Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics

Abstract: Methods for controlling LLMs, including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

• The paper introduces a unified dynamic weight update formalism that compares and optimizes varied LLM steering techniques including fine-tuning, LoRA, and activation steering.
• It develops a preference–utility framework with log-odds metrics to quantify trade-offs between targeted behavior and generation coherence.
• The study proposes SPLIT, a joint preference–utility optimization objective that mitigates utility decay while advancing model preference.

A Unified Analysis of LLM Steering via Dynamic Parameter Dynamics

Introduction and Motivation

Methods for exerting control over the behavior of LLMs—including local weight fine-tuning, parameter-efficient interventions (e.g., LoRA), and activation-based steering—are usually studied separately. This fragmentation obscures the underlying connections, prevents systematic comparison, and may hinder effective development of new control strategies. The paper "Why Steering Works: Toward a Unified View of LLM Parameter Dynamics" (2602.02343) addresses this gap by providing a comprehensive mathematical and empirical framework that unifies these disparate approaches.

Crucially, the authors introduce a "preference–utility" analytical framework, which decomposes the effects of interventions into two independent axes: preference, quantifying the model's propensity toward a desired concept or target attribute, and utility, measuring the validity and coherence of the generated output regardless of preference. By establishing a dynamic parameter update view and studying how these two axes interact under various interventions, the work elucidates both the effectiveness and the trade-offs inherent in LLM steering.

Figure 1: Intervention methods—including local weight fine-tuning, LoRA, and activation steering—can be recast as dynamic weight updates, producing systematic changes in preference and utility with varying intervention strength.

Unified Dynamic Weight Update Formalism

The central technical contribution is the formulation of a unified dynamic weight update model for LLM control. The interventions are cast by introducing perturbations $\Delta\mathbf{W}$ and $\Delta\mathbf{b}$ to a linear transformation in the model:

$$\mathbf{h}_{i+1} = (\mathbf{W} + m_1\, \Delta\mathbf{W}) \mathbf{h}_i + (\mathbf{b} + m_2\, \Delta\mathbf{b})$$

where the magnitudes $m_1$ and $m_2$ control the strength of intervention.

Within this view:

• Local weight fine-tuning: modifies $\mathbf{W}$ and/or $\mathbf{b}$ on selected layers.
• LoRA: introduces a low-rank parameterization for $\Delta\mathbf{W}$.
• Activation steering: adds a vector perturbation to $\mathbf{b}$ or hidden activations, typically interpreted as a bias shift.

This abstraction enables direct comparison and theoretical analysis across techniques that would previously require separate discussions.

Preference–Utility Dynamics: Measurement and Empirical Observations

To characterize the response of LLMs to different interventions, the authors define two log-odds metrics:

• Preference log-odds: Difference in log-likelihood assigned to polarity-paired outputs; operationalizes directional control.
• Utility log-odds: Log-odds that either polarity output (positive or negative) is valid; operationalizes generation coherence and task fidelity.

Experimentally, across diverse models and intervention types, the following regular patterns emerge:

• As the intervention strength $m$ increases, preference log-odds show a consistent three-phase dynamic: linear increase for small $|m|$, a transition phase as out-of-manifold effects accrue, and convergence or plateau as maximal concept expression is reached.
• Utility log-odds typically peak near $m=0$ (no intervention), and reliably degrade with larger $|m|$ as the model is pushed away from its training manifold.

These dynamics are robust over architectures (e.g., Gemma-2-9B-IT, Qwen-2.5-7B-IT) and intervention forms.

Figure 2: Systematic dependence of preference (solid) and utility (dashed) log-odds on steering strength for various intervention types.

Mechanistic Explanation: Activation Manifold Hypothesis

To explain the empirical preference–utility dynamics, the authors propose an activation manifold hypothesis. LLM activations largely reside on a low-dimensional, training-induced manifold, and interventions (e.g., steering) translate activations along certain directions.

• For small perturbations, activations stay near-manifold, so model outputs are valid (utility maintained), and preference can be smoothly increased.
• For larger interventions, activations are moved off-manifold, causing a validity decay: a monotonic, heavy-tailed drop in utility, which is well modeled by a rational quadratic attenuation term.

The increase in preference is governed by the projection of the steering direction onto the target concept axis, but as validity decays, the realized control saturates and may even collapse.

Figure 3: Illustration of steering along a concept direction, showing projection gain and validity decay as activations deviate from the learned activation manifold.

Theoretical fits to log-odds as a function of $m$ achieve $R^2 > 0.95$ in nearly all settings, strongly supporting the manifold-based decay hypothesis.

SPLIT: Joint Preference–Utility Optimization

Building on the mechanistic decomposition, SPLIT (Steering with Preference–UtiLity IntervenTion) is proposed as a training objective that jointly maximizes the gap in preference (using a hinge loss) while minimizing losses on both positive and negative outputs to preserve utility.

Numerical results demonstrate that SPLIT consistently advances preference metrics without the typical trade-off of utility degradation, across multiple LLMs, datasets (e.g., PowerSeeking, AxBench), and intervention paradigms.

Figure 4: Replication of unified preference–utility dynamics in a different LLM (Qwen-2.5-7B-IT) and dataset, affirming robustness.

Figure 5: Effects of varying steering scale on powerseeking behavior; SPLIT yields optimal balance of preference advancement and utility retention.

Implications, Limitations, and Future Research

This work provides a systematic, interpretable foundation for understanding and improving LLM control mechanisms. By recasting control methods in the unified dynamic parameter update view and introducing log-odds-based measurement, practitioners can compare, analyze, and optimize interventions using common language and metrics.

Key implications:

• Utility-preserving interventions are feasible across activation and parameter spaces, as long as off-manifold excursions are quantitatively minimized.
• Explicit trade-off modeling enables principled improvement of existing parameter-efficient and activation-based control methods.
• The activation manifold framework extends to evaluating failure modes, including oversteering, incoherence, and loss of task alignment under strong interventions.

Limitations arise from the reliance on paired-output evaluation, the focus on attribute-level concept control (rather than multi-turn or complex behaviors), and possible deviation from the idealized activation manifold picture in very large or heterogeneous model families.

Future directions include extending the analysis to richer behavioral axes, investigating dynamic/adaptive intervention schedules, and generalizing the manifold hypothesis to accommodate more varied LLM architectures and pretraining regimes.

Conclusion

This study furnishes a unified theoretical and empirical account of LLM parameter dynamics under intervention, clarifying why and how different steering methods operate, and enabling the design of improved preference–utility trade-offs. The manifold-based mechanism and SPLIT objective equip the field with robust tools for principled, effective, and reliable LLM control (2602.02343).