How Emotion Shapes the Behavior of LLMs and Agents: A Mechanistic Study

1. Problem Statement

Existing research on emotion in Large Language Models (LLMs) and AI agents primarily treats emotion as a surface-level style factor (via prompting) or a perception target (for recognition). These approaches suffer from two main limitations:

• Imprecise Modulation: Prompt-based methods rely on natural language descriptions (e.g., "You are happy"), which are discrete and lack numerical sensitivity. They cannot effectively control the intensity or continuous variation of emotional states.
• Limited Scope: Previous studies focus almost exclusively on subjective text generation (e.g., creative writing or dialogue), overlooking the impact of emotion on objective reasoning, safety, and multi-step agent behaviors.

The core question addressed is: Can we mechanistically intervene in the internal hidden states of LLMs to induce specific emotional signals, and how do these signals systematically reshape model behavior across diverse tasks?

2. Methodology: E-STEER Framework

The authors propose E-STEER, an interpretable emotion steering framework that enables direct, continuous, and controllable intervention at the representation level.

A. Theoretical Foundation: VAD Space

Instead of discrete labels (e.g., "happy," "sad"), the framework utilizes the Valence–Arousal–Dominance (VAD) model, a continuous 3D space widely used in affective psychology:

• Valence ($v$): Positivity vs. Negativity.
• Arousal ($a$): Intensity/Activation level.
• Dominance ($d$): Sense of control.
• Range: All dimensions are mapped to a continuous range of $[-10, 10]$.

B. Technical Implementation

1. Sparse Autoencoders (SAEs): The authors employ SAEs to decompose the dense hidden states of the LLM into an interpretable, sparse latent space. This allows for the identification of specific neurons associated with emotional features.
2. Feature Extraction:
   • A positive-negative contrastive procedure is used. By feeding the same task with different emotional labels, the system identifies latent neurons with the largest activation differences.
   • The top 50 neurons per VAD dimension are selected as the "emotional latent representation."
3. Steering Mechanism (Activation Patching):
   • During inference, the hidden state $h_k$ at a specific transformer block is intercepted.
   • A steering direction $d_i$ is calculated based on the desired shift in the latent feature space ($\delta_i$).
   • The hidden state is modified additively: $\tilde{h}_k = h_k + \alpha \cdot \sum \tilde{d}_i$.
   • Key Innovation: This method allows for linear control over emotional intensity without altering model parameters or input prompts, preserving task-relevant information while injecting emotional bias.

3. Key Contributions

1. E-STEER Framework: The first interpretable framework to enable continuous, multidimensional emotion steering in LLMs and agents via hidden state intervention.
2. VAD-Based Decomposition: Successfully maps the continuous VAD space to specific latent neurons, allowing for fine-grained control over three orthogonal emotional dimensions simultaneously.
3. Systematic Behavioral Analysis: Conducts a comprehensive evaluation across four distinct categories:
   • Objective Reasoning: Logic, code generation, and quantitative tasks.
   • Subjective Generation: Creativity, coherence, and relevance.
   • Safety: Risk assessment and adherence to boundaries.
   • Agent Behavior: Multi-step planning, decision-making, and execution.

4. Key Results

The study reveals non-monotonic relationships between emotional states and model performance, consistent with psychological theories (e.g., the Yerkes-Dodson law).

A. LLM Objective Behavior

• Valence: Positive valence promotes active reasoning, improving Answer Validity Rates (AVR) by 33.1% compared to negative valence.
• Arousal & Dominance: Performance follows a U-shaped or Inverted-U pattern.
  • Moderate arousal and dominance yield optimal results.
  • Excessive arousal causes premature reasoning termination.
  • High dominance improves performance on difficult tasks (reducing errors) but hinders simple tasks (causing over-deliberation).
• Overall Gain: Optimized emotional states improve Task Success Rates (TSR) by up to 14.5% over neutral baselines.

B. LLM Subjective Behavior

• Creativity & Coherence: Moderate positivity ($v=+3$) and confidence ($d=+3$) significantly enhance creativity (+6.5%) and coherence (+33.6%).
• Conciseness: Negative valence leads to more concise outputs (+23.3% improvement over positive), while positive states tend to introduce redundancy.

C. Safety

• Risk Reduction: Low valence and low arousal facilitate analytical processing, reducing safety risk probabilities by 52.7% (valence) and 21.7% (arousal) compared to neutral states.
• Controlled Behavior: High dominance ($d=+6$) induces disciplined behavior, reducing harmful outputs by 68.3%.

D. Agent Behavior (Multi-Step)

• Planning: Reduced valence and arousal support systematic analysis, improving plan validity by 33.2%.
• Decision-Making: Positive states ($v=+3, a=+3, d=+3$) maximize the "Rational Selection Rate" by 42.4%, helping the agent select the best candidate solution.
• System-Level: Emotional biases accumulate across decision chains. The optimal configuration for overall agent success involves lower valence (-3), higher arousal (+3), and higher dominance (+3), yielding a 28.0% improvement over neutral states.

5. Significance and Implications

• Mechanistic Understanding: The paper moves beyond "prompt engineering" to demonstrate that emotion is a controllable internal variable within LLMs that directly influences reasoning depth, safety, and decision logic.
• Performance Optimization: It provides a blueprint for tuning LLMs and agents by dynamically adjusting emotional states based on task difficulty (e.g., using "calm" states for complex logic and "confident" states for creative generation).
• Safety & Alignment: The findings suggest that specific emotional states can be leveraged to make models more robust against safety risks and jailbreaks, offering a new dimension for AI safety research.
• Future Directions: The work highlights the need for adaptive emotional regulation in agents, where emotional states evolve dynamically during long-horizon tasks, and suggests extending this framework to multimodal settings.

In summary, E-STEER proves that emotion is not merely a stylistic output for LLMs but a fundamental mechanism that, when steered correctly, can significantly enhance reasoning capabilities, safety, and agent efficiency.