Tune an AI to act more like people and it starts to look human, but does it think like one?
Does adjusting steered mechanisms make LLM agents match human behavior more closely?
This explores whether deliberately tuning the controls on an LLM agent (conditioning it on profiles and goals, steering it, finetuning it on human data) makes its behavior look more like a person's, and whether 'looks more like' is the same as 'works like'.
This explores whether deliberately tuning the controls on an LLM agent (conditioning it on profiles and goals, steering it, finetuning it on human data) makes its behavior look more like a person's, and whether 'looks more like' is the same as 'works like'. The corpus says yes on the measurable surface, with a catch about what that closeness proves.
The yes case is fairly strong. LLMs finetuned on psychology experiment data predict human decisions better than the theory-driven cognitive models built for that purpose, and their embeddings pick up individual differences Can language models learn to model human decision making?. Conditioning works too: give a user simulator a session-level profile and a turn-level intent, and its synthetic conversations become hard to tell from real ones, according to crowdworkers and discriminator models Can controlled latent variables make LLM user simulators realistic?. Persona-based simulations reproduced 84 of 111 published marketing-experiment effects Can AI personas reliably replicate human experiment results?.
Steering also fixes specific human-versus-machine gaps. Humans stay on their own goals through a conversation, and LLM user simulators often don't. Breaking a simulator's goal into profile, policy, task, requirements and preferences, each with tracked status, reduces that drift Why do LLM user simulators fail to track their own goals?. Steering was only the first stage there. Supervised finetuning and then GRPO were needed to make the goal alignment stick, so a light adjustment gets you partway and the rest takes training.
The closeness has limits. The persona replications track how strong the original evidence was, and they got unreliable on marginal effects, with both false positives and false negatives Can AI personas reliably replicate human experiment results?. Simulated social competence also collapses when agents hold private information. One model playing every side looks fine, and it fails once each agent knows something the others don't Why do LLMs fail when simulating agents with private information?. So tuning helps most in the easy conditions.
The deeper catch is that matching human actions shows prediction, not explanation. Many different reasoning paths produce the same behavior, so a well-tuned agent can match people without thinking like them Can matching human actions prove an LLM simulation explains behavior?. If you want to diagnose or compare interventions, you have to keep the reasoning traces, not just score the actions. Some human-like traits also sit in the training objective rather than in adjustable settings. Conversational agents are passive by design because alignment rewards responding over initiating Why can't conversational AI agents take the initiative?. Much of the fix may be scaffolding rather than model tweaks: reliability tends to come from externalizing memory, skills and protocols into a harness Where does agent reliability actually come from?, and a single model prompted to play several personas can reproduce multi-agent dynamics Can branching prompts replicate what multi-agent systems do?. In short, adjusting the mechanisms reliably makes agents match human outputs more closely, but the corpus gives little evidence that it makes them match human reasoning.
Sources 9 notes
LLMs finetuned on psychology experiment data predict human behavior more accurately than theory-driven models in decision tasks, capture individual differences in their embeddings, and transfer learning across tasks without task-specific design.
RecLLM demonstrates that conditioning an LLM simulator on session-level (user profile) and turn-level (user intent) latent variables produces synthetic conversations measurable as realistic via crowdsource discrimination, discriminator models, and classifier-ensemble distribution matching.
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
The UGST framework breaks user goals into profile, policy, task, requirements, and preferences—each with explicit status tracking. A three-stage method (steering, SFT, GRPO) progressively internalizes goal alignment, reducing the misalignment that corrupts RL training signals.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
Show all 9 sources
Matching human actions proves only prediction, not explanation. Since behavior underdetermines reasoning, simulations used to explain, diagnose, or compare interventions must preserve the reasoning traces behind actions, not just reproduce the actions themselves.
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Research shows single LLMs using dynamic persona simulation achieve multi-agent cognitive synergy without multiple model instances. Solo Performance Prompting validates that structured prompting techniques map directly to multi-agent debate architectures, enabling equivalent outcomes through structural equivalence.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Goal Alignment in LLM-Based User Simulators for Conversational AI
- Is this the real life? Is this just fantasy? The Misleading Success of Simulating Social Interactions With LLMs
- LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents
- Do Role-Playing Agents Practice What They Preach? Belief-Behavior Consistency in LLM-Based Simulations of Human Trust
- From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration
- A Looming Replication Crisis in Evaluating Behavior in Language Models? Evidence and Solutions
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation