Source-linked AI summary
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, Guozhou Zheng, Huajun Chen, Shumin Deng
TL;DR
LLMs deployed in socially sensitive domains can exhibit unpredictable behaviors that make controllability an important research problem. SteerEval introduces a hierarchical benchmark spanning behavioral domains and specification levels, finding that steering performance systematically degrades as objectives become deeper and more tightly specified.
Problem
LLM steering benchmarks are often narrow, while heterogeneous concepts and formats limit systematic cross-method evaluation of controllability.
Method
SteerEval organizes language features, sentiment, and personality into L1–L3 specification levels and evaluates steering with synthetic, manually validated preference data.
Results
Steering performance degrades systematically and predictably as control objectives become deeper and more tightly specified.
Takeaways & Limitations
SteerEval makes boundaries and failure modes of existing steering methods explicit for developing more reliable and interpretable behavioral control.
Takeaways & Limitations
The benchmark does not cover multi-turn dialogue, tool use, long-context interaction, or safety-critical domains.
Abstract
from arXiv · showhide
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.
1 Introduction
LLMs deployed in socially sensitive settings can behave unpredictably, creating risks that make reliable behavioral control essential. SteerEval addresses this need with a hierarchical benchmark linking behavioral intent to increasingly concrete expression constraints and evaluating steering across domains and levels.
- Motivation: LLMs used in education, healthcare, and decision support can show misaligned intent, unintended sentiment shifts, and inconsistent personality expression.These behaviors are presented as tangible risks in real-world settings.
- Benchmark: SteerEval is a hierarchical benchmark for systematically evaluating LLM steerability.
- Contribution: SteerEval uses a principled and interpretable hierarchy to connect behavioral goals with concrete textual outputs for systematic steering evaluation.
- Benchmark: The benchmark covers language features, sentiment, and personality, with L1 specifying what to express, L2 how to express it, and L3 how to instantiate it.The hierarchy links high-level behavioral intent to concrete textual outputs.
- Findings: Steering performance often degrades as behavioral constraints become more fine-grained and precise, although some methods remain reliable at coarse levels.
2 Preliminary
The paper formulates steering as inference-time intervention that should improve expression of a target concept while preserving response quality. It positions SteerEval as a broader, structured alternative to benchmarks that are narrow, heterogeneous, or not concept-targeted.
- Steering Task: Without steering, the model generates ˆy = M(x); a steering method conditions on g to construct an inference-time intervention I_g.
- Steering Methods: Prompt-based steering prepends a concept prompt p_g, whereas activation-based steering modifies intermediate activations with a concept-specific vector.
- Evaluation: Steering is evaluated by concept expression alongside instruction following and fluency, using the open-source EASYEDIT framework.
- Existing Benchmarks: Prior benchmarks often target limited behaviors or tasks, while heterogeneous concepts and formats make cross-method comparison difficult.
- SteerEval: SteerEval addresses these gaps with multiple behavioral domains, three granularity levels, concept-targeted preference data, and concept-aligned evaluation sets.
3 Hierarchical Steering Benchmark
SteerEval tests behavioral steering across three domains and progressively stricter L1–L3 specifications, moving from intent to strategy to verifiable surface evidence. Its benchmark combines automated synthesis with paired data, quality controls, and manual validation.
- Design Principles: Steering targets are tested under progressively stricter granularity constraints to probe where control methods remain robust or break down.
- Design Principles: The benchmark spans personality, sentiment, and language features, with each concept organized into three increasingly specific levels.
- Granularity Hierarchy: L1 specifies what to express, L2 specifies how to express it, and L3 turns the strategy into atomic, verifiable surface constraints.
- Granularity Hierarchy: L3 provides unambiguous realization evidence but may interfere with instruction following, making it the strictest setting.
- Data Synthesis: The synthesis pipeline generates hierarchical concepts, rewrites questions to reduce concept cues, and produces minimally edited matching and nonmatching answer pairs.
- Quality Assurance: Data quality combines automated format and integrity validation with calibrated, dual independent manual review and consensus resolution.
- Dataset Statistics: The core benchmark contains 7,560 paired samples across 3 domains, 3 levels, and 8 concepts per level, with 70 training, 30 test, and 5 validation samples per concept.
4 Experiments
The experiments compare prompt-based and activation-based steering across models, domains, granularity levels, and evaluation metrics. Prompting is strongest overall, while activation-based control becomes substantially less reliable at finer-grained levels.
- Experiment Settings: The evaluation covers Gemma-2-9B-Instruct, Qwen-2.5-7B-Instruct, and Llama-3.1-8B-Instruct using prompt-based, activation-based, and unsteered baselines.Prompt baselines include 0-shot and 3-shot prompting; activation baselines include PCA, DiffMean, and RePS.
- Evaluation: The benchmark reports Concept, Instruction, Fluency, and harmonic-mean scores on open-ended generations.The harmonic mean combines the three component scores and downweights low performance in any single dimension.
- Overall Comparison: Prompt (0/3-shot) reaches HM=3.10/3.12 on Gemma-2-9B-Instruct, exceeding PCA 1.11, DiffMean 1.98, RePS 2.56, and Vanilla 0.81.The passage reports consistent conclusions on Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct, with few-shot prompting providing additional improvement.
- Overall Comparison: RePS is consistently stronger than PCA and DiffMean among activation-based methods, but still trails prompting overall.RePS directly trains a steering vector from data, whereas PCA and DiffMean are training-free baselines.
- Level-wise Analysis: Activation-based HM on Gemma-2-9B-Instruct drops from 1.67/2.76/2.94 at L1 to 0.05/0.07/1.72 at L3 for PCA/DiffMean/RePS, while prompting remains around 3.0.Activation methods can match or outperform prompting at L1 but fall behind substantially at L2 and L3.
- Domain-wise Analysis: RePS has level-averaged HM of approximately 2.43 for personality, 2.37 for sentiment, and 2.25 for language features across the three models.These results indicate domain-dependent transfer for activation-based steering, unlike the more stable prompting results.
5 Analysis
The analyses examine few-shot demonstrations, steering-factor scaling, and case studies across the three granularity levels. Coarse control is generally easier, whereas fine-grained constraints expose trade-offs and persistent failures.
- Few-shot Analysis: A few demonstrations often provide most prompting gains for L1/L2 targets before performance saturates.The analysis attributes this pattern to helping the model infer the task and disambiguate underspecified instructions.
- Few-shot Analysis: For L3 targets, additional demonstrations are typically less helpful and can hurt by introducing idiosyncratic surface cues.The passage presents this as a plausible explanation for weaker gains at the finest level.
- Scaling with the Steering Strength: Increasing activation strength improves Concept Score but can reduce Instruction following and Fluency, producing a Harmonic Mean peak at moderate strengths.The pattern reflects a trade-off between enforcing the target concept and retaining general response quality.
- Scaling with the Steering Strength: Stronger activation steering can outperform prompting for well-calibrated L1 targets, but L2/L3 gains are smaller and less consistent.The evaluated activation methods do not reliably deliver fine-grained control under stronger specification constraints.
- Case Study: L1 steering usually enhances concept expression while largely preserving instruction following and fluency across prompts.The case study characterizes coarse-grained concept guidance as typically easy to steer without harming general quality.
- Case Study: L2 creates a recurring tension: methods may preserve instruction following and fluency while missing the specified concept, or realize the style while sacrificing those capabilities.At this level, the target constrains the manner of expression rather than only the broad concept.
- Case Study: L3 remains difficult even when general capabilities degrade, because concept scores often stay low under the token-level constraint.The case study therefore distinguishes atomic constraint satisfaction from easier L1–L2 steering.
6 Related Work
Prior steering work includes prompt-based and activation-based methods, but existing evaluations often cover limited behaviors or lack explicit hierarchical behavioral structure. SteerEval is motivated by the open question of systematic, predictable control across concept levels.
- Steering Methods: Prompt-based steering guides generation with designed instructions or examples, while activation-based steering modifies hidden activations using learned concept directions.These are presented as two common steering families in the related work discussion.
- Evaluation Gaps: Existing methods are often evaluated on limited behaviors or small task sets, including sentiment, safety, and personas.The cited prior evaluations do not establish systematic control across broad behavioral domains and granularity levels.
- Evaluation Gaps: AXBENCH improves cross-method comparability but uses SAE-derived concepts without domain or granularity structure or concept-targeted preference pairs.Its evaluation prompts are sampled from Alpaca-Eval rather than tailored to specific concepts.
- Open Problem: Whether model behavior can be controlled systematically, predictably, and hierarchically remains an open question requiring evaluation across concept levels.The paper frames this gap as the need for a hierarchical steering benchmark.
7 Conclusion
The paper introduces SteerEval as a hierarchical benchmark for evaluating behavioral steering across domains and concept granularities. Its results show predictable degradation as control objectives become deeper and more tightly specified, making current limits explicit.
- Contribution: SteerEval evaluates LLM steering across behavioral domains and concept-granularity levels using high-quality synthetic preference data.The benchmark is designed to support systematic evaluation of controllability.
- Conclusion: Steering performance degrades systematically and predictably as control objectives become deeper and more tightly specified.The benchmark reveals boundaries and failure modes of existing steering methods.
- Implications: By making these limits explicit, SteerEval provides a principled foundation for more reliable, robust, and interpretable behavioral control.This conclusion stays within the benchmark’s stated role as a foundation for future approaches.
Limitations
The paper identifies limitations in benchmark coverage, experimental scope, steering-method tuning, and LLM-based evaluation. These constraints particularly affect generalization to realistic settings and interpretation of fine-grained control results.
- Coverage of concepts and domains: The benchmark covers a limited set of concepts and domains, excluding multi-turn dialogue, tool use, long-context interaction, and safety-critical domains.The pipeline is described as extensible, but extending coverage to these settings remains future work.
- Experimental setting: The experiments study single-turn prompts and single-concept control, excluding composition, sequential steering, and iterative steering common in real use.
- Method tuning: Steering results depend on extraction choices and coefficient selection, with no claim of optimal tuning for every concept, especially at L2/L3.The authors sweep steering strengths but do not establish optimal settings across all concepts.
- LLM-as-a-judge: LLM-based judges may be biased, prompting-sensitive, and unreliable for fine-grained compliance, so scores are approximate rather than definitive ground truth.
Ethics Statement
The benchmark characterizes controllability boundaries across domains and granularity levels, while its extensible pipeline introduces potential misuse risk. The authors recommend deployment monitoring and capability-retention checks and report that the benchmark contains no personal data.
- The extensible benchmark pipeline implies misuse risk, motivating safety monitoring and capability-retention checks during deployment.
- The benchmark characterizes controllability boundaries across behavioral domains and granularity levels.
- The benchmark does not collect or include personal data, and the authors do not anticipate significant ethical or societal impacts.
A Dataset Case
The dataset case illustrates how SteerEval entries are structured and annotated. Figure 7 presents the fields used to define concepts, pose probing questions, and provide contrastive steering responses.
- Figure 7 shows the field specifications of a representative data entry.
- Each entry includes domain and concept definitions that specify the behavioral target.
- The entry also contains a probing question and contrastive responses used for model steering.
B Detailed Experimental Setup
The experiments apply steering at fixed mid-to-late layers for three instruction-tuned models and tune steering factors per concept on validation data. The remaining hyperparameters follow prior frameworks, with experiments run on three NVIDIA A800 GPUs for one week.
- Layer selection: Steering is applied at a single mid-to-late layer for each evaluated model.The selected layers are the 20th, 14th, and 12th for Gemma-2-9B-Instruct, Qwen-2.5-7B-Instruct, and Llama-3.1-8B-Instruct, respectively.
- Steering-factor search: The study searches for an optimal steering factor for each concept on the validation set when applying PCA, DiffMean, and RePS.Detailed steering-factor values are reported in the referenced tables.
- Compute and hyperparameters: All experiments use three NVIDIA A800 GPUs over one week, with other hyperparameters kept consistent with AxBench and RePS.
C Detailed Experiment Results
The benchmark reports concept, instruction, fluency, and harmonic-mean scores on a 0–4 scale, while validating an LLM-based judge against human annotations. Agreement is strongest for fine-grained L3 concepts, whereas coarser levels show more interpretational variability.
- CS, IS, FS, and HM are reported for all domains and granularity levels on a 0–4 scale.
- The study supplements LLM judging with manual concept-score evaluation of 432 randomly sampled generations across baselines.
- Agreement between the LLM-based judge and human annotations is highest at L3, where concepts are more defined and less ambiguous.
- L1 and L2 show slightly lower judge–human agreement because their coarser concepts permit greater interpretational variability.
- A consistent LLM-based evaluator supports fair comparisons across steering methods and granularity levels.
D.2 Synthetic Benchmark Bias and Data Reliability
The benchmark combines human review, sampled data-quality checks, and structured prompts to constrain synthetic concept and question generation. Its reliability evidence includes a 94.1% verification pass rate and Cohen’s κ of 0.82.
- Data validation: Concepts generated by LLMs are independently reviewed by multiple annotators and retained only with unanimous agreement.
- Data validation: 94.1% of a randomly sampled 20% dataset subset passed binary human verification.
- Benchmark structure: The data-entry schema separates hierarchical domain and concept fields from contrastive matching and not_matching responses used for steering.
- Data validation: Cohen’s κ = 0.82 indicates strong inter-annotator agreement and high data validity.
- Prompt constraints: Generation prompts impose broad scope, clear boundaries, abstract terminology, and relevance to model steering while excluding specific concepts.
- Hierarchical synthesis: The synthesis instructions define L1 as coarse, domain-specific categories that specify what orientation to take without prescribing implementation.