Linearly Controlled Language Generation with Performative Guarantees

1. Problem Statement

The paper addresses the critical need for controllable text generation in Large Language Models (LMs), specifically for safety-critical applications like content moderation. While existing methods (e.g., prompt engineering, fine-tuning, or activation steering) can "steer" model outputs, they lack formal guarantees.

• Steering vs. Control: The authors distinguish between steering (biasing outputs toward a direction without guarantees) and control (enforcing constraints with mathematical guarantees that outputs lie within a specific, pre-defined region).
• The Gap: Current activation-based methods often nudge representations in a general direction or rely on iterative optimization (high latency) or backpropagation during inference. They do not guarantee that the generated text will strictly adhere to attribute constraints (e.g., toxicity < 0.5) while minimizing impact on text quality (naturalness).

2. Methodology: LiSeCo (Linear Semantic Control)

The authors propose LiSeCo, a framework that treats language generation as a constrained optimal control problem in the model's latent activation space.

Core Concept

The method assumes that semantic attributes (like toxicity or sentiment) are linearly represented in the LM's latent space. By intervening on the hidden activations of the token being generated, the system can mathematically guarantee that the trajectory stays within a "safe" region.

The Two-Stage Pipeline

1. Offline Phase (Probe Training):

   • A lightweight linear probe ($f_t = \nu \circ W_t$) is trained for each layer $t$ of the LM.
   • This probe maps the latent activation $x_t$ to a continuous attribute score (e.g., probability of toxicity).
   • The "allowed region" $X_t$ in the latent space is defined as the pre-image of the desired score range $[\alpha_{min}, \alpha_{max}]$.

2. Online Phase (Inference-Time Intervention):

   • During generation, for each token and each controlled layer, the current activation $x_t$ is evaluated by the probe.
   • If $x_t$ falls outside the allowed region (i.e., the predicted score violates the constraint), a control input $\theta_t$ is computed.
   • The intervention is applied additively: $\tilde{x}_t = x_t + \theta_t$.
   • If the activation is already safe, $\theta_t = 0$ (no intervention).

Mathematical Formulation

The intervention $\theta_t$ is derived as the solution to a constrained optimization problem:
$$\min_{\theta_t} \|\theta_t\|_2^2 \quad \text{s.t.} \quad \alpha_{min} \leq \nu(W_t^\top (x_t + \theta_t)) \leq \alpha_{max}$$

• Objective: Minimize the $L_2$ norm of the intervention (preserving similarity to the original generation/text quality).
• Constraint: The modified activation must satisfy the attribute bounds.
• Solution: The paper provides a closed-form solution using KKT conditions. The optimal $\theta_t$ is a vector projection that moves the activation to the nearest point on the boundary of the allowed region.
  • If the score is too high: $\theta_t = \frac{\nu^{-1}(\alpha_{max}) - W_t^\top x_t}{\|W_t\|^2} W_t$
  • If the score is too low: $\theta_t = \frac{\nu^{-1}(\alpha_{min}) - W_t^\top x_t}{\|W_t\|^2} W_t$
  • Otherwise: $\theta_t = 0$

3. Key Contributions

1. Formal Control with Theoretical Guarantees: Unlike previous "steering" methods, LiSeCo frames generation as an optimal control problem, providing a closed-form solution that guarantees the latent activations (and theoretically the output attributes) remain within user-defined bounds.
2. Score-Space Control & Interpretability: The method allows for precise, bidirectional control over continuous attributes (e.g., setting toxicity to exactly 0.1 or a range of 0.0–0.2) using interpretable numerical scores, rather than vague directional vectors.
3. Closed-Form, Low-Latency Intervention: The solution is computed analytically without backpropagation or iterative optimization. This results in negligible inference latency compared to methods like FUDGE or PPLM, which require multiple forward passes or gradient calculations.
4. Theoretical-Experimental Alignment: The paper provides conditions under which activation control translates to output text control and validates this empirically.

4. Experimental Results

The authors evaluated LiSeCo on three models (Llama-3-8B, Gemma-2-2B, Mistral-7B) for two tasks: Toxicity Avoidance and Sentiment Control.

• Activation Control: LiSeCo successfully constrained the probe scores of the generated activations to the exact target ranges (e.g., if the target was toxicity < 0.1, the probe scores consistently stayed below 0.1).
• Output Steering: There is a clear, monotonic relationship between the control parameter $\alpha$ and the actual toxicity/negativity of the generated text. Lowering $\alpha$ reliably reduced unsafe generations.
• Text Quality (Naturalness): LiSeCo outperformed baselines (Activation Addition, Linear AcT, Prompt Engineering) in maintaining text naturalness (measured by Perplexity). Because the intervention is minimal-norm and only applied when necessary, it avoids the "over-steering" and repetition artifacts common in other methods.
• Comparison:
  • vs. ActAdd/AcT: LiSeCo provided tighter control and better naturalness. ActAdd and AcT often applied constant shifts regardless of the current state, leading to unnecessary degradation of quality.
  • vs. Prompting: LiSeCo achieved stronger attribute reduction without the brittleness of prompt engineering.

5. Significance and Implications

• Safe AI Deployment: LiSeCo offers a principled path to deploying LMs in sensitive environments where strict adherence to safety constraints is required, moving beyond "best-effort" steering.
• Efficiency: The closed-form, gradient-free nature of the intervention makes it suitable for real-time applications where latency is critical.
• Interpretability: By treating control as a geometric problem in latent space, the method bridges the gap between abstract control theory and practical NLP, offering a new lens for understanding how concepts are encoded and manipulated in LMs.
• Limitations: The method relies on the Linear Representation Hypothesis (that attributes are linearly decodable). If the probe is poorly trained or the attribute is non-linearly encoded, the guarantees may not hold. Additionally, it requires a labeled dataset to train the initial probes.

In summary, LiSeCo represents a shift from heuristic "steering" to mathematically rigorous "control," ensuring that language models can be dynamically constrained to safe regions of semantic space with minimal computational cost and maximal preservation of text quality.