Evaluate the question of whether LLM's can employ decision heuristics, or if they can…
- User
Evaluate the question of whether LLM's can employ decision heuristics, or if they can only generate text that is consistent with those heuristics. Or in other words, does requiring the use of a heuristic by an LLM only bias the generated tokens to produce text that appears to have used the heuristic without altering the model’s underlying function?
- Assistant
Large Language Models (LLMs) like GPT-3.5 have demonstrated behaviors that align with human decision heuristics, such as anchoring, representativeness, availability, framing effects, and the endowment effect. These findings suggest that LLMs can exhibit heuristic-driven responses, not merely generate text that appears heuristic-based.
A study by Suri et al. (2024) investigated whether ChatGPT displays decision heuristics similar to humans. The researchers conducted four studies where ChatGPT was influenced by random anchors in making estimates (anchoring heuristic), judged the likelihood of two events occurring together as higher than either event alone (representativeness and availability heuristics), found an item more efficacious when its features were presented positively rather than negatively (framing effect), and valued an owned item more than a newly found identical item (endowment effect). Human participants showed similar effects in each study. The authors concluded that the presence of these heuristics in an LLM, which lacks cognitive and affective processes, invites consideration of the possibility that language may play a role in generating these effects in humans[1].
Another study explored how LLMs learn task-specific heuristics from in-context examples. The researchers found that LLMs can learn and apply heuristics from demonstrations, leading to improved performance in tasks like document-level event argument extraction. This suggests that LLMs can internalize and utilize heuristics beyond merely generating text that appears heuristic-driven[2].
These studies indicate that LLMs can employ decision heuristics, not just produce text consistent with them. When prompted to use a specific heuristic, an LLM's responses are influenced by that heuristic, reflecting an alteration in the model's function rather than a superficial bias in token generation.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.