Robust Steering

Robust Steering is a Python package for steering language-model representations with robust control. It models representation dynamics across transformer layers and uses finite-horizon H∞ control to steer toward semantic targets under distribution shift.

The package brings together controller synthesis, activation interventions, and evaluation across truthfulness, safety, multilingual reasoning, and long-context tasks. It supports closed-loop steering without updating model weights, with LQR and activation-steering baselines for comparison.

Installation · Quick start · Benchmarks · CLI · Documentation

Features

• Robust controller synthesis. Finite-horizon H∞ control with disturbance geometry estimated from representation-dynamics residuals.
• Representation-level steering. Semantic target calibration, reduced-state dynamics, and activation interventions during generation.
• A shared controller interface. H∞, LQR, PID, and activation-addition controllers, with model adapters separated from controller mathematics.
• Benchmark integration. Task-specific calibration and evaluation of steering effectiveness, transfer, and capability preservation.

Installation

Requires Python 3.11 or newer.

git clone https://github.com/rgs2151/controller.git
cd controller
python -m pip install -e .

Import the package as robust_steerability. A Conda environment is also provided in environment.yml:

conda env create -f environment.yml
conda activate robust-steerability

Language-model benchmarks require CUDA GPUs and access to the selected models. Benchmark-specific requirements are documented in the guides below.

Quick start

Synthesize and apply an H∞ controller using the standalone control API. This example runs on CPU:

import torch
from robust_steerability import FiniteHorizonControlProblem, HInfinityController

# Four-step dynamics with a two-dimensional state.
identity = torch.eye(2)
layers = identity.repeat(4, 1, 1)

problem = FiniteHorizonControlProblem(
    dynamics=0.9 * layers,
    control_channels=layers,
    disturbance_channels=0.1 * layers,
    state_costs=layers,
    control_costs=layers,
    terminal_cost=identity,
)
controller = HInfinityController.synthesize(problem, device="cpu")

state_deviation = torch.tensor([[1.0, -0.5]])
intervention = controller.intervention(0, state_deviation)

FiniteHorizonControlProblem specifies the dynamics, control and disturbance channels, and performance costs. The synthesized controller maps the current state deviation to an intervention. In language-model applications, runtime policies connect this interface to model representations and semantic targets.

See the controller API and package guide for integration details.

Benchmarks

The repository includes benchmarks for both steering the intended behavior and preserving capabilities while steering a separate behavior.

Benchmark	Steering target	Evaluation
TruthfulQA	Truthfulness	English and Spanish-input questions, with English output.
HarmBench	Safe refusal	Direct harmful requests and jailbreak variants.
MGSM	Spanish-language output	Mathematical accuracy across input languages.
L-CiteEval	Positive sentiment	Answer and citation quality in long contexts.
RealToxicityPrompts	Reduced toxicity	Toxicity and response quality.

Each benchmark provides artifact preparation, calibration, evaluation, and scoring through python -m robust_steerability.benchmarks.<benchmark>. The linked guides contain supported models, commands, and evaluation protocols.

Command-line interface

Run benchmark commands from the repository root. The following example evaluates Llama-3-8B on TruthfulQA, comparing the original model, A-LQR, and H∞ on English and Spanish-input questions:

# Prepare model and dataset artifacts.
python -m robust_steerability.benchmarks.truthfulness artifacts \
  --model llama8b --devices auto

# Calibrate the steering methods.
python -m robust_steerability.benchmarks.truthfulness calibrate \
  --model llama8b --methods alqr,h_infinity --devices auto

# Evaluate the selected methods and conditions.
python -m robust_steerability.benchmarks.truthfulness evaluate \
  --model llama8b --methods original,alqr,h_infinity \
  --datasets id,spanish --devices auto

# Score the generated responses.
python -m robust_steerability.benchmarks.truthfulness score \
  --model llama8b --methods original,alqr,h_infinity \
  --datasets id,spanish --scorers default --devices auto

--devices auto uses the available CUDA GPUs; use --devices cuda:0 to select a single GPU. API-based judges require OPENAI_API_KEY. Model access and scorer requirements are described in each benchmark guide.

HarmBench, MGSM, L-CiteEval, and RealToxicityPrompts use the same stage interface through the harmful, mgsm, lciteeval, and toxicity modules, respectively. Consult each module's help for supported model keys and options:

python -m robust_steerability.benchmarks.truthfulness --help
python -m robust_steerability.benchmarks.harmful --help

Documentation

• Package guide — model integration, calibration, and runtime architecture.
• Controller API — synthesis interfaces and mathematical conventions.
• Benchmark guide — evaluation pipelines and benchmark configuration.
• Reproducibility conventions — scientific settings and baseline choices.
• Baseline implementations — source methods and attribution.

Repository structure

The implementation lives in robust_steerability/. Benchmark specifications and results are in benchmarks/; supporting figures and analyses are maintained separately in figs/ and parking/. Reference material remains in ref/.

The Python package is independent of the supplementary analysis folders. Incorporated reference implementations retain their upstream attribution and applicable licenses.

Citation

If you use Robust Steering in your research, please cite the accompanying manuscript:

@unpublished{fallah2026mindcontrol,
  title  = {Mind Control of {LLMs} via Closed-Loop Stimulation: Out-of-Distribution Robustness},
  author = {Fallah, Kasra and Singha, Rudramani and Ghanei, Hannah and Sajda, Paul and Anderson, James and Zabeh, Erfan},
  year   = {2026},
  note   = {Manuscript}
}