Mind the Performance Gap: Capability-Behavior Trade-offs in Feature Steering

Feature steering has emerged as a promising approach for controlling LLM behavior through direct manipulation of internal representations, offering advantages over prompt engineering. However, its practical effectiveness in real-world applications remains poorly understood, particularly regarding potential trade-offs with output quality. We show that feature steering methods substantially degrade model performance even when successfully controlling target behaviors, a critical trade-off. Specifically, we evaluate Goodfire's Auto Steer against prompt engineering baselines across 14 steering queries (covering innocuous and safety-relevant behaviors) on 171 Massive Multitask Language Understanding (MMLU) questions using Llama-8B and Llama-70B, measuring accuracy, coherence, and behavioral control. Our findings show that Auto Steer successfully modifies target behaviors (achieving scores of 3.33 vs. 2.98 for prompting on Llama-8B and 3.57 vs. 3.10 on Llama-70B), but causes dramatic performance degradation: accuracy on the MMLU questions drops from 66% to 46% on Llama-8B and 87% to 73% on Llama-70B, with coherence falling from 4.62 to 2.24 and 4.94 to 3.89 respectively. Simple prompting achieves the best overall balance. These findings highlight limitations of current feature steering methods for practical deployment where task performance cannot be sacrificed. More broadly, our work demonstrates that mechanistic control methods face fundamental capability-behavior trade-offs that must be empirically characterized before deployment.

Do Large Language Models Show Biases in Causal Learning?

Causal learning is the cognitive process of developing the capability of making causal inferences based on available information, often guided by normative principles. This process is prone to errors and biases, such as the illusion of causality, in which people perceive a causal relationship between two variables despite lacking supporting evidence. This cognitive bias has been proposed to underlie many societal problems, including social prejudice, stereotype formation, misinformation, and superstitious thinking. In this research, we investigate whether large language models (LLMs) develop causal illusions, both in real-world and controlled laboratory contexts of causal learning and inference. To this end, we built a dataset of over 2K samples including purely correlational cases, situations with null contingency, and cases where temporal information excludes the possibility of causality by placing the potential effect before the cause. We then prompted the models to make statements or answer causal questions to evaluate their tendencies to infer causation erroneously in these structured settings. Our findings show a strong presence of causal illusion bias in LLMs. Specifically, in open-ended generation tasks involving spurious correlations, the models displayed bias at levels comparable to, or even lower than, those observed in similar studies on human subjects. However, when faced with null-contingency scenarios or temporal cues that negate causal relationships, where it was required to respond on a 0-100 scale, the models exhibited significantly higher bias. These findings suggest that the models have not uniformly, consistently, or reliably internalized the normative principles essential for accurate causal learning.

Are UFOs Driving Innovation? The Illusion of Causality in Large Language Models

Illusions of causality occur when people develop the belief that there is a causal connection between two variables with no supporting evidence. This cognitive bias has been proposed to underlie many societal problems including social prejudice, stereotype formation, misinformation and superstitious thinking. In this research we investigate whether large language models develop the illusion of causality in real-world settings. We evaluated and compared news headlines generated by GPT-4o-Mini, Claude-3.5-Sonnet, and Gemini-1.5-Pro to determine whether the models incorrectly framed correlations as causal relationships. In order to also measure sycophantic behavior, which occurs when a model aligns with a user's beliefs in order to look favorable even if it is not objectively correct, we additionally incorporated the bias into the prompts, observing if this manipulation increases the likelihood of the models exhibiting the illusion of causality. We found that Claude-3.5-Sonnet is the model that presents the lowest degree of causal illusion aligned with experiments on Correlation-to-Causation Exaggeration in human-written press releases. On the other hand, our findings suggest that while mimicry sycophancy increases the likelihood of causal illusions in these models, especially in GPT-4o-Mini, Claude-3.5-Sonnet remains the most robust against this cognitive bias.