CoBRA: Programming Cognitive Bias in Social Agents
Paper URL: arXiv:2509.13588
arXiv subjects: cs.AI (Artificial Intelligence), cs.CE (Computational Engineering, Finance, and Science), cs.CY (Computers and Society)
Giving an AI agent a persona is not the same as specifying how it will behave. A prompt such as “act like an economist” relies on a model’s shifting associations with that role. It does not say how the agent should respond when an outcome is framed as lives saved rather than lives lost, nor does it give a researcher a way to measure whether the behavior is consistent. CoBRA addresses this problem by turning vague behavioral descriptions into observable and adjustable targets.
Liu, Shang, and Jin introduce CoBRA, a toolkit for LLM-based social simulation. It has two parts. The Cognitive Bias Index (CBI) measures an agent’s response pattern in adapted classic social-science experiments. The Behavioral Regulation Engine then adjusts the agent until it reaches a specified score. The system can intervene through numerical prompts, changes to hidden-state activations, or model-parameter updates. Rather than asking for an agent with a loosely defined disposition, a researcher can request a measurable level of an observed tendency within the toolkit’s testbed.
The paper evaluates four tendencies: authority effect, bandwagon effect, confirmation bias, and framing effect. For each, the authors use two adapted experimental paradigms and compute a CBI from response probabilities. In the framing task, for example, an agent receives equivalent outcomes stated positively and negatively. Its tendency to prefer the positive framing produces a score from 0 to 4. A lower score means less demonstrated framing preference in that experiment. It does not mean that the agent is generally unbiased, rational, or expert.
Across several open and API-based models, CoBRA produced smoother and more reproducible control curves than qualitative persona prompts or informal descriptions such as low, medium, and high bias. The authors also report that some calibrated framing controls preserve their relative ordering when transferred between related paradigms. An open-ended emotional-contagion demonstration shows an ordered response pattern after calibration, whereas qualitative specifications yield overlapping effects. Together, these results make a credible case that CoBRA can regulate selected response tendencies within its evaluated settings.
The paper’s most useful conceptual distinction is between reproducibility and validity. CoBRA may reliably reproduce a CBI score, but that alone does not prove that the score captures the full psychological construct it is named after. Classic experiments were validated with people, not automatically with language models responding to adapted multiple-choice items. Likewise, transfer between closely related paradigms shows rank preservation within a benchmark family, not generalization to real conversation, work, or social life.
The evidence is therefore strongest for a behavioral control layer, not a theory of machine cognition or a validated social simulator. CoBRA controls observable context sensitivity; it does not show that an agent has human-like cognitive mechanisms. The paper does not evaluate human participants, workplace use, long-running multi-turn conversations, or usability for social-science researchers. Its synthetic emotional-contagion example extends the evidence beyond questionnaires, but it still uses generated posts and automated sentiment analysis.
These limits do not weaken the core systems contribution. They clarify its promise. CoBRA offers a more disciplined way to specify agent behavior than relying on persona prompts, and it makes that specification testable. For HCI, the next questions concern how people understand such controls, when they trust them, and who bears responsibility when an agent is intentionally designed to exhibit a behavioral tendency. Future work should test multiple interacting tendencies, sustained dialogue, human-facing deployments, and the ecological validity of the measures. The central design challenge is not merely controlling agent behavior, but making that control intelligible, accountable, and appropriate in the settings where agents are used.