Language models are now used to play simulated users who talk to dialogue systems, for example to stress-test a restaurant-booking assistant against many different customers. Researchers want these simulated users to have personalities, such as a careless user or a nervous one, and a promising way to set personality is activation steering: adding a fixed direction vector to the model's internal hidden states during generation, so that the trait is dialed in numerically instead of requested in a prompt. Steering has a hidden cost, though. It pushes every token the same way, including the tokens where the user states what they actually want. A nervous user who forgets their party size, or a careless user who books the wrong city, is no longer testing the system. The simulator has simply stopped playing its assigned task. Recent work on steered user simulators reports trait expression and judged satisfaction, but it does not measure whether the steered user stays faithful to its goal.
We propose using the explicit task state that task-oriented dialogue already has, namely the user's goal frame and the running record of which slots have been filled, for two jobs. First, it is an objective instrument for measuring goal drift caused by steering. Second, it is a gate that decides turn by turn, or span by span, when to apply the persona vector at full strength and when to release it. The central claim is that persona and goal compete for the same activations. If that is right, a state-gated simulator should reach a strictly better tradeoff between trait expression and goal fidelity than any constant steering strength. We also test a second claim: a simulator first fine-tuned on in-domain user turns absorbs steering with less damage to goal fidelity and fluency than a general instruction-tuned model.
Core Claim and What Is New
The closest prior work is Activation-Steered Personas in Task-Oriented LLM Agent Simulations. It applies fixed Big Five trait vectors to simulated users on the Schema-Guided Dialogue Restaurants and Movies domains, and it adapts a system-side steering coefficient from the raw dialogue text. It reports that trait expression depends on domain and steering strength, and that adding system-side steering to steered users improves pairwise-judged truthfulness by +8.3 points and satisfaction by +2.7 points. It does not condition user-side steering on any structured state, and it does not measure whether steered users keep their goals. PERSONA also adjusts traits turn by turn (PERSONA-FLOW), predicting changes from raw text, and it optimizes trait adherence rather than task fidelity. Programming Refusal with Conditional Activation Steering (CAST) gates a steering vector on a similarity test against an internal condition vector. On Hermes 2 Pro, ungated steering pushed harmless-prompt refusal to 91.8%, while CAST reduced it to 2.4%. This is the clearest evidence in the file that when to steer matters as much as what to steer. Our gate differs from CAST's: it reads an external, inspectable task state instead of a latent similarity score, and the thing it protects is task content rather than refusal behavior. Classic AI as Scaffolding for LLM Social Agents motivates the design. Its EpisodeSim architecture argues that persistent explicit state (roles, obligations, commitments) is what keeps social simulation coherent, but it controls behavior only through prompts. We connect that kind of state to an activation-level control.
Concretely, at layer $l$ and token position $t$ we generate with
$$h_{l,t}' = h_{l,t} + g(s_t, t)\,\alpha\, v,$$
where $h_{l,t}$ is the hidden state, $v$ is a trait vector extracted by contrastive activation addition (the mean difference of hidden states between high-trait and low-trait examples, as in Steering Llama 2 via Contrastive Activation Addition), $\alpha$ is the steering strength, $s_t$ is the current semantic state (the goal frame plus the slots filled so far), and $g \in [0,1]$ is the gate. Constant steering is the special case $g \equiv 1$.
Research Program
We build three gates of increasing granularity. The turn-level gate uses the dialogue act the user is about to perform, which the state determines (for example, informing a required slot, confirming a booking, expressing an open preference, or chit-chat). It applies full steering on preference and social turns and reduced steering on inform and confirm turns. The span-level gate keeps steering on throughout, but sets $g$ to a small value while the model is emitting a value that matches a slot in the goal frame. The match is detected by string-matching against the schema's candidate values as tokens are produced. The learned gate fits $g$ as a small function of state features. It is tuned on a development split to maximize trait detectability subject to a goal-fidelity floor.
The state comes from two sources. In the oracle condition we use Schema-Guided Dialogue's gold user goals and gold dialogue state annotations (Towards Scalable Multi-Domain Conversational Agents). In the realistic condition the user simulator knows its own goal by construction, so only the "slots filled so far" part needs predicting, and we predict it with an LLM state tracker. This addresses the oracle-dependence concern that Baked-in State Probing raises for state abstractions. For reference, the original SGD baseline tracker reached joint goal accuracy of only 0.356 on seen and 0.254 on unseen services. This is why we report the predicted-state condition separately instead of assuming clean state.
Domain fine-tuning enters as a controlled backbone factor. Each base model is used in two versions: as released, and after LoRA fine-tuning on SGD user turns conditioned on the goal frame. We extract steering vectors separately on each backbone, and we also test vectors transferred from base to fine-tuned. The question is narrow and empirical. Behavioural Asymmetry Across Activation Interventions for Big Five Personality Control in LLMs found that negative-pole steering roughly doubled perplexity on Llama-3.1-8B. BehaviorBench found that domain-adapted behavioral models match population distributions better than general models. Neither tells us whether a domain-adapted simulator is more or less robust to steering, and our design measures that directly.
Baselines
The naive reference point is the unsteered simulator, which gives the goal-fidelity ceiling and the trait-expression floor. The second reference is prompt-only persona conditioning, with the persona description and goal frame both in the system prompt. Behavioural Asymmetry reports a system-prompting ceiling of about 72% trait expression on BIG5-CHAT, so prompting is a serious competitor. It is run with the same models, decoding settings and goals, with prompt wording tuned on the same development split and the same number of trials as our gate.
The strongest existing method is constant user-side activation steering as in Activation-Steered Personas, swept over the full range of $\alpha$. The gated methods must beat this whole curve, not one setting of it. The comparison is fair because gated and constant runs use identical vectors, layers (layer 14 for Llama-3.1-8B, as Behavioural Asymmetry found best; the analogous layer chosen on development data for Qwen2.5-7B) and goals. We also report each method's average effective steering strength, defined as $\alpha$ times the mean of $g$, so a gate cannot "win" just by steering less. For the system-side downstream metrics, the reported number to beat is the +8.3 point truthfulness and +2.7 point satisfaction gain of their full user-plus-system steering over user-only steering. The file has no reported number for their task metrics ("slight" improvement only), and no reported result found on Lacuna for goal fidelity of steered users. That gap is part of what this work fills.
The key ablation isolates whether the information in the state matters. A shuffled gate applies the same number of reduced-steering turns or spans at random positions. A latent gate in the style of CAST switches steering off when the hidden state is similar to a condition vector built from slot-informing turns, which needs no external state. Beating constant steering is a success only if the state-gated method also beats the shuffled gate. Beating the latent gate is what justifies maintaining explicit state at all.
Why not just…?
Why not just lower the steering strength? This is the most serious threat. If goal drift and trait expression both scale smoothly with $\alpha$, a modest constant strength might sit on the same frontier as any gate. We take this seriously by comparing whole Pareto curves at matched effective steering strength. If the constant-$\alpha$ curve is not dominated, the central claim is false and we will report it as such.
Why not just put the persona and goal in the prompt? Prompting keeps goals visible, but prior work reports that prompt-conditioned traits are hard to regulate and verify across turns (Activation-Steered Personas), and that prompting plateaus around 72% expression (Behavioural Asymmetry). Our prompt baseline will show where it lands on the same two axes. If it matches gated steering, the activation-level machinery is unnecessary for this use.
Why not just gate on the model's own internal state, as CAST does? CAST shows latent gating works for refusal, and it avoids running a tracker. Our view is that slot commitment is a fine-grained, value-specific event that a single similarity threshold may blur. That is a hypothesis, and the latent-gate ablation settles it directly.
Why not just fine-tune a separate simulator per trait? PERSONA reports steering at 9.60 versus 9.61 for SFT on PersonalityBench, so fine-tuning is not clearly better at trait expression. It also needs persona-labeled task dialogues that SGD lacks, and it cannot change intensity mid-dialogue. As a check, we will still train one SFT simulator for the two most steerable traits on synthetic persona-rewritten SGD dialogues and place it on the frontier.
Proposed Experiments
We run on the SGD Restaurants and Movies domains, matching the seed setting, and add one unseen-service domain to test whether the span gate generalizes through the schema. Models are Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, each in base and SGD-LoRA versions. Trait vectors are built for the traits Activation-Steered Personas found observable (careless, outgoing, compassionate, nervous) and for the hard ones (consistent, solitary), using BIG5-CHAT pairs. Each simulated user talks to a fixed, unsteered system agent.
Goal fidelity is the fraction of goal-frame slot values the user states correctly, scored by exact match against the schema. Unauthorized slot changes are counted separately from omissions, and dialogue-level goal completion is reported alongside. Trait expression is measured with a classifier trained on BIG5-CHAT and with pairwise LLM judgments following the seed protocol. Fluency is measured as perplexity under the unsteered model.
The decisive checks are as follows. Does state gating dominate the constant-$\alpha$ Pareto curve on goal fidelity versus trait expression at matched effective strength? Does it beat the shuffled and latent gates? Does the advantage survive when the predicted state replaces the gold state? Does the fine-tuned backbone shift the whole frontier outward, or does fine-tuning instead make vectors less effective because they must be re-extracted? We expect hard-to-observe traits to show little gain, since there is little expression to protect. A gain concentrated on expressive traits would support the competition account. Uniform gains, or none at all, would point instead to gating simply acting as a lower average $\alpha$.
Workshop authors to connect with
- Activation-Steered Personas in Task-Oriented LLM Agent Simulations: Maryam Shoaeinaeini, Bernardo Rodrigues, Adib Mosharrof, A.B. Siddique, Brent Harrison
- BehaviorBench: Benchmarking Foundation Models for Behavioral Science Tasks: Jin Huang, Yutong Xie, Wanli Song, Xingjian Zhang, Walter M. Yuan, Matthew O. Jackson, Qiaozhu Mei
- Classic AI as Scaffolding for LLM Social Agents: Anatole Gershman
References
- Activation-Steered Personas in Task-Oriented LLM Agent Simulations
- PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra
- Programming Refusal with Conditional Activation Steering
- Classic AI as Scaffolding for LLM Social Agents
- Steering Llama 2 via Contrastive Activation Addition
- Towards Scalable Multi-Domain Conversational Agents: The Schema-Guided Dialogue Dataset
- Baked-in State Probing
- Behavioural Asymmetry Across Activation Interventions for Big Five Personality Control in LLMs
- BehaviorBench: Benchmarking Foundation Models for Behavioral Science Tasks