Interpreting and Steering LLM Agents for Social Simulations
Jiayue Gaveal Fan, Arul Murugan, Shreyas Krishnan, Abhishek Nagaraj
cs.LG, cs.AI, cs.CL
2026-09-15
On Llama-3.3-70B, probes place lottery switch points from ~30-200 tokens with ~2-token error; strong prompts barely move risk but still beat SAE on creativity.
Social scientists already run LLMs as cheap experimental subjects: lotteries, ultimatum games, brainstorming, dozens of agents in a single sweep. The bottleneck is the black box. Induction needs a mechanism you can name. Deduction needs a mechanism you can turn up or down. Persona prompts change the costume. The model then writes a post-hoc story for its choice, and that story need not be the internal representation that produced the action.
The Berkeley DIAL lab ports two tools from mechanistic interpretability into social simulation: sparse autoencoders (SAEs), which unpack hidden activations into human-readable features, and linear probes, which learn a single addable direction for a labeled trait. The baselines are persona, chain-of-thought, and few-shot prompts on the same tasks. The backbone is Llama-3.3-70B-Instruct.
Four tasks, two primitives. Preferences: a lottery (sure 50 tokens versus 50% chance of n, else 0) and an ultimatum game (the other player offers n out of 100; accept and split, reject and both get 0). Capabilities: unusual uses for a brick, and product improvements for a stapler. Forty agents per condition. Risk-neutrality in the lottery is n=100, where expected values match.
The SAE is Goodfire's dictionary on layer 50 of 80, trained on LMSYS Chat. After Claude labels the features and 3.5% harmful ones are dropped, 65,536 dimensions remain. Steering adds or subtracts a scalar α in [-1, 1] on a chosen feature, then decodes back into activation space. The probe is ℓ2-regularized logistic regression at layer 48, the best layer in a cross-validated sweep. Preference labels come from the model's own choices (risky = high risk-taking). Creativity labels come from an independent GPT-5 judge, split at the median. Steering adds a unit-normalized weight vector scaled by λ. Held-out probe accuracy on preference trials is 82%.
The prompting ladder is baseline, persona, zero-shot CoT, 6-shot, and few-shot plus CoT. Lottery and ultimatum also sweep the fraction ρ of demonstrations that take the target action. Creativity adds a high-temperature condition. Open-ended answers are scored blind, with a length-controlled Torrance rubric, by five judges: GPT-5, Claude Sonnet 4.6, Gemini 2.5 Pro, Kimi K2.6, DeepSeek V4 Pro. Mean pairwise Spearman is 0.77.
On the lottery, baseline agents are risk-averse and switch near 125 tokens. A "risk-taking" persona slams the switch point to about 20, in an all-or-nothing way: "barely risky" does nothing, "slightly risky" goes extreme. SAE steering is graded: moderate activations (0.6/0.4) switch near 75, stronger ones (0.7/0.5) near 65-70. A stronger prompting ladder barely moves risk: switch points stay in 97-101 across doses, and CoT makes the agent more cautious, switching near 140. Probes place the switch anywhere from about 30 to 200 tokens, with mean absolute error of about 2 tokens against the target.
On ultimatum, baseline agents start accepting once the offer exceeds about 30 (an even split is 50). An altruistic persona drops the threshold to 10. Few-shot CoT does show a dose response, moving acceptance from about 40 down to 10. Probes slide the threshold between 30 and 60, with saturation at extreme λ.
Creativity reverses the ranking. On the brick task, SAE steering is Δ=-0.23 versus baseline, persona +0.04, high temperature -0.10; zero-shot CoT is +1.52 and few-shot plus CoT +1.48. On the stapler, SAE is +0.45, persona +0.34, persona plus CoT +0.59. A probe can push the GPT-5 score from about 4.0 at target 3 to about 8.2 at target 9, always below the perfect-control diagonal, with a wider gap at high targets. A brick-trained probe still steers stapler, paperclip, and bowl. The same qualitative pattern appears on Qwen-2-7B-Instruct, with the probe at layer 17.
The practical split is clean. Use SAE features when the question is which internal concepts the model is actually using. Use a probe when a predefined trait needs a calibrated dose. Persona prompts are cheap, but they do not give a smooth knob on risk, and they invite treating a generated rationale as a mechanism. Strong prompting still wins on open-ended generation. The point of the interpretability tools here is a controllable internal direction, not a higher creativity score.
This is not a recipe for production agent personas. It is closer to a genetically modified mouse for computational social science: turn one mechanism up or down, then read the psychometric curve.
The authors flag three issues. SAE steering strength is still art; too much produces nonsense, and features are not fully disentangled, so boosting one can drag others along. The study stays at the individual agent. The trait set is small, and more abstract social mechanisms may not sit on a single linear direction.
A few claims sit on thinner ice. Creativity-probe labels come from an LLM judge, and Figure 11's achieved scores are GPT-5; the five-judge panel mainly covers SAE and prompting, so the evaluation loop is not fully cut. Preference-probe labels are the same model's own choices, so the learned direction may be "how this model talks when it picks the gamble," not human risk preference. Prompting is nearly useless on the lottery and already dose-responsive on ultimatum, so the claim that interpretability beats prompting holds sharply for risk and not as a blanket ranking. The main run is tied to Llama-3.3-70B and an external SAE dictionary; a new backbone needs a new probe and a new dictionary.