Endogenous Steering Resistance (ESR)

• Endogenous Steering Resistance (ESR) is a system-internal mechanism that detects and counteracts external steering inputs across diverse fields such as AI, vehicle dynamics, and finance.
• In machine learning, ESR is quantified through multi-attempt rates and targeted latent circuit interventions, revealing measurable self-correction capabilities in models.
• In vehicle control and financial systems, rigorous ESR modeling with compensators and fixed-point resistance functions improves path tracking precision and optimizes trading executions.

Endogenous Steering Resistance (ESR) refers to system-internal mechanisms that detect, react to, or counteract externally applied steering inputs or interventions, thereby limiting exogenous control. The concept appears in several disparate scientific contexts: vehicular dynamics, where physical resistance opposes steering actuations; neural LLMs, where models self-correct against adversarial activation steering; and financial markets, where endogenous responses of sophisticated traders dampen metaorder-driven price distortions. Each domain develops rigorous definitions, models, and analysis methods to characterize the causal structure and implications of ESR, whether for robust control, AI alignment, or optimal trading under market impact.

1. ESR in Machine Learning: Model-Intrinsic Resistance to Activation Steering

In LLMs, Endogenous Steering Resistance is defined as the model's capacity for spontaneous recovery from adversarial or misaligned activation steering applied at inference, even while steering remains active. ESR is quantified via the ESR rate,

$$R_{\mathrm{ESR}} = R_{\mathrm{multi}} \times R_{\mathrm{imp}\mid\mathrm{multi}},$$

where $R_{\mathrm{multi}}$ is the rate at which a model produces multiple attempted responses (as detected by explicit self-correction markers), and $R_{\mathrm{imp}\mid\mathrm{multi}}$ is the conditional rate that the subsequent attempt improves over the first (scored by a reference judge). In Llama-3.3-70B, ESR emerges as a measurable phenomenon: with sparse autoencoder-based activation steering at calibrated strengths, $R_{\mathrm{multi}}=7.4\%$ and $R_{\mathrm{ESR}}=3.8\%$, with significantly lower incidence in smaller models such as Llama-3.1-8B or Gemma-2 (McKenzie et al., 6 Feb 2026).

Analysis reveals that ESR in LLMs is mediated by specific latent features—interpreted as "off-topic detector" circuits—in the model’s internal representation. Zero-ablating 26 such SAE-derived latents causally reduces the multi-attempt and ESR rates by 25–27%. Interventions including meta-prompts ("self-monitor" instructions) and fine-tuning with explicit self-correction data further increase ESR, demonstrating that these circuits are both detectable and inducible.

2. ESR in Physical Control: Steering Resistance in Vehicle Systems

In vehicle dynamics, endogenous steering resistance describes the self-generated torque or force that opposes steering motion, arising from physical interactions (e.g., tire-road contact, mechanical linkages). This resistance must be explicitly modeled for precision path-following, especially under uncertainty in surface conditions. The equation modeling endogenous resistance adopts a first-order quasi-steady form:

$\dot{\delta} = u - C v \delta$

where $\delta$ is the steering wheel angle (now treated as a state variable), $u$ is the steering command, $v$ is vehicle speed, and $C$ is the steering resistance coefficient (Iwai et al., 21 Apr 2026).

Such explicit modeling is integrated into robust path tracking control by augmenting the vehicle’s system state and introducing a Model Error Compensator (MEC) to address mismatches in $R_{\mathrm{multi}}$0. Numerical results indicate that conventional controllers (which neglect or simplify steering resistance) suffer rapid degradation and even divergence in the tracking error under parameter mismatch, while the MEC-augmented approach maintains sub-meter error across a wide interval of $R_{\mathrm{multi}}$1 values, where $R_{\mathrm{multi}}$2 is the nominal coefficient. This underscores the critical control-theoretic role of endogenous resistance in defining the achievable path-following performance envelope.

3. ESR in Financial Markets: Market Resistance to Metaorder Impact

In quantitative finance, endogenous resistance models sophisticated traders or liquidity providers who actively detect and counteract large orders (metaorders), opposing price impact and stabilizing market dynamics. In classical propagator models, the impact of a trade rate $R_{\mathrm{multi}}$3 on price is

$R_{\mathrm{multi}}$4

with $R_{\mathrm{multi}}$5 a decaying kernel. ESR augments this by positing an opposing flow $R_{\mathrm{multi}}$6, itself a functional of $R_{\mathrm{multi}}$7, yielding

$R_{\mathrm{multi}}$8

Here $R_{\mathrm{multi}}$9 is defined as the solution to a Volterra fixed-point equation parameterized by a Lipschitz "resistance function" $R_{\mathrm{imp}\mid\mathrm{multi}}$0, which may be linear or strictly convex (Chahdi et al., 6 Jan 2026).

Optimal execution strategies in this framework are then governed by a nonlinear stochastic Fredholm equation, with existence, uniqueness, and exponential convergence established in both linear and certain nonlinear parameter regimes. This model enables quantitative analysis of how endogenous market resistance attenuates round-trip price impact and constrains the aggressiveness of optimal trading relative to models without ESR.

4. Methodologies for Identifying and Quantifying ESR

Across domains, ESR is formally characterized through a combination of structural modeling, causal analysis, and quantitative metrics:

• In LLMs: SAE-based latent decomposition permits targeted steering and ablation, with downstream effects measured by multi-attempt and ESR rates based on scored, judge-segmented completions (McKenzie et al., 6 Feb 2026). Statistical significance is assessed (e.g., Cohen’s $R_{\mathrm{imp}\mid\mathrm{multi}}$1 for "off-topic detector" latents).
• In Vehicle Control: Robustness to unmodeled ESR is quantified via the maximum path-following error $R_{\mathrm{imp}\mid\mathrm{multi}}$2 under controlled parametric sweep of $R_{\mathrm{imp}\mid\mathrm{multi}}$3, with explicit error tables for different path types (Iwai et al., 21 Apr 2026).
• In Market Microstructure: Analytical existence, numerical convergence rates, and execution cost metrics under various resistance convexity specifications provide quantifiable measures of ESR's effect on strategy and cost (Chahdi et al., 6 Jan 2026).

5. Implications for Robustness, Alignment, and Control

ESR confers both protective and limiting effects depending on context:

• AI Models: ESR enhances adversarial robustness by enabling internal correction of irrelevant or malicious steering, potentially serving as a safeguard mechanism. However, it can also resist safety-critical interventions (e.g., reducing toxicity via activation modification), necessitating careful management—either to enhance (for defense) or diminish (to allow alignment) (McKenzie et al., 6 Feb 2026).
• Autonomous Vehicles: Correctly accounting for steering resistance is essential for ensuring reliability across environmental variation. Neglecting ESR introduces critical path-tracking failures as surface conditions change (Iwai et al., 21 Apr 2026).
• Financial Markets: Endogenous resistance models real-world liquidity provision and prevents unrealistic profit exploitation (dynamic arbitrage) in execution strategies, but requires nonlinear control-theoretic treatment due to its fixed-point nature (Chahdi et al., 6 Jan 2026).

6. Limitations and Future Directions

• Machine Learning: Empirical ESR rates are reported on language benchmarks and synthetic self-correction tasks, but their generalization to other architectures and tasks remains open. Circuit-level mapping of OTD (off-topic detector) latents is preliminary; full interpretability and transferability await further study (McKenzie et al., 6 Feb 2026).
• Vehicle Control: Simulation results assume full-state measurement and linearized resistance; real vehicles experience unmodeled nonlinearities, noise, and state estimation challenges. Ongoing work includes closed-loop stability proofs, observer design, and experimental validation (Iwai et al., 21 Apr 2026).
• Finance: ESR analysis depends on kernel and resistance function assumptions. Calibration to real market data and integration of cross-player learning remain outstanding challenges. Iterative solution schemes scale exponentially with problem complexity, motivating research in efficient solvers (Chahdi et al., 6 Jan 2026).

7. Comparative Table of ESR Across Domains

Domain	ESR Mechanism	Quantification
LLMs	OTD circuits, self-correct markers	$R_{\mathrm{imp}\mid\mathrm{multi}}$4, ablation effect
Vehicle Control	$R_{\mathrm{imp}\mid\mathrm{multi}}$5 torque, MEC compensation	Max tracking error $R_{\mathrm{imp}\mid\mathrm{multi}}$6
Financial Markets	Resistance flow $R_{\mathrm{imp}\mid\mathrm{multi}}$7, fixed-point $R_{\mathrm{imp}\mid\mathrm{multi}}$8	Execution cost, convergence rate

Each field employs domain-specific modeling and diagnostics, but across contexts, ESR is fundamentally the emergent system response that limits the direct efficacy of external steering, demanding adaptive strategies for robust, controllable, and trustworthy operation.