How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities

Ziwen Xu; Kewei Xu; Haoming Xu; Haiwen Hong; Longtao Huang; Hui Xue; Ningyu Zhang; Yongliang Shen

cs.CLMar 3, 2026

How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities

Authors: Ziwen Xu, Kewei Xu, Haoming Xu, Haiwen Hong, Longtao Huang, Hui Xue, Ningyu Zhang, Yongliang Shen,

Organizations: Zhejiang University · Alibaba Group

Abstract

Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerEval, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.

Figures & tables

Figure 1: Behavioral control targets can be organized by granularity. For example, the target of autonomy progresses from a high-level objective ( Level 1 ), to a constrained manner of expression ( Level 2 ), and finally to a directly checkable surface realization ( Level 3 ).

Figure 2: Example cases from the three domains Personality , Sentiment , and Language Features across the L1 $\sim$ L3 hierarchies. Taking Language Features as an example, the core steering goal is to increase redundancy. At Level 1 (L1) , the model is guided to express the general intent “Increase redundancy”, shifting from “Concise phrasing” to “Elaborative repetition”. At Level 2 (L2) , the steering specifies a strategy for realization, moving from a “Single expression” to a “Rephrased restatement”. At Level 3 (L3) , atomic, verifiable markers are enforced, requiring the inclusion of “ (i.e., ”. These examples illustrate how each level progressively constrains model outputs from abstract intent to concrete surface evidence. Further details are provided in § 3 .

Level	Frequency	Abstraction	Description
L1	High	Highest	What to express
L2	Medium	Moderate	How to express it
L3	Low	Lowest	How to instantiate it

Table 1: Relationship between granularity levels, their typical occurrence frequency in natural text, and abstraction. Finer-grained targets are less frequent and less abstract, but more directly verifiable.

Figure 3: Automated data synthesis pipeline.

Figure 4: The hierarchical structure and sample distribution of our dataset.

Method	Language Features						Personality						Sentiment
	L1		L2		L3		L1		L2		L3		L1		L2		L3
	CS	HM	CS	HM	CS	HM	CS	HM	CS	HM	CS	HM	CS	HM	CS	HM	CS	HM
Gemma-2-9b-Instruct
Vanilla	1.16	1.38	0.95	1.14	0.14	0.15	0.45	0.58	0.79	1.01	0.05	0.06	1.40	1.61	1.18	1.40	0.00	0.00
Prompt (0-shot)	2.53	2.72	2.84	3.03	2.85	3.21	2.57	2.99	3.02	3.21	2.87	3.17	2.87	3.18	3.15	3.39	2.57	2.99
Prompt (3-shot)	2.32	2.60	2.99	3.14	2.88	3.19	2.71	3.10	2.94	3.27	3.18	3.47	2.97	3.35	2.94	3.24	2.37	2.71

Table 2: Performance across domains, granularity levels, and metrics. Each domain includes three granularity levels L1 to L3. We report Concept Score (CS) on a 0–4 scale and Harmonic Mean (HM) on the same scale. HM is the harmonic mean of Concept Score, Instruction Score, and Fluency Score. Best and second-best results are highlighted within each model block.

Panel of Figure 5 — Figure 5: Experimental results in terms of few-shot analysis and steering strength.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Field specifications of the data entry. The domain and concept fields define the hierarchical subject matter, which is probed by the question . Matching and not_matching serve as contrastive responses for model steering.

Concept	Gemma-2-9B-it			Qwen-2.5-7B-it			Llama-3.1-8B-it
Concept	PCA	DiffMean	RePS	PCA	DiffMean	RePS	PCA	DiffMean	RePS
L1_1	3	4	10	4	4	3	4	3	7
L2_1	4	7	20	5	6	5	4	5	8
L3_1	1	1	22	1	1	8	1	1	8
L1_2	2	7	22	8	7	8	6	3	8
L2_2	4	6	10	1	8	8	1	3	3

Appendix

Table 3: Steering factors in the Language Features domain when applying steering vectors across all concepts and models.

Concept	Gemma-2-9B-it			Qwen-2.5-7B-it			Llama-3.1-8B-it
Concept	PCA	DiffMean	RePS	PCA	DiffMean	RePS	PCA	DiffMean	RePS
L1_1	1	6	18	1	6	3	3	3	8
L2_1	1	4	10	1	5	3	4	3	4
L3_1	1	1	16	1	1	8	1	1	7
L1_2	4	5	12	1	5	2	5	4	2
L2_2	7	4	10	5	4	6	8	4	5

Appendix

Table 4: Steering factors in the Personality domain when applying steering vectors across all concepts and models.

Concept	Gemma-2-9B-it			Qwen-2.5-7B-it			Llama-3.1-8B-it
Concept	PCA	DiffMean	RePS	PCA	DiffMean	RePS	PCA	DiffMean	RePS
L1_1	7	8	10	7	6	6	5	6	1
L2_1	5	8	18	4	8	5	3	1	1
L3_1	1	1	10	8	1	1	1	1	1
L1_2	1	6	12	4	6	5	4	5	4
L2_2	4	5	10	7	3	3	4	3	3

Appendix

Table 5: Steering factors in the Reasoning Patterns domain when applying steering vectors across all concepts and models.

Concept	Gemma-2-9B-it			Qwen-2.5-7B-it			Llama-3.1-8B-it
Concept	PCA	DiffMean	RePS	PCA	DiffMean	RePS	PCA	DiffMean	RePS
L1_1	1	8	12	2	8	3	7	8	8
L2_1	8	8	10	8	7	2	7	7	4
L3_1	1	6	10	1	1	1	1	1	1
L1_2	1	8	12	5	8	6	3	7	3
L2_2	5	5	14	1	8	8	8	8	4

Appendix

Table 6: Steering factors in the Sentiment domain when applying steering vectors across all concepts and models.

Method	L1				L2				L3
Method	CS	IS	FS	HM	CS	IS	FS	HM	CS	IS	FS	HM
Gemma-2-9b-Instruct
Vanilla	1.16	3.94	3.70	1.38	0.95	3.94	3.71	1.14	0.14	3.92	3.68	0.15
Prompt (0-shot)	2.53	3.93	3.83	2.72	2.84	3.92	3.78	3.03	2.85	3.93	3.66	3.21
Prompt (3-shot)	2.32	3.96	3.93	2.60	2.99	3.89	3.82	3.14	2.88	3.95	3.53	3.19
PCA	1.94	3.31	3.60	1.85	1.45	3.35	3.62	1.51	0.13	3.91	3.70	0.15

Appendix

Table 7: Detailed experimental results for the Language Features domain across three granularity levels (L1–L3). We report the Concept Score (CS), Instruction Score (IS), Fluency Score (FS), and their Harmonic Mean (HM). All metrics are evaluated on a 0–4 scale.

Method	L1				L2				L3
Method	CS	IS	FS	HM	CS	IS	FS	HM	CS	IS	FS	HM
Gemma-2-9b-Instruct
Vanilla	0.45	3.86	3.72	0.58	0.79	3.90	3.72	1.01	0.05	3.97	3.65	0.06
Prompt (0-shot)	2.57	3.83	3.92	2.99	3.02	3.73	3.91	3.21	2.87	3.90	3.59	3.17
Prompt (3-shot)	2.71	3.87	3.96	3.10	2.94	3.84	3.91	3.27	3.18	3.99	3.60	3.47
PCA	1.33	3.22	3.45	1.48	1.51	2.52	3.25	1.20	0.05	3.94	3.66	0.06

Appendix

Table 8: Detailed experimental results for the Personality domain across three granularity levels (L1–L3). We report the Concept Score (CS), Instruction Score (IS), Fluency Score (FS), and their Harmonic Mean (HM). All metrics are evaluated on a 0–4 scale.

Method	L1				L2				L3
Method	CS	IS	FS	HM	CS	IS	FS	HM	CS	IS	FS	HM
Gemma-2-9b-Instruct
Vanilla	0.82	3.88	3.64	1.04	0.78	3.90	3.64	0.94	0.00	3.95	3.60	0.00
Prompt (0-shot)	3.03	3.86	3.59	3.08	3.53	3.90	3.60	3.52	2.90	3.95	3.61	3.17
Prompt (3-shot)	3.47	3.98	3.68	3.60	3.28	3.94	3.66	3.34	3.33	3.99	3.60	3.54
PCA	1.30	3.70	3.60	1.46	1.76	3.68	3.37	1.84	0.01	3.95	3.67	0.02

Appendix

Table 9: Detailed experimental results for the Reasoning Patterns domain across three granularity levels (L1–L3). We report the Concept Score (CS), Instruction Score (IS), Fluency Score (FS), and their Harmonic Mean (HM). All metrics are evaluated on a 0–4 scale.

Method	L1				L2				L3
Method	CS	IS	FS	HM	CS	IS	FS	HM	CS	IS	FS	HM
Gemma-2-9b-Instruct
Vanilla	1.40	3.74	3.67	1.61	1.17	3.85	3.77	1.40	0.00	3.77	3.70	0.00
Prompt (0-shot)	2.87	3.86	3.92	3.18	3.15	3.88	3.90	3.39	2.57	3.81	3.85	2.99
Prompt (3-shot)	2.97	3.93	3.95	3.35	2.94	3.93	3.95	3.24	2.37	3.89	3.90	2.71
PCA	1.86	3.42	3.67	2.01	1.68	3.37	3.65	1.75	0.00	3.75	3.63	0.00

Appendix

Table 10: Detailed experimental results for the Sentiment domain across three granularity levels (L1–L3). We report the Concept Score (CS), Instruction Score (IS), Fluency Score (FS), and their Harmonic Mean (HM). All metrics are evaluated on a 0–4 scale.

Level	Spearman	Pearson	QWK
L1	0.81	0.83	0.83
L2	0.83	0.84	0.84
L3	0.96	0.95	0.92
All	0.86	0.87	0.87

Appendix

Table 11: Correlation with human judgments and Quadratic Weighted Kappa (QWK).

LLM Evaluation

Large language model evaluation measures what models can do, how reliably they do it, and whether benchmark scores predict real-world performance.

See all 2,408 papers

RealControlBench: A Benchmark for Interpretability-Informed Fine-Grained Control of LLMs and Agents
Oct 19, 2026Xinying Cai, Bin Hu, Mingyu Jin +16Language Model Steering Language Model-Based Control
University of California, San Diego · Rutgers University · NEC Laboratories America +8
FaithSteer-BENCH: A Deployment-Aligned Stress-Testing Benchmark for Inference-Time Steering
Mar 18, 2026Zikang Ding, Qiying Hu, Yi Zhang +4Language Model Steering Language Model Robustness
University of Electronic Science and Technology of China · Mohamed bin Zayed University of Artificial Intelligence · South China University of Technology
Measuring Activation Control in Large Language Models
Aug 21, 2026Marek Mateusz Kowalski, Joshua Fonseca Rivera, Uzay Macar +1LLM Safety Evaluation LLM Auditing
UK AI Security Institute

Oct 19, 2026J/K move · Enter open · S save

RealControlBench: A Benchmark for Interpretability-Informed Fine-Grained Control of LLMs and Agents

Xinying Cai, Bin Hu, Mingyu Jin, Qiming Zhang+15

University of California, San Diego · Rutgers University · NEC Laboratories America+8

Language Model Steering Language Model-Based Control AI Agent Benchmarks LLM Agent Evaluation

Open paper AskPDF arXiv

Existing steering benchmarks are built around simplified, isolated generation tasks and evaluate a single behavioral attribute under controlled settings that do not reflect the diversity and complexity of practical deployment. As a result, performance on these benchmarks provides only a limited picture of how steering methods perform in the real world. We therefore introduce RealControlBench, a comprehensive benchmark for evaluating fine-grained steering of LLMs and LLM agents across diverse behavioral targets and interaction settings. RealControlBench extends steering evaluation beyond isolated generations to complex, execution-grounded, multi-turn agent interactions, covering five complementary dimensions: instruction-following control, agentic tool-use decision control, reasoning-performance steering, reasoning-length control, and trajectory-conditioned safety awareness. Rather than measuring steering success alone, the benchmark jointly evaluates target controllability and capability preservation, including instruction following, reasoning performance, and downstream agent utility and safety. It further examines whether steering effects remain specific and robust across prompt variations, domains, and model architectures. Our evaluation shows that existing steering methods remain unreliable across realistic control objectives: they often fail to consistently induce the intended behavior and can substantially degrade the model's underlying task performance. These results suggest that strong steering effects in controlled settings do not necessarily translate into reliable control in more realistic evaluations. By providing a unified, model-agnostic steering evaluation framework, RealControlBench aims to support the development of steering methods that are not only effective but also precise, robust, and useful in realistic AI deployments.

Abstract

LLM Evaluation

Explore similar work

RealControlBench: A Benchmark for Interpretability-Informed Fine-Grained Control of LLMs and Agents

FaithSteer-BENCH: A Deployment-Aligned Stress-Testing Benchmark for Inference-Time Steering

Measuring Activation Control in Large Language Models