Anthropocentric bias in language model evaluation

Evaluating the cognitive capacities of large language models (LLMs) requires overcoming not only anthropomorphic but also anthropocentric biases. This article identifies two types of anthropocentric bias that have been neglected: overlooking how auxiliary factors can impede LLM performance despite competence (“auxiliary oversight”), and dismissing LLM mechanistic strategies that differ from those of humans as not genuinely competent (“mechanistic chauvinism”). Mitigating these biases necessitates an empirically-driven, iterative approach to mapping cognitive tasks to LLM-specific capacities and mechanisms, which can be done by supplementing carefully designed behavioral experiments with mechanistic studies.

💡 Research Summary

This paper presents a critical methodological analysis of the evaluation of large language models (LLMs), arguing that assessing their cognitive capacities requires overcoming not only anthropomorphism (attributing human qualities without justification) but also a less-discussed “anthropocentric bias.” Anthropocentric bias is defined as the tendency to evaluate LLMs by human standards unjustly, potentially dismissing genuine competencies that operate differently from human cognition.

The authors begin by framing the issue through the classic competence/performance distinction from cognitive science. They observe a troubling asymmetry in current LLM evaluation: while inferences from good performance to competence are treated cautiously due to fears of anthropomorphism, inferences from bad performance to a lack of competence are often drawn hastily. This one-sided skepticism, they argue, is itself a form of anthropocentric bias.

The core contribution is the identification and detailed analysis of two specific, neglected types of anthropocentric bias:

1. Auxiliary Oversight: This is the tendency to overlook how an LLM’s performance failure on a task might be caused by “auxiliary factors” rather than a lack of the core competence being tested. The paper breaks this down further:

   • Mismatched Auxiliary Task Demands: When experimental conditions for humans and LLMs are not equitable. For example, humans may receive instructions, training, and context for a task (like judging recursively nested grammatical structures), while LLMs are evaluated “zero-shot.” This imposes a heavier auxiliary demand (understanding the task itself) on the LLM, confounding the comparison of core linguistic competence.
   • Test-time Computational Bottlenecks: Limitations on the computational resources an LLM can use during inference can mask its underlying capabilities. A prime example is the difference between asking an LLM for a direct answer versus allowing it to generate a “chain of thought.” Performance can dramatically improve with more “thinking” tokens, as seen in models like OpenAI’s o1 or latent reasoning models that refine internal states iteratively. A failure under constrained compute does not necessarily indicate a lack of competence.

2. Mechanistic Chauvinism: This bias involves dismissing an LLM’s problem-solving strategy as not genuinely competent simply because it differs from the mechanistic strategies employed by humans. For instance, if an LLM solves a reasoning task using statistical pattern matching rather than explicit, rule-based symbolic manipulation, a mechanistically chauvinistic evaluator might deny it has true reasoning ability. The authors argue that competence should be defined functionally—as a system’s computational capacity to meet a normative standard under fair conditions—rather than being tied to human-like mechanistic implementations.

To mitigate these biases, the paper advocates for an empirically-driven, iterative methodology. This involves a continuous loop between carefully designed behavioral experiments (to map what LLMs can do) and mechanistic studies using interpretability tools (to understand how they do it). These two approaches inform and refine each other. Behavioral experiments can reveal capacities that mechanistic studies then seek to explain, while mechanistic insights can guide the design of fairer behavioral tests that control for auxiliary factors.

In conclusion, the paper calls for a shift in perspective: LLMs should be evaluated not as imperfect proxies for human cognition, but as unique computational systems with their own profiles of strengths, weaknesses, and operational mechanisms. Developing a fair and accurate understanding of their capacities requires a methodology specifically tailored to them, one that consciously guards against the subtle pitfalls of anthropocentric reasoning.