acceptodds
Under review as a conference paper at ICLR 2027

Motivationally Grounded Role-Playing Agents: Stress-Driven Psychodynamic Modeling

Abstract

Role-playing agents (RPAs) powered by large language models have become central to interactive character simulation, yet existing approaches often rely on surface-level persona descriptions and stylistic imitation, lacking a principled account of the internal psychological dynamics that drive role-consistent behavior. This limitation leads to shallow personality modeling and character drift in extended interactions. We propose a motivationally grounded framework that explicitly models a character's internal motivational structure. Grounded in psychodynamic, attachment, moral, existential, and interpersonal theories, we construct a structured character psychology that includes a taxonomy of core anxieties, a corresponding set of compensatory desires, three personality organization types, and a mapping from each personality organization and stress level to the character's available defense mechanisms. These structures are embedded in a dual-process generation architecture. On SocialBench, our method outperforms models of comparable scale. A controlled probe study on a purpose-built stress dataset further confirms that the framework's internal mechanisms operate as intended, with defense tier governing modulation intensity and defense selection producing distinct surface forms from the same impulse. These results suggest that explicitly modeling character psychology offers a scalable and effective pathway toward deeper role-playing fidelity.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Role-Playing Agents, Large Language Models, Motivational Modeling, Psychodynamic Framework

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NjkbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PsyDrive, a psychology-grounded framework for role-playing agents that integrates core anxieties, compensatory desires, personality organization, and defense mechanisms into a stress-driven response generation architecture. The authors evaluate the framework through SocialBench comparisons, ablation studies, and stress-probe analyses, reporting improved benchmark performance and differences in generated responses across psychological conditions.

Critical strengths
  1. An important and interesting research direction. The paper addresses the limitations of static persona-based role-playing agents by exploring how internal motivations and stress-driven psychological dynamics can shape character responses. This is a worthwhile direction for developing more context-sensitive and consistent role-playing agents.

  2. A concrete implementation with empirical performance improvements. The authors operationalize psychological constructs within a structured response-generation architecture and provide benchmark comparisons and ablation studies. PsyDrive shows notable improvements over its base model on SocialBench.

Critical weaknesses
  1. Insufficient theoretical justification for integrating psychological constructs. The paper draws on multiple psychological theories to construct its framework, but the rationale for selecting and connecting these constructs remains insufficiently established. In particular, it is unclear how the proposed categories and mappings are derived from the underlying theories, or why stress-driven and anxiety-based mechanisms are prioritized over other motivational processes. The framework appears to rely heavily on predefined design choices without sufficient justification for their psychological validity.

  2. Misalignment between the research claims and empirical evaluation. Although the paper aims to model psychologically grounded motivations and authentic character responses, the evaluation primarily relies on SocialBench metrics that do not directly assess these constructs. The stress-probe analyses largely demonstrate that the predefined mechanisms operate as designed, rather than establishing their psychological validity.

Questions for authors
  1. How were the psychological constructs and their relationships selected and justified? In particular, what motivates the focus on anxiety and stress over other motivational processes?
  2. What evidence supports the psychological validity of the proposed taxonomy and the mappings between core anxieties, compensatory desires, personality organization, and defense mechanisms?
  3. How does performance on SocialBench demonstrate improvements in psychological depth or motivational coherence, beyond conventional role-playing capabilities?
  4. How do the stress-probe analyses establish psychological validity rather than simply confirming that the predefined rules operate as intended?
  5. Why was an uncensored Draft model used for the mechanistic analyses but not for the main experiments? How might this difference affect the interpretation of the findings?
1 comment

We thank the reviewer for the careful reading. The new analyses are in the supplementary material; we summarize the conclusions here.

Q1. Core anxieties, compensatory desires, personality organization, and defenses are grounded in psychodynamic, attachment, moral, existential, and interpersonal theory (Appendices A–D). We prioritize anxiety/stress because interpersonal pressure makes defense mechanisms observable and yields an executable state-to-tier mapping. The taxonomy is a principled, executable subset tailored to role-playing rather than an exhaustive motivational theory, and we will state this scope explicitly.

Q2. The new evidence supports operational and construct validity in generation. The internal mechanism is checkpoint-invariant, the defense–impulse interaction survives alignment, the paired intervention shows that varying only the defense layer causally changes delivery, and tier lowering produces a graded response. Taken together, these establish causal, graded surface effects rather than a check that the rules run as designed (Table 24).

Q3. SocialBench measures role-playing and social consistency rather than psychological depth directly. Component isolation localizes the gain: defense-directed rewrite has the highest marginal efficiency, and gains over RichProfile concentrate in cross-turn memory dimensions, where RichProfile fails long-context consistency despite scoring higher on style and positive preference. This pattern is consistent with state accumulation and defense modulation (Table 29).

Q4. The paired and dose-response analyses manipulate only the defense layer while holding draft, history, and state fixed. Defense direction moves arousal and pragmatic markers in the predicted direction, tier lowering yields a monotonic pragmatic-marker gradient, and the aligned null is explained by the measured copy-rate compression. This is causal evidence about the mechanism rather than a confirmation of the implementation (Table 25).

Q5. We used the uncensored draft originally because aligned drafts and rewrites were too similar, which compressed surface observability. The main benchmark uses Qwen2.5-7B-Instruct, and the new Instruct probe shows the mechanism is present under both checkpoints, with only surface observability reduced under alignment (Table 21).

Reviewer_csPsOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PsyDrive combines social-threat perception, recurrent psychological states, and a draft-and-rewrite pipeline to generate stress-dependent character behavior. On Qwen2.5-7B, it raises SocialBench average accuracy from 65.12% to 76.13%. Ablations and synthetic stress probes examine the contributions of its components and their effects on response style.

Critical strengths
  1. The framework makes character state explicit and updates it across turns. Connecting perceived threats to motivations and response strategies gives a concrete, inspectable design for role-playing agents without fine-tuning.
  2. The reported benchmark improvement is substantial, particularly on long-context comprehension, which rises from 34.64% to 77.73%. Component ablations also provide useful evidence about which parts of the implementation affect performance.
Critical weaknesses
  1. The mechanism study uses an uncensored draft checkpoint, whereas all main experiments use Qwen2.5-7B-Instruct. Appendix H explains that aligned drafts and rewrites were too similar. The probe therefore does not establish that the same mechanism explains the benchmark gains. Both settings need comparable behavioral analyses.
  2. The baselines include CoT, ToT and self-correction, but their scores are adopted from prior work and generation costs are not matched. Removing rewrite costs only 0.38 average points. Rerun representative baselines and compare rich-profile prompting and generic state-conditioned draft/rewrite at similar token budgets. 3.The probe offers limited causal evidence. Personality organization affects defense tiers through predefined capacities and thresholds, while the "same impulse" comparison groups different drafts into five broad categories. Hold the draft and history fixed, intervene on defense selection, and compare the intended mappings with shuffled or generic alternatives.
  3. Psychological coherence lacks independent validation, as the limitations acknowledge. SocialBench accuracy and phrase counts do not establish believable stress responses or breakdowns; absolute sentiment logits are also an unvalidated arousal measure. Blinded human ratings and expert checks of selected constructs would directly test the central behavioral claims.
Questions for authors
  1. Which probe effects survive with the aligned draft model used in the main experiments? Does the uncensored variant produce comparable SocialBench gains, and which exact checkpoint was used?
  2. How does PsyDrive compare with rich-profile prompting and a generic state-conditioned draft/rewrite pipeline under matched token budgets? What additional benefit does the full psychological structure provide?
  3. With identical drafts and histories, do interventions on defense selection produce the predicted changes? Are the proposed mappings more effective than shuffled or generic mappings?
  4. Do blinded human judgments support improved character coherence under stress and credible breakdown behavior? Can the proposed arousal measure be validated against independent annotations?
  5. Do the main gains and probe effects remain under repeated seeds, character-clustered uncertainty and threshold variation? Does the method help a second backbone or a larger model?
1 comment

We thank the reviewer for the careful reading. The new analyses are in the supplementary material; we summarize the conclusions here.

Q1. Both the main experiments and the new probe use Qwen2.5-7B-Instruct. Under Instruct, internal mechanism effects survive while surface observability is compressed, and the compression is measurable through the copy rate. The uncensored variant serves only as a mechanism-discovery setting; we do not use it to explain benchmark gains (Tables 19–21).

Q2. We added four isolation configurations with a per-sample token ledger: RichProfile, Generic Two Stage, PsyDrive$_{-def}$, and full PsyDrive. Average accuracy is monotonic, but the more informative result is marginal efficiency: defense-directed rewrite yields +2.25+2.25 average points for roughly +276+276 tokens per sample, while gains over RichProfile cost 2.42×2.42\times the tokens and concentrate in CM Short/Long. The net +0.38+0.38 rewrite effect decomposes into a negative generic-rewrite component and a positive defense-direction component (Table 29).

Q3. The paired intervention holds draft, history, role profile, and psychological state fixed and varies only the defense layer, with the generic condition as control. It reproduces the predicted restraint and mitigation effects, and tier lowering yields a graded pragmatic-marker response. A shuffled-mapping control would be a useful extension, which we leave for future work (Tables 24–25).

Q4. Blinded human ratings are outside the scope of the present revision; our claims are scoped to mechanism-level causal effects rather than perceived coherence. The sentiment-logit arousal measure is auxiliary: it is non-significant under the aligned checkpoint and shows no reliable tier slope, so the main conclusions do not depend on it (Table 22).

Q5. Multi-seed runs, character-clustered uncertainty, threshold sweeps, and a second backbone are not yet in place, and we will mark them as limitations in the revision. We still expect the method to transfer: each stage depends only on instruction following, a capability that current LLMs share across families and scales, and Qwen2.5-7B already follows the defense-directed rewrite reliably. Within-sample paired tests, stratified checks, and the exclusion of copied samples address part of the concern, and we leave cross-backbone and multi-seed verification to future work.

Reviewer_P9L5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PsyDrive, a role-playing framework that integrates core anxieties, compensatory desires, personality organization, and defense mechanisms. It maintains psychological states across turns and generates responses through a draft-and-rewrite pipeline. Using Qwen2.5-7B-Instruct, PsyDrive improves average SocialBench accuracy from 65.12% to 76.13%. A separate stress-probe study examines relationships between internal states and surface expression.

Critical strengths

S1.Connecting threat perception, accumulated stress, and defensive expression provides useful interfaces for inspecting and intervening in character behavior across turns.

S2.The reported 11.01% improvement over the backbone, together with component ablations, suggests practical value.

Critical weaknesses

W1.The probe uses an uncensored Draft model, whereas the main experiments use the standard Instruct model. Consequently, probe findings cannot directly explain the benchmark improvements. Moreover, personality–defense associations are partly imposed by predefined rules, and comparing different texts within broad impulse categories does not establish the effect of applying different defenses to the same draft. These analyses provide limited support for the paper’s stronger mechanistic claims.

W2.Removing Social Perception and Psychological State retains approximately 81.5% of the overall improvement; removing rewriting reduces average accuracy by only 0.38%, without reported multi-seed uncertainty. The paper lacks targeted controls with equivalent character information and inference budgets, as well as independent evaluation of character believability or long-term consistency.

Questions for authors

Q1.Do the mechanistic findings hold under the standard Instruct configuration used in the main experiments?

Q2.Can paired interventions that hold the draft fixed and vary only the defense mechanism demonstrate reliable effects?

Q3.Under matched character information and inference budgets, does PsyDrive consistently outperform static psychological profiles and generic two-stage rewriting? Is there direct evidence of improved character believability?

1 comment

We thank the reviewer for the careful reading. The new analyses are in the supplementary material; we summarize the conclusions here.

Q1. We reran the probe under Qwen2.5-7B-Instruct with identical prompts, candidate pools, and extraction code. The internal mechanism is checkpoint-invariant: tier assignment, state accumulation, breakdown delivery, and the personality-organization × defense-tier contingency all reproduce, including the structural zero, and the defense–impulse interaction survives. What alignment changes is surface observability, and the cause is measurable: the rewrite stage copies the draft far more often under Instruct (30.6%30.6\% vs. 2.9%2.9\%), and the copy rate rises with tier. The null result of the aligned intervention follows from this compression rather than from an inactive defense layer (Tables 19–21).

Q2. We added a paired intervention in which model, sampling configuration, draft, history, role profile, and psychological state are held fixed and only the defense layer varies. Specifying a defense makes the rewrite more restrained and more mitigated: the arousal delta turns negative (pBonf=1.1×10−4p_{\text{Bonf}}=1.1\times10^{-4}), the pragmatic-marker delta more than quadruples (+0.103+0.103 vs. +0.022+0.022, pBonf<10−8p_{\text{Bonf}}<10^{-8}), while polarity stays null. Lowering the tier produces a monotonic pragmatic-marker gradient (p<0.001p<0.001) (Tables 24–25).

Q3. For matched information and budget, yes. Four isolation configurations give monotonic average accuracy: RichProfile 72.0572.05 < Generic Two Stage 72.7572.75 < PsyDrive$_{-def}$ 73.8873.88 < PsyDrive 76.1376.13. Defense-directed rewrite is nearly free and has the highest marginal efficiency, +2.25+2.25 average points for roughly +276+276 tokens per sample (8.148.14 vs. 1.031.03 points per 1K tokens relative to RichProfile), whereas generic rewrite alone hurts. SocialBench targets role-playing and social consistency rather than perceived believability, so we read these gains as evidence for the mechanism and treat an independent believability study as future work (Table 29).

Comments on the paper

1 comment

Dear Area Chair,

We thank the reviewers for their constructive feedback. The additional analyses and clarifications directly address the main concerns and support the paper's central claims. We summarize them below; details are in the paper (marked in blue).

Responses to Reviewers P9L5, csPs, and Njkb.

Checkpoint and alignment effects. We reran the probe under Qwen2.5-7B-Instruct with identical prompts, candidate pools, and extraction code. The internal mechanism is checkpoint-invariant: tier assignment, state accumulation, breakdown delivery, the personality-organization × defense-tier contingency, and the defense–impulse interaction all reproduce. Alignment mainly compresses surface observability, measurable via a higher copy rate under Instruct (30.6%30.6\% vs. 2.9%2.9\%). The aligned null therefore follows from this compression, not from an inactive defense layer (Tables 19–21). The uncensored variant is only a mechanism-discovery setting; the main benchmark uses Qwen2.5-7B-Instruct.

Causal evidence for the defense layer. A paired intervention holds model, sampling configuration, draft, history, role profile, and psychological state fixed and varies only the defense layer, with the generic condition as control. It reproduces the predicted restraint and mitigation effects, and tier lowering yields a graded pragmatic-marker response (Tables 24–25). This provides causal, graded surface evidence for the mechanism rather than merely confirming the implementation.

Efficiency and component isolation. Four isolation configurations give monotonic average accuracy: RichProfile 72.0572.05 < Generic Two Stage 72.7572.75 < PsyDrive$_{-def}$ 73.8873.88 < PsyDrive 76.1376.13. Defense-directed rewrite has the highest marginal efficiency (+2.25+2.25 average points for roughly +276+276 tokens per sample), whereas generic rewrite alone hurts. Gains over RichProfile concentrate in cross-turn memory dimensions, consistent with state accumulation and defense modulation (Table 29).

Theory, construct validity, and scope. Constructs are grounded in psychodynamic, attachment, moral, existential, and interpersonal theory (Appendices A–D). The taxonomy is a principled, executable subset tailored to role-playing, not an exhaustive motivational theory, and we have clarified this scope explicitly. New evidence supports operational and construct validity in generation: the mechanism is checkpoint-invariant, the defense–impulse interaction survives alignment, the paired intervention causally changes delivery, and tier lowering produces a graded response. SocialBench measures role-playing and social consistency rather than perceived believability directly; our claims are scoped to mechanism-level causal effects, supported by component isolation.

We hope these additional experiments and clarifications address the reviewers' main concerns and further strengthen the paper.

Best,

The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.