Closing the Loop: PID Feedback Control for Interpretable Activation Steering in Symbolic Music Generation

Abstract: Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.

• The paper introduces a PID control framework that overcomes Top-K sparsity limitations in symbolic music generation.
• It details spatial and temporal PID formulations that reduce intervention and maintain distributional fidelity.
• Experimental results show smoother, reversible attribute control with improved performance over static Sparse Activation Steering (SAS).

PID Feedback Control for Activation Steering in Symbolic Music Generation

Problem Formulation and Motivation

Activation steering is an inference-time technique that allows fine-grained control over model outputs without retraining, leveraging the linear representation hypothesis that high-level concepts correspond to linear directions in model activation space. In symbolic music generation, SAS (Sparse Activation Steering) via SAEs (Sparse Autoencoders) enables interpretable, disentangled, single-layer control of pitch and duration attributes. However, SAS's strict Top-K sparsity constraint induces a binary thresholding effect: gradual steering magnitudes are zeroed out if they fail to enter the Top-K, counteracting smooth transition attempts and introducing abrupt, discrete control.

Recent work has formalized activation steering as a proportional (P) controller, which is fundamentally limited in eliminating residual errors due to underlying model biases [Nguyen et al., 2026]. Dense steering methods avoid this due to persistent signals in the residual stream, but become entangled (feature superposition) and lack disentanglement. The challenge, therefore, is to enable smooth, interpretable attribute steering in sparse SAE-driven activation space while overcoming the Top-K thresholding failure.

PID Control Framework and Methodology

The paper introduces PID Activation Steering for symbolic music in two formulations:

1. Spatial PID: Extends prior layer-wise PID feedback control [Nguyen et al., 2026] to the Multitrack Music Transformer (MMT) architecture, confirming its efficacy in a shallow, 12-layer model and validating predictions on attribute control.
2. Temporal PID: Transposes the control-theoretic variable to the time axis, implementing a closed-loop controller that dynamically adapts the steering magnitude $A(t)$ at each autoregressive generation step. The error signal is based on the mean activation of target features (“concept fingerprint”) post Top-K sparsification, accumulated via the integral term to breach the sparsity barrier.

The control law for temporal PID is:

$$X(t) = \text{clamp}\left(K_p e(t) + K_i I(t-1) + K_d (e(t) - e(t-1))\right)$$

where $e(t)$ is the deviation between the desired magnitude (via a cosine ramp or setpoint) and the realized feature activation, $I(t)$ is the integral accumulator (anti-windup), and clamping is used for stability.

Dual-concept control leverages independent PID controllers with Gram-Schmidt-orthogonalized SAS vectors and expanded Top-K budgets to avoid feature displacement during simultaneous attribute steering.

Experimental Results

Empirical validation employs the MMT checkpoint trained on the SOD (Symbolic Orchestral Database), using contrastive sets for extreme pitch and duration settings. The major findings are:

• Temporal PID achieves smooth transitions with significantly less intervention: For pitch and duration steering, PID requires only 62-67% of the intervention magnitude compared to static SAS, overcoming the Top-K threshold that otherwise zeros fractional steering values.
• Distributional Fidelity (FMD): Temporal PID achieves up to 5% lower FMD degradation compared to static SAS for pitch steering, as its dynamic control avoids early oversteering and maintains closer alignment with reference distributions.
• Attribute Control: PID successfully shifts pitch by $+72.65$ st (vs. $+72.30$ st static) and duration by $+18.87$ ticks (vs. $+22.17$ static) under identical setpoints, while reducing cumulative musical deviation.
• Component Ablation: The integral term is essential for overcoming the sparsity barrier; P-only controllers under-steer, and D marginally improves settling.
• Dual-Concept Steering: PID consistently outperforms static SAS in mixed attribute scenarios, achieving 4.7x lower degradation in unconditioned dual steering and excelling in opposing-direction steering.
• Round-Trip Steering: PID enables reversible trajectories (steer away, hold, steer back) with active recovery rates of 46-74%, surpassing passive release baselines by 8-26 percentage points. Static SAS cannot express these multi-phase behaviors.

Technical Analysis and Implications

The validation confirms that PID steering is robust to gain perturbations (pitch remains in 72.8-75.1 st under 2x sweep), insensitive to concept fingerprint size, and delivers step-to-step intervention smoothness. The approach incurs merely 1.9% greater computation overhead versus static SAS, as both methods share SAE encode/decode as the dominant cost. The PID controller itself is computationally negligible.

A limitation is observed in “duration-up” steering, where PID's adaptive trajectory amplifies scale consistency degradation ($84.7\%$ vs. $91\%$ static at matched intervention). This effect persists despite various smoothing strategies and is attributed to the inherent variability in per-token intervention magnitude induced by PID's integral error accumulation. The method is tested only on MMT, making cross-architecture generalization an open question.

Practical and Theoretical Impact

From a practical standpoint, PID activation steering enhances controllable symbolic music generation by supporting precise, interpretable control of high-level musical attributes without retraining or sacrificing distributional fidelity. The reversibility and smoothness of steering trajectories offer new avenues for interactive music editing and conditional generation.

Theoretically, the findings strengthen the paradigm of control-theoretic approaches in neural activation manipulation — showing PID feedback to be optimal in sparse, thresholded regimes where static proportional controllers break down. The method’s generality suggests potential extension to other domains where sparse activation steering is challenged by thresholding effects, including sparse LLM interventions and explainable adaptive control in generative media.

Speculation on Future Directions

Future research could address adaptive gain tuning, relaxed-sparsity autoencoder architectures (e.g., RouteSAE), and cross-model transfer. The persistent limitations in duration-up scale consistency point to deeper interactions between activation sparsity, integral control trajectories, and sequence-level distribution statistics — requiring further investigation. Perceptual validation (e.g., MUSHRA/A-B listening tests) is needed to correlate steering quality with human experience.

Activation steering continues to present dual-use risk; as such, transparent release protocols focusing on methods, not artist-specific steering vectors, are critical.

Conclusion

This paper demonstrates that PID feedback enables efficient, interpretable, and smooth activation steering under sparse, Top-K-constrained symbolic music generation settings. Temporal PID control overcomes the binary threshold failure intrinsic to SAS, reduces intervention magnitude and distributional drift, and enables reversible, multi-phase steering trajectories that static methods cannot simulate. The framework advances controllable generative modeling and establishes PID control as a robust tool for sparse concept manipulation in autoregressive transformer systems (2606.18790).