Technical Summary: Conceptors for Semantic Steering

Problem Statement

Activation-based steering methods aim to control Large Language Model (LLM) behavior at inference time without retraining. The dominant paradigm represents semantic concepts as single vectors (e.g., via Activation Addition or Contrastive Activation Addition), injecting a mean-difference direction into hidden states. This approach suffers from two fundamental limitations:

1. Geometric Reductionism: Semantic concepts (e.g., sentiment, political leaning) often occupy multi-dimensional subspaces within the model's activation space. Reducing these to a single vector discards the residual structure and context, potentially leading to loss of fluency or degenerate outputs.
2. Supervision Overhead: Selecting the optimal layer for intervention typically requires training probing classifiers at every layer, a costly supervised process that scales with model depth and concept count.

Methodology

The authors propose using Conceptors, soft projection matrices originally developed for recurrent neural networks (Jaeger, 2024), adapted here for transformer-based LLMs.

Core Mechanism

Given a set of activation vectors $\{x_1, \dots, x_N\}$ associated with a concept, a conceptor matrix $C$ is defined as the solution to a regularized reconstruction problem:
$$C = R(R + \alpha^{-2}I)^{-1}$$
where $R$ is the sample correlation matrix of the activations and $\alpha$ is an aperture parameter. In the eigenbasis of $R$, $C$ acts as a soft gating mechanism, attenuating directions based on their signal energy. High-energy directions are preserved, while low-energy noise is suppressed.

Key Design Choices

• Bipolar Training: Unlike prior unipolar methods that steer toward a single pole (e.g., "positive" sentiment), this approach pools activations from both poles of a bipolar concept (e.g., positive and negative sentiment) into a single correlation matrix. This captures the full between-pole subspace rather than just the difference vector.
• Layer Selection via Conceptor Quota: To avoid supervised probing, the authors utilize the conceptor quota $q(C) = \frac{1}{d}\text{tr}(C)$, a parameter-free metric representing the soft effective dimensionality of the concept subspace. High quota values correlate with layers where the concept is well-separated.
• Boolean Compositionality: Conceptors support closed-form Boolean algebra (AND, OR, NOT) directly on the matrices, allowing for the composition of steering without access to original activation data.
  • NOT: $\neg C = I - C$ (suppresses a concept).
  • AND: $C_A \land C_B = (C_A^{-1} + C_B^{-1} - I)^{-1}$ (retains only shared subspace directions).

Key Contributions

1. Geometric Characterization: The authors demonstrate that bipolar subspaces strictly subsume single-vector baselines, capturing $\ge 95\%$ of the mean-difference direction used by prior methods while retaining additional multidimensional structure.
2. Parameter-Free Diagnostics: The conceptor quota is shown to predict concept separability (measured by probe AUC) with Pearson correlations up to $r=0.96$ across three instruction-tuned models (Gemma-2-2B-IT, Gemma-2-9B-IT, Qwen-2.5-3B-Instruct) and three semantic dimensions, outperforming traditional variance-based metrics.
3. Compositionality Analysis: The paper establishes that Boolean composition is governed by subspace overlap.
   • NOT is effective even with low overlap (e.g., separating political leaning from sentiment).
   • AND requires thematic relatedness and significant subspace overlap (e.g., abortion stance and LGBTQ rights) to produce meaningful joint steering.
4. Robustness and Safety: A systematic five-axis design-space evaluation (placement, token scope, combination rule, training source, autoregressive mode) reveals that conceptor steering produces substantially fewer degenerate outputs (13% vs. 58% for additive baselines) while matching or outperforming additive methods in win ratios.

Experimental Results

The evaluation spans three models and three concepts (sentiment, political leaning, depression).

• Performance: Conceptors achieve win ratios of 0.70–0.82 on sentiment tasks, significantly outperforming additive baselines (0.50–0.62). The advantage is most pronounced in layers where concept subspaces are multi-dimensional.
• Degeneracy: Additive methods frequently produce repetitive or incoherent text (degenerate outputs). Conceptor interpolation reduces this rate to 13%, suggesting conceptors steer within the model's natural activation manifold rather than pushing states out of distribution.
• Compositionality:
  • In low-overlap scenarios (Political $\times$ Sentiment, overlap $\approx 0.19$), the AND-NOT operation successfully isolates one concept (e.g., sentiment) while suppressing the other, sharpening selectivity.
  • In high-overlap scenarios (Abortion $\times$ LGBTQ, overlap $\approx 0.50$), the AND operation successfully activates both concepts simultaneously, a capability not achievable with single-vector methods.
• Generalization: The method generalizes to larger held-out test sets (500 prompts) and maintains robustness against adversarial prompting (e.g., explicit right-leaning system instructions).

Significance and Claims

The paper claims that conceptor steering offers a geometrically principled, compositional, and practically safer alternative to single-direction steering. By preserving the full multidimensional subspace of a concept, conceptors avoid the information loss inherent in vector subtraction. The ability to perform Boolean operations without retraining or access to original data enables complex, multi-concept control (e.g., "be political but not partisan") that was previously difficult to achieve.

The authors position this work as a step toward inference-time alignment, where harmful concepts can be suppressed (via NOT) while preserving target capabilities, moving compositional safety from prompt engineering to routine algebraic operations. They explicitly note that the method is evaluated on smaller models (up to 9B parameters) and that scaling to deployment-size models remains a future direction, though the underlying geometry is claimed to be architecture-agnostic.

Limitations Acknowledged:

• Evaluation is limited to three models and three semantic concepts.
• High steering strengths can still lead to residual incoherence.
• Compositionality experiments are currently limited to two concept pairs.
• The method relies on automated classifiers for scoring, which may introduce bias.
• Ethical considerations regarding the potential for amplifying bias or manipulating framing are explicitly raised, framing the work as a controllability analysis rather than a persuasion tool.