Can we make LLM social simulations interpretable and steerable?

The abstract grants that LLM-based simulations are "powerful for understanding human behavior" and then argues they are limited as social science because the models are black boxes. It names two gaps. Interpretability is "the ability to assign clear mechanisms driving observed behavior." Steerability is "the ability to mute or amplify specific theoretically meaningful mechanisms of action to drive specific model behavior." What the excerpt describes is a framework for comparing "alternate methods for looking into the LLM black-box," not a single new technique.

Three method families are compared: (1) prompt-based manipulation, (2) SAE-derived feature steering, and (3) probe-based direction steering. They are applied to two "foundational components of human behaviors": preferences (risk attitudes, altruism) and capabilities (divergent creativity, product innovation). Both are "operationalized using four classic economic and creative tasks implemented as natural-language interactions." The design logic is that a simulated agent's behavior should trace to a named mechanism, and that the mechanism should be adjustable the way a social scientist varies a theoretical parameter. The discussion passage is cut off, but it points to "hybrid approaches, such as multi-probe steering or low-rank" variants as future work.

In the vault's terms, the probe-based arm belongs to the family described in Can high-level concepts replace circuit-level analysis in AI?: a concept treated as a direction in activation space and tested by manipulation. The target here differs, since risk attitude or altruism replaces truthfulness or honesty. The SAE arm has a neighbor in Can we trigger reasoning without explicit chain-of-thought prompts?, where one steered feature overrode a prompt-level instruction. If something similar held for social traits, prompt manipulation would be the shallowest of the three interventions, but this paper's excerpt neither tests nor claims that. Why do large language models explore less effectively than humans? is another case of using SAE features to attribute a decision-making tendency to internal mechanisms, which is the kind of attribution the paper's interpretability gap asks for.

The excerpt is silent on results. It gives no models, sample sizes, effect sizes or ranking of the three methods, and it does not say whether steered agents reproduce human behavior better. The introduction concedes that simulations "might not always faithfully match human behavior," and steerability does not by itself address that fidelity problem. What can be said at this strength is that the paper proposes an evaluation design, with the same behavioral dimensions run through three intervention families. Until its results are in view, the design is something to cite as a framing of the problem and not as a finding about which method works.