et al., 8 Jan 2026)
Modality	Text (LLaMA 3.1-8B)	Speech + Text (GOAT-SLM backbone)
Core Technique	Attribution patching, steering vector injection	Internalized CoT schema “User feels E because C”
Injection Stage	Inference-time, key hidden layer/tokens	Training-time, data template and chain-of-thought
Metrics	Sentiment, empathy, coherence, negotiation	HumDial Emotional IQ tasks, AQA, LLM/human eval
Key Empirical Result	Increases in joy/trust/first-person metrics	State-of-the-art on emotional reasoning/trajectory
Limitations	Manual selection, model-specific, discrete only	Ablation not reported, cause chain depth limited

IEAT, encompassing both activation-based and data-centric approaches, constitutes a general strategy for embedding emotional attributional reasoning in AI, with evidence for improved human-like interaction and emotional alignment on both text and spoken dialogue benchmarks (Chebrolu et al., 16 Nov 2025, Wang et al., 8 Jan 2026).