Order Is Not Control

2026-06-11 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The paper "Order Is Not Control" challenges the assumption that identifying order-inducing objects in AI alignment, interpretability, and steering studies equates to control. It posits that true control necessitates a "receiver-gated response law," defined as a denominator-indexed operator that maps material state, action/drive, bath, and receiver state to response displacement, sinks, effort, and basin projection. This law is local, with interventions admitted, saturated, sign-changing, leaky, or overdriven based on various medium and receiver conditions. Evidence is presented from biological systems like Mouse ALM, C. elegans, and zebrafish, alongside LLM panels. LLM response vectors show predictability at 72.8-73.7% component-sign accuracy, rising to 84.3-84.8% on nonzero components, with held-out observers predicting system-effect and target/oracle families at 93.6% and 91.7% accuracy. The work describes a driven-dissipative response-system at the mesoscopic control level, supporting local admitted control and measurable stochastic response operators.

Key takeaway

For AI Scientists developing steerable models, understanding that "order is not control" is crucial. You should focus on implementing explicit receiver-gated response laws rather than merely identifying order-inducing objects. This approach enables more precise, admitted control over model outputs, as demonstrated by predictable LLM response vectors. Consider designing systems that measure and bound effort for target movement.

Key insights

The paper distinguishes "order" from "control," defining control via local, receiver-gated response laws demonstrated across biological and LLM systems.

Principles

• Control requires a receiver-gated response law.
• Response laws are local and context-dependent.
• Finite effort must move a target under bounded conditions.

Method

The paper identifies control by analyzing receiver-gated response laws across biological and LLM panels, measuring response vector predictability and observer accuracy for system effects and target families.

In practice

• Predict LLM response vectors using component-sign accuracy.
• Reshape susceptibility with constitution-conditioned adapters.
• Separate opportunity from action policies via stochastic operators.

Topics

• AI Alignment
• LLM Control
• Neural Perturbation
• Response Laws
• System Steering
• Interpretability

Best for: AI Scientist, Research Scientist

Related on AIssential

• Week Ending 2.8.2026 — AI research is advancing model robustness, efficiency, and real-world applicability across diverse domains.
• Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning — Agentic Chain-of-Thought Steering (ACTS) adaptively guides LLM reasoning via a controller agent to optimize efficiency and control.
• When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search — Activation steering variability often reflects search difficulty, not representational feasibility, addressable via geometry-guided optimization.
• Your LLM Can’t Track a Light Switch. A 1958 Math Trick Fixes It. — A 1958 mathematical technique enables dynamic positional encoding, fixing static attention mechanism flaws in LLMs.
• Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control — LLM transformer dynamics are locally linear, enabling efficient, model-based optimal control for activation steering.
• What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal — Steering vectors primarily influence LLM behavior by interacting with the attention mechanism's OV circuit.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.