Beyond Prompt: Fine-grained Simulation of Cognitively Impaired Standardized Patients via Stochastic Steering¶

Conference: ACL 2026 arXiv: 2604.12210 Code: N/A Area: Medical Imaging Keywords: standardized patient simulation, cognitive impairment, steering vector, stochastic modulation, clinical training

TL;DR¶

This paper proposes StsPatient, which extracts domain-specific steering vectors from contrastive instruction/response pairs and applies a Stochastic Token Modulation (STM) mechanism to control injection probability, enabling simulation of standardized patients across different cognitive impairment domains and severity levels. Compared to prompt engineering methods, StsPatient achieves an average improvement of 11.23% in clinical authenticity and surpasses the best baseline by 18.54% in severity controllability.

Background & Motivation¶

Background: Patients with cognitive impairment (e.g., Alzheimer's disease, mild cognitive impairment) exhibit deficits of varying degrees across multiple cognitive domains such as memory and attention, which significantly affect their language patterns. Clinical practitioners require specialized training to communicate with such patients, and traditional approaches rely on human actors to portray standardized patients (SPs).

Limitations of Prior Work: (1) Cognitive impairment is highly heterogeneous—the same diagnosis may manifest as deficits in different domains (attention, memory, executive function, etc.) at varying severity levels ranging from mild to severe, making full coverage by human actors costly and impractical. (2) Existing LLM-based SP methods primarily rely on prompt engineering; however, prompts are inherently discrete and coarse-grained, making it impossible to precisely control the degree of deficit in specific cognitive domains. (3) Traditional steering vector methods control intensity via a scaling coefficient \(\alpha\), but the relationship between \(\alpha\) and behavioral output is highly nonlinear and unstable.

Key Challenge: Achieving fine-grained, stable, and controllable simulation across the combinatorial space of multiple cognitive domains × multiple severity levels, given that prompts are too coarse and traditional steering vectors are too unstable.

Goal: To design a framework that can (1) extract behavioral modulation signals specific to different cognitive domains and (2) stably control deficit manifestation along a continuous severity spectrum.

Key Insight: Inspired by the probabilistic nature of synaptic transmission in biological neuroscience—where synaptic strength is regulated not by signal amplitude but by the probability of neurotransmitter release—this work analogously proposes to modulate not the magnitude of the steering vector, but the probability of its application at each token.

Core Idea: Fix the magnitude of the steering vector at a level sufficient to trigger observable deficits, and use Bernoulli sampling to determine whether the steering vector is injected at each token; the probability \(s\) directly maps to severity level.

Method¶

Overall Architecture¶

The framework consists of two stages: (1) Domain-specific steering vector extraction—a LLM synthesizes a contrastive dataset (impaired vs. healthy), and the steering vector is computed as the mean difference in hidden states; (2) Stochastic Token Modulation (STM) at inference—a Bernoulli distribution with probability \(s\) determines whether the steering vector is injected at each token generation step.

Key Designs¶

Domain-Specific Steering Vector Extraction (Dual-Channel Contrastive):
- Function: Captures the representational direction of linguistic deficits specific to a given cognitive domain.
- Mechanism: Two types of contrastive subsets are constructed: a Prompt contrastive subset (system instruction pairs such as "play a patient with memory impairment" vs. "play a healthy person") and a Response contrastive subset (impaired vs. healthy response pairs for the same clinical question). The steering vector is computed as \(\mathbf{v}_d = \text{mean}(\mathbf{h}^+ - \mathbf{h}^-)\) from the hidden state differences of contrastive pairs, and the layer maximizing the distance between positive and negative sample embedding centroids is automatically selected.
- Design Motivation: The dual-channel design captures deficit characteristics at both the "instruction intent" and "behavioral representation" levels simultaneously, providing more comprehensive coverage than a single source.
Stochastic Token Modulation (STM):
- Function: Enables stable and controllable deficit simulation along a continuous severity spectrum.
- Mechanism: Severity \(s \in [0,1]\) is defined. A line search first identifies the minimum effective scaling coefficient \(\alpha^*\) (searched within \([1, 6]\), ensuring deficits are observable without causing incoherent output). At inference, each token generation step samples \(z_t \sim \text{Bernoulli}(s)\), and \(\alpha^* \cdot \mathbf{v}_d\) is injected into the hidden state only when \(z_t = 1\).
- Design Motivation: Traditional methods adjust \(\alpha\) to control intensity, but the relationship between \(\alpha\) and output behavior is highly nonlinear—small increases in \(\alpha\) may produce no change, while large increases may cause output collapse. STM shifts the control variable from amplitude to probability; statistically, larger \(s\) results in more modulated tokens, yielding smoother and more predictable effects.
Automatic Parameter Selection:
- Function: Eliminates manual tuning, exposing only the severity level \(s\) as the user-facing control knob.
- Mechanism: \(\alpha^*\) is determined automatically via line search (subject to validity and coherence criteria), and the optimal layer \(l^*\) is selected automatically by maximizing the distance between positive and negative sample embedding centroids.
- Design Motivation: Traditional steering vector methods require manual tuning of \(\alpha\) and layer selection, limiting practical usability.

Key Experimental Results¶

Main Results (GPT-4 Therapist Scenario, LLM + Human Evaluation)¶

Method	CDC↑(LLM)	CDC↑(Human)	IDI↓(LLM)	IDI↓(Human)	Auth↑	Tra↑
Direct Prompt	0.54	0.68	0.47	0.42	3.32	3.40
PATIENT-ψ	0.50	0.60	0.52	0.48	3.83	3.96
Roleplay-doh	0.58	0.68	0.44	0.38	3.78	3.72
StsPatient	Best	Best	Best	Best	Best	Best

Ablation Study¶

Configuration	Description
w/o STM (scaling \(\alpha\) only)	Severity control is unstable; high \(\alpha\) causes output collapse
Prompt contrastive only	Lacks behavioral-level information; deficit manifestation is less natural
Response contrastive only	Lacks intent-level signal; domain specificity is reduced
Full StsPatient	Best across all metrics

Key Findings¶

StsPatient achieves an average improvement of 11.23% across all metrics and surpasses the best baseline by 18.54% in severity controllability.
STM is critical: Traditional scaling methods frequently produce incoherent output when \(\alpha > 4\), whereas STM maintains linguistic integrity even at \(s = 0.9\).
Steering vectors for different cognitive domains do encode distinct deficit characteristics: the attention deficit vector and memory deficit vector point in clearly different directions in the representation space.
The correspondence between severity \(s\) and clinical scores is monotonic, and while not strictly linear, it satisfies the requirements of educational simulators.

Highlights & Insights¶

The analogy from synaptic transmission probability to token modulation probability is highly elegant—it transfers a control principle from neuroscience to LLM behavioral control. This "probabilistic gating" approach is broadly applicable to any scenario requiring continuously controllable behavioral modulation.
The domain specificity of steering vectors demonstrates that LLM hidden state spaces genuinely encode linguistic features corresponding to different cognitive domains, offering insights for interpretability research.
Inference-time intervention without fine-tuning makes the method plug-and-play across different LLMs.

Limitations & Future Work¶

The mapping between severity \(s\) and standard clinical scores (e.g., MMSE) is not directly linear.
Cognitive domains are currently defined manually; whether the latent dimensions of cognitive impairment can be discovered automatically remains an open question.
The stability of steering vectors depends on the quality of the synthesized contrastive data.
Validation is conducted only on English data; cross-lingual generalization (e.g., linguistic features of Chinese-speaking patients with cognitive impairment) has yet to be explored.
No comparison has been made against real clinical data (e.g., recorded conversations with actual AD patients).

vs. Prompt-based SP: Prompts provide discrete, coarse-grained control, whereas StsPatient operates in a continuous representation space for finer-grained control.
vs. Traditional SV methods (Rimsky et al.): Traditional methods scale \(\alpha\), leading to instability; STM resolves this core issue through probabilistic control.
vs. PATIENT-ψ: PATIENT-ψ focuses on narrative control without domain-specific deficit simulation; StsPatient enables precise control over which domain is impaired.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ The biologically inspired design of the STM mechanism is highly original; domain-specific application of steering vectors is also a first.
Experimental Thoroughness: ⭐⭐⭐⭐ Both LLM-based and human evaluations are included, but comparison against real clinical data is absent.
Writing Quality: ⭐⭐⭐⭐⭐ Motivation and methodology are clearly articulated with intuitive illustrations.
Value: ⭐⭐⭐⭐ Practically significant for clinical AI training; the STM approach is transferable to other behavioral control scenarios.