AI Steerability 360 Toolkit

Updated 4 July 2026

AI Steerability 360 is a unified framework for steering large language models across multiple control surfaces, enabling method composition and reproducible benchmarking.
The toolkit supports four distinct control surfaces—input, structural, state, and output—each offering tailored abstractions and intervention techniques.
It enables systematic trade-off analysis and rigorous evaluation through a composable pipeline and use-case specific metrics.

AI Steerability 360 is an extensible, open-source Python library for steering LLMs. It provides a unified framework for lightweight, deliberate control of model behavior across four model control surfaces—input, structural, state, and output—while supplying a common steering pipeline for composing multiple steering methods and an evaluation layer built from use case and benchmark classes. The toolkit is Hugging Face native, released under the Apache 2.0 license, and available at https://github.com/IBM/AISteer360 (Miehling et al., 8 Mar 2026).

1. Scope, motivation, and toolkit lineage

The toolkit was introduced in response to a fragmented steering landscape. The paper states that many steering methods exist but use different semantics and APIs, that current libraries often cover only one intervention surface—especially activation or state steering—and that fair comparison is difficult because methods are designed differently and have different requirements. It also emphasizes that practical steering commonly involves stacked or composed interventions, including examples such as SFT followed by DPO, DPO followed by chain-of-thought prompting, or multiple activation steering methods together. AI Steerability 360 was therefore built to provide a common interface for steering methods, support multiple control surfaces, enable composition, support benchmarking and trade-off analysis, and lower the barrier to developing and comparing steering methods in a controlled, reproducible way (Miehling et al., 8 Mar 2026).

A broader toolkit pattern is visible in earlier “360” systems. AI Fairness 360 organizes fairness work around a standard machine-learning workflow with dataset representations, metrics, explainers, and algorithms, and presents itself as a common platform for sharing and evaluating fairness algorithms (Bellamy et al., 2018). AI Explainability 360 similarly combines an open-source toolkit with a taxonomy intended to help different stakeholders choose appropriate explanation methods rather than assume a single explanation style is sufficient (Arya et al., 2019). This suggests that AI Steerability 360 adopts a related design philosophy: not merely implementing isolated steering algorithms, but building a reusable architecture for method development, composition, and evaluation.

Within steerability research specifically, the toolkit addresses a problem already visible in earlier work: steerability is measurable but unstable across prompts, contexts, model versions, and evaluation schemes. “AI Text-to-Behavior: A Study in Steerability” treated steerability as a quantitative property of prompt-induced behavior and explicitly warned that metric proposals degrade rapidly as models evolve (Noever et al., 2023). AI Steerability 360 operationalizes that concern by making steering a pipeline-level and benchmark-level object rather than a single-number score.

2. Control-surface taxonomy and formalization

The paper defines steering as any deliberate control of model behavior and formalizes four intervention surfaces. These surfaces separate methods by where they act on the inference or training stack rather than by a single algorithmic family.

Control surface	Formal notation	Typical effect
Input control	$p_{\theta}(\sigma(x))$	Prompt rewriting, few-shot insertion, instruction augmentation
Structural control	$p_{\theta'}(x)$	Fine-tuning, adapter layers, weight merging
State control	$p^{h}_{\theta}(x)$	Activation steering, attention steering, residual interventions
Output control	$d(p_{\theta})(x)$	Logit adjustment, constrained decoding, reward-guided search

Input control changes what goes into the model without changing the model itself. The notation $p_{\theta}(\sigma(x))$ uses $\sigma(x)$ for a prompt adapter or transformation applied before inference. In the toolkit, input controls subclass InputControl and override get_prompt_adapter() (Miehling et al., 8 Mar 2026).

Structural control modifies the model’s parameters or architecture. The notation $p_{\theta'}(x)$ denotes a modified model state after steering. The paper lists fine-tuning, adapter layers, weight merging, and task or weight arithmetic as representative cases. Structural controls subclass StructuralControl, and training or weight-modification logic lives in steer() (Miehling et al., 8 Mar 2026).

State control modifies internal activations, hidden states, or attentions during the forward pass without permanently changing the parameters. The notation $p^{h}_{\theta}(x)$ uses $h$ for hooks inserted into the computation, and the paper explicitly describes the intervention as ephemeral and occurring at inference time. State controls subclass StateControl and override get_hooks() (Miehling et al., 8 Mar 2026).

Output control intervenes during decoding. The notation $d(p_{\theta})(x)$ denotes a decoding or transformation function applied on top of the base model. The paper associates this surface with logit adjustment, constrained decoding, alternative sampling methods, reward-guided search, and other decoding-time alignment methods. Output controls subclass OutputControl and override generate() (Miehling et al., 8 Mar 2026).

This four-part taxonomy also clarifies how AI Steerability 360 relates to adjacent steerability paradigms. Prompt-conditioned persona steering in OCEAN experiments belongs to input control, since the model is instructed to “Act in the role of extreme [trait]…” and then evaluated for behavioral alignment (Noever et al., 2023). Steering-vector methods, by contrast, sit squarely within state control: Dialz describes steering vectors as directions in hidden-state space that encode concepts such as honesty, positivity, sycophancy, or stereotypes and are added or subtracted during generation to amplify or weaken a concept (Siddique et al., 4 May 2025).

3. Pipeline architecture, reusable abstractions, and composition

A central abstraction in the toolkit is the SteeringPipeline class. The paper assigns it two main roles: a common interface that defines how controls interact with a model, and a composition mechanism that allows multiple controls to be combined into a single operation. Its core methods are steer(), which performs any needed training or setup, and generate(), which runs inference analogously to Hugging Face model generation. The typical workflow is to instantiate a steering control, pass it into a SteeringPipeline, call steer() if setup is required, and then call generate() for inference (Miehling et al., 8 Mar 2026).

The paper’s concrete state-control example is Contrastive Activation Addition (CAA). CAA learns a steering direction from contrastive examples—positive completions exhibiting a target behavior and negative completions without it—computes a direction vector from residual-stream differences, and injects that vector into hidden states during generation. The example configuration uses VectorTrainSpec(method="mean_diff", accumulate="last_token"), together with a signed multiplier=-10.0, layer_id=15, and token_scope="all". The paper uses this configuration to reduce sycophancy: the unsteered model tends to agree with the user’s viewpoint, whereas the steered model becomes more balanced and less sycophantic (Miehling et al., 8 Mar 2026).

Composition is presented as a first-class property rather than an implementation convenience. The toolkit allows multiple controls to be stacked in one pipeline so that prompt changes, structural modifications, activation interventions, and decoding-time adjustments can be studied jointly. The paper cites a notebook combining PASTA, a state-control method, with DeAL, an output-control method, on TruthfulQA, and reports that this composite steering gives better truthfulness-informativeness trade-offs than using either method alone (Miehling et al., 8 Mar 2026). This makes composition an object of empirical study: complementarity, interference, order effects, and side-effect trade-offs can all be benchmarked rather than assumed.

For activation steering, the toolkit further decomposes many methods into four reusable parts: Estimator, Selector, Transform, and Gate. The Estimator learns a steering artifact, usually a direction vector; the Selector chooses where to intervene, such as which layer or head; the Transform applies the intervention to hidden states; and the Gate decides when the transform fires. The paper states that ActAdd, ITI, and CAA are implemented with this pattern. ActAdd uses SinglePairEstimator, produces a positional direction sequence, and injects only during the initial forward pass. CAA uses MeanDifferenceEstimator, uses multiple contrastive pairs, and produces a non-positional direction. ITI uses ProbeMassShiftEstimator, trains per-head probes across layer-head pairs, uses an accuracy-based TopKHeadSelector, and applies HeadAdditiveTransform. All three use an AlwaysOpenGate, meaning steering occurs on every forward pass (Miehling et al., 8 Mar 2026).

This architecture places AI Steerability 360 in dialogue with end-to-end activation-engineering platforms such as Dialz, which organizes steering-vector research around Datasets, Vectors, Scores, and Visualize, and supports the full workflow from contrastive-pair construction through vector computation, inference-time intervention, scoring, and token-level visualization (Siddique et al., 4 May 2025). A plausible implication is that AI Steerability 360 generalizes beyond that state-steering workflow by embedding it in a broader, multi-surface control framework.

4. Use cases, benchmarks, and trade-off analysis

Evaluation in AI Steerability 360 is organized through two classes: UseCase, which defines the task, and Benchmark, which compares steering pipelines on that task. A use case specifies the evaluation task, how to generate outputs from evaluation data, and how to score outputs with metrics; implementing a use case requires generate() and evaluate(). The paper’s running example is instruction following, based on IFEval. Each evaluation datum includes an ID, a prompt with split natural-language instructions, an instruction_id_list, and required arguments in kwargs. Example instructions include writing at least 500 words, including specific keywords, and not using commas (Miehling et al., 8 Mar 2026).

The instruction-following use case is evaluated with two metrics. StrictInstruction is a custom metric that measures adherence to the instructions, and RewardScore is a generic metric that measures response quality using a reward model. The paper presents this pairing as an instance of multi-objective evaluation: one metric measures the targeted behavior, while the other monitors response quality (Miehling et al., 8 Mar 2026). This is important because the toolkit does not treat successful steering as a scalar objective divorced from collateral effects.

The Benchmark class supports fixed controls, variable controls, repeated trials, and runtime overrides for inference-time-only parameters. For fixed-control benchmarking, the paper compares a baseline model and a model steered with PASTA. PASTA selectively rescales attention scores at certain heads, downweights tokens outside the target span, and steers the model’s focus toward important prompt tokens. In the benchmark example, runtime_overrides is used because the substrings to emphasize are known only when the prompt is available. The benchmark also uses num_trials=10 to handle stochastic generation variability (Miehling et al., 8 Mar 2026).

For variable-control benchmarking, the paper introduces ControlSpec, which specifies a control class, fixed parameters, and variable parameters to sweep. Its example varies PASTA’s alpha over [5, 10, 15, 20, 25, 30], while keeping head_config and scale_position fixed. The variable space can also be defined via Cartesian grids or lambda-based functional relationships. The resulting trade-off analysis includes a Pareto frontier plot, and the paper notes a sweet spot around $p_{\theta'}(x)$ 0, beyond which both response quality and instruction-following performance worsen (Miehling et al., 8 Mar 2026).

This evaluation design resonates with other steerability benchmarks that emphasize multidimensional rather than binary success. “Visual AI and Linguistic Intelligence Through Steerability and Composability” treats multimodal control as a family of task types across 14 major tasks and 800 sequences of human-machine instruction, using iterative task completion and assembly or disassembly success as practical scoring criteria (Noever et al., 2023). “When is Your LLM Steerable?” adds a three-way outcome taxonomy—UnderSteer, SuccSteer, and OverSteer—and uses macro-F1 to evaluate a predictor trained on 1.4M steered generations (Fan et al., 10 Jun 2026). Together, these works suggest that AI Steerability 360’s use-case framework is well matched to a field in which prompt adherence, quality preservation, multimodal consistency, and hyperparameter sensitivity all matter simultaneously.

5. Implementation model, reproducibility, and stated limitations

The toolkit is Hugging Face native and open source under the Apache 2.0 license. The paper explicitly points to the public GitHub repository at https://github.com/IBM/AISteer360 and to repository notebooks and paths such as examples/notebooks/control_caa/caa.ipnyb, evaluation/use_cases/instruction_following/use_case.py, and algorithms/state_control/pasta/control.py (Miehling et al., 8 Mar 2026). This positions the library as both a research platform and a reproducible software artifact.

The paper emphasizes several practical advantages. These include a unified interface across input, structural, state, and output controls; support for composition; benchmarking support; trade-off analysis; reusable abstractions, especially for activation steering; Hugging Face-native integration; and open-source availability (Miehling et al., 8 Mar 2026). The intended users are researchers building new steering methods, practitioners comparing steering methods fairly, and those studying behavior change, side effects, and method interactions under task-specific metrics.

The limitations section is equally explicit. First, the toolkit depends on Hugging Face, which the authors chose for flexibility and access to internals, but describe as slower than optimized inference runtimes such as vLLM. Second, inference speed is identified as a limitation, especially for large-scale evaluation. Third, state steering requires Hugging Face generation because of API restrictions. Fourth, best-parameter selection remains hard both conceptually and computationally: the toolkit helps explore parameters but does not solve optimization fully. Fifth, unknown side effects remain a central concern: a steering method can improve the target behavior while degrading other behaviors, and benchmarks may miss unmonitored dimensions (Miehling et al., 8 Mar 2026).

The paper mentions possible future support for vLLM.hook and possible hyperparameter-optimization tools (Miehling et al., 8 Mar 2026). This suggests that AI Steerability 360 is designed as an extensible base layer rather than a closed catalog of methods. That stance is consistent with earlier “360” toolkits: AIF360 stresses extensibility, common abstractions, benchmarking, and testing infrastructure (Bellamy et al., 2018), while AIX360 pairs a broad method set with a taxonomy that exposes coverage gaps and motivates further additions (Arya et al., 2019).

6. Position within steerability research and adjacent toolkits

AI Steerability 360 emerged in a research environment where steerability had already been probed from multiple angles but lacked a unified software framework. An early quantitative strand used behavioral scoring. “AI Text-to-Behavior: A Study in Steerability” employed the OCEAN personality model to gauge prompt-induced behavioral alignment and reported that conscientiousness and neuroticism were the most distinctly evoked traits, openness was linguistically ambiguous, and extroversion and agreeableness overlapped noticeably. The same paper used historical figure simulations to show instructibility and persona internalization, while warning that rapid model advances and opaque training techniques can cause metric proposals to degrade quickly (Noever et al., 2023). AI Steerability 360 can be read as a response to precisely that instability: it supplies abstractions for method comparison and benchmarking rather than treating any one metric as definitive.

A second strand developed state-steering toolkits. Dialz presents an end-to-end Python framework for steering vectors on open-source LLMs, with modules for contrastive-pair datasets, steering-vector computation, activation scoring, and visualization. It supports PCA-based and mean-difference steering vectors, applies interventions at selected layers during inference, and demonstrates stereotype mitigation and layer-wise hallucination analysis (Siddique et al., 4 May 2025). AI Steerability 360 subsumes comparable state-control workflows within a broader taxonomy that also covers prompt, structural, and decoding interventions.

A third strand concentrated on multimodal and sequential evaluation. “Visual AI and Linguistic Intelligence Through Steerability and Composability” defines steerability as the model’s ability to be guided toward a specific outcome or follow a set of instructions, and composability as the ability to combine disparate elements to create something new. Its 14-task, 800-dialog setup reveals substantial difficulty differences across task families, from low-difficulty bartender and spreadsheet tasks to high-difficulty negotiation, genetic programming, game self-play, and satellite image analysis (Noever et al., 2023). This suggests that AI Steerability 360’s use-case abstraction can, in principle, serve not only single-turn text steering but also interaction-heavy, multimodal, and sequential evaluation regimes.

A fourth strand addresses prediction and optimization of steering success. “When is Your LLM Steerable?” introduces ASTEER, a testbed with 1.4M steered generations spanning 150 concepts, three LLMs, and two steering methods, and trains a GBDT classifier on early hidden-state features to predict UnderSteer, SuccSteer, or OverSteer. The paper reports around 0.7 macro-F1 on unseen concepts and shows that prediction-guided steering-strength search can approach near-optimal performance while using only a small fraction of exhaustive decoding cost (Fan et al., 10 Jun 2026). This provides a natural complement to AI Steerability 360’s benchmark layer: the toolkit can compare controls, while predictor-based methods can reduce the cost of exploring their operating regimes.

Finally, steerability has been formalized beyond LLMs. ReSteer studies multitask robot policies and defines steerability as the ability to respond properly to a new instruction at an arbitrary intermediate state. It introduces the Steerability Coverage Ratio, uses conditional mutual information as a rollout-free proxy for instruction sensitivity, and reports an 11% improvement over 18k rollouts in simulation together with 2.2× higher steerability in real-world experiments (Chen et al., 18 Mar 2026). A plausible implication is that AI Steerability 360’s control-surface abstraction belongs to a broader research shift in which steerability is treated as a measurable systems property rather than an informal prompting skill.

7. Conceptual significance

The central contribution of AI Steerability 360 is architectural. Its novelty does not lie only in implementing individual steering methods, but in framing steering as a software and evaluation problem defined by intervention surface, pipeline composition, use-case-specific metrics, and explicit trade-off analysis. That framing is consequential because steering is not monolithic: prompt adaptation, model editing, activation manipulation, and decoding control have different semantics, operating costs, and failure modes, yet often need to be combined in practice (Miehling et al., 8 Mar 2026).

The toolkit also formalizes an important methodological shift. Rather than asking whether a model is steerable in the abstract, it asks which control surfaces are being used, how methods compose, on which tasks they are evaluated, which side effects are monitored, and how parameter sweeps change the Pareto frontier between target behavior and other qualities. This suggests that steerability should be analyzed as a multidimensional systems property, analogous to how fairness and explainability toolkits moved their respective fields away from one-off demonstrations toward shared abstractions, metrics, and benchmarkable workflows (Bellamy et al., 2018, Arya et al., 2019).

In that sense, AI Steerability 360 occupies a specific place in the research landscape. It is a unifying LLM toolkit for steering across major intervention surfaces, composition patterns, and evaluation regimes; a bridge between isolated steering methods and controlled comparison; and an extensible platform for studying not only whether steering works, but also when it works, how it fails, and what it costs (Miehling et al., 8 Mar 2026).