Anticipating User Interpretations in Emotional Support Conversations via Inference-Time User-Side Simulation¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: Dialogue Systems
Keywords: Emotional Support Conversation, User-Side Simulation, Cognitive World Model, Inference-Time Look-Ahead, Selective Revision
TL;DR¶
Addressing the frequent conversational breakdown where sensible advice is misperceived by help-seekers as dismissive, judgmental, or pressuring, this paper presents Cognitive Look-ahead Simulation (CLS), an inference-time framework that infers structured user states, anticipates immediate user inner thoughts and state transitions via a lightweight cognitive world model, and selectively rewrites high-risk drafts before delivery, substantially reducing severe failure rates and improving multi-turn dialogue outcomes.
Background & Motivation¶
The success of an emotional support dialogue system cannot be judged purely by the surface-level linguistic quality of its responses; rather, it hinges fundamentally on how those responses are interpreted and received by the user. While recent advances in large language models have markedly improved empathetic expression, dialogue history modeling, and multi-turn support strategy planning, the vast majority of existing conversational agents remain essentially reactive. They condition on dialogue context and directly generate the next turn in a single pass, without explicitly evaluating how a concrete candidate utterance is likely to land with the user.
This reactive generation paradigm exposes a critical tension in emotionally supportive dialogue: conversational failures are predominantly failures of user interpretation rather than failures of phrasing. In psychological support, advice that appears completely reasonable, factually sound, and well-intentioned in isolation can still be heard by a vulnerable, distressed, or defensive user as preachy, patronizing, dismissive, or pressuring. Moreover, the user's internal psychological state is only partially observable; the assistant cannot directly read the user's cognitive schema, core beliefs, or unvoiced needs from surface utterances alone. Prior self-refinement techniques (such as Self-Refine or CRITIC) rely almost exclusively on assistant-side feedbackโcritiquing whether the response conforms to guidelines or is politeโwhile completely missing the user-side psychological uptake.
To address this challenge, the authors formulate emotional support response generation as a problem of candidate response evaluation under user-side interpretation uncertainty. Core idea: expand single-pass response generation into an inference-time "GenerateโSimulateโEvaluateโRevise" pipeline that infers a structured user-state representation, uses a Cognitive World Model as an operational one-step predictor of the user's immediate inner thought and state transition under a candidate draft, and selectively applies targeted revision to high-risk drafts.
Method¶
Overall Architecture¶
Cognitive Look-ahead Simulation (CLS) operates entirely at inference time between initial drafting and final response delivery. When a user utterance arrives, CLS first parses the dialogue history to infer a two-tiered Structured User-State Representation, separating session-level stable profiles from dynamic turn-level snapshots. Conditioning on both the history and this structured state, the generator produces an initial candidate draft. Next, an operational one-step Cognitive World Model (CWM) acts as a virtual user simulator, predicting the user's immediate inner thought and next-turn state shift if that draft were sent as is. Finally, a hybrid risk evaluator inspects these simulated consequences; if a trust drop, emotional resistance, or need mismatch is detected, the draft is selectively rewritten before delivery; otherwise, the original draft is passed through untouched.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Dialogue History & Current Input"] --> B["Two-Tiered User-State Inference<br/>infer session profile & turn snapshot"]
B --> C["Candidate Reply Generation<br/>draft response conditioned on state"]
C --> D["Cognitive World Model Simulation<br/>predict inner thought & next-turn state"]
D --> E{"Hybrid Risk Evaluation<br/>check trust drop & need mismatch"}
E -->|high risk detected| F["Selective Targeted Revision<br/>rewrite draft with uptake & hints"]
E -->|low risk| G["Final Transmitted Reply"]
F --> G
Key Designs¶
1. Two-Tiered User-State Inference: Decoupling Session Profiles and Turn Snapshots
Because a user's internal psychological landscape is unobservable and different psychological factors evolve at vastly different tempos, CLS introduces an explicit structured representation \(S_t = (\Sigma, z_t)\). The session-level profile \(\Sigma\) captures relatively persistent cognitive tendencies within the session, structured into cognitive schema (core beliefs, self-view, world-view, interpretation bias), coping and resistance patterns (coping style, resistance triggers), and motivational context (hidden needs, motivation hooks, taboos). In contrast, the turn-level snapshot \(z_t = (\tau_t, e_t, n_t)\) tracks fast-moving state variables at each turn: a coarse trust level \(\tau_t \in \{\text{low}, \text{medium}, \text{high}\}\), the user's immediate emotional tone \(e_t\), and their active conversational need \(n_t\). This decoupling prevents the model from conflating temporary emotional turbulence with fundamental personality traits, providing a transparent and steerable anchor for generating the initial draft \(a_t \sim \pi_\theta(\cdot \mid \Sigma, z_t, \bar{H}_t)\).
2. Cognitive World Model Simulation: Operational One-Step User Uptake Prediction
Rather than relying on assistant-centric critique or attempting intractable long-horizon tree searches over open-ended dialogue, CLS instantiates a Cognitive World Model (CWM) as a localized one-step user-side predictor. Conditioned on the structured state \(S_t\), dialogue history \(\bar{H}_t\), and candidate reply \(a_t\), the CWM predicts immediate user-side consequences: $\((I_{t+1}, z_{t+1}) \sim p_\phi(\cdot \mid \Sigma, z_t, \bar{H}_t, a_t)\)$ Here, \(I_{t+1}\) is a predicted user uptake formulated as a first-person natural-language inner thought describing how the user would privately interpret the reply (e.g., "He is preaching at me again; nobody actually gets it"), while \(z_{t+1}\) represents the predicted next-turn state transition. By concretizing latent psychological reactions into inspectable natural language and state transitions, the system obtains a direct preview of potential pragmatic friction before committing to an utterance.
3. Selective Targeted Revision: Risk-Gated Harm Reduction
Blindly revising every response adds substantial latency and risks over-hedging. CLS deploys a hybrid evaluator \((d_t, \kappa_t) = G(\Sigma, z_t, a_t, I_{t+1}, z_{t+1})\) combining an LLM critic with rule-based transition alarms. Revision (\(d_t = 1\)) is triggered only when concrete risk signals emerge: the predicted inner thought \(I_{t+1}\) conveys feeling judged, dismissed, pressured, or misunderstood; the predicted trust level \(\tau_{t+1}\) drops below \(\tau_t\); the draft violates conversational taboos or resistance triggers in \(\Sigma\); or the response mismatches the active need \(n_t\). Upon triggering, the rewriter constructs the final reply using the original draft, the anticipated negative interpretation, and revision hints \(\kappa_t\): $\(a^*_t = \begin{cases} R_\theta(\Sigma, z_t, \bar{H}_t, a_t, I_{t+1}, \kappa_t), & d_t = 1 \\ a_t, & d_t = 0 \end{cases}\)$ This selective gating ensures that revisions function as surgical harm reduction specifically targeting interactionally dangerous drafts while leaving safe drafts untouched.
A Worked Example¶
Consider an interaction with a 17-year-old student distressed by study avoidance and gaming addiction: - User Utterance: "I can't focus lately; I just want to squad up and play. It's so frustrating." - Inferred User State: The session profile \(\Sigma\) records core belief "I'm not in control", interpretation bias "advice sounds like criticism", hidden need "autonomy and being understood rather than lectured", and taboo "being told what to do". The turn snapshot \(z_t\) records trust as medium, emotion as frustrated, and active need as clarity about choices. - Initial Draft: The assistant drafts a standard problem-solving response: "You should set an alarm. Play less and study more." - CWM Simulation: The CWM projects the user's inner thought \(I_{t+1}\): "Here we go again with the same old preaching... Stop lecturing me!" The projected trust level drops from medium to low. - Evaluation & Revision: The evaluator flags high risk (\(d_t=1\)) due to preachy tone and taboo violation. The rewriter drops prescriptive directives and pivots to emotional validation and reflective exploration: "That gaming rush is hard to replace. What else makes you feel 'Godlike'?" - Outcome: The revised reply disarms defensiveness, validates the underlying psychological driver, and prompts the user to constructively reflect: "If only I could go 'Godlike' on an exam too..."
Loss & Training¶
While CLS is primarily an inference-time framework requiring no training of the underlying generator, deploying proprietary frontier models (like Gemini 2.5 Flash) as the CWM incurs latency and privacy overhead. To enable local, cost-effective inference, the authors construct a Distilled-CWM. They collect 500 teacher dialogue trajectories on the SAGE training split using Gemini 2.5 Flash as the CWM, recording intermediate structured states, simulated inner thoughts, and evaluation traces. A Qwen2.5-7B-Instruct student model is then fine-tuned via standard autoregressive cross-entropy loss on these structured traces using LoRA (rank \(r=8\), \(\alpha=16\), learning rate \(3 \times 10^{-5}\) over 3 epochs), yielding a compact, open-weights world model tailored for user-side psychological simulation.
Key Experimental Results¶
Main Results¶
Evaluation is conducted on the SAGE benchmark using an LLM-based user simulator, the Sentient Agent instantiated with DeepSeek-V3.2, across \(N=100\) distinct test profiles. Each profile defines persona, background, explicit goal, and hidden intention. Key metrics include Avg (mean outcome score in \([0, 100]\)), Succ (percentage of dialogues scoring \(\ge 90\)), and Fail (percentage scoring \(\le 10\)).
| Model Category | Backbone / System | Configuration | Outcome Avg | Absolute Gain (\(\Delta\)) | Success Rate (Succ) | Failure Rate (Fail) |
|---|---|---|---|---|---|---|
| Open-Source | Qwen2.5-7B-Instruct | Vanilla (Direct Prompting) | 8.02 | โ | 0% | 92% |
| Open-Source | Qwen2.5-7B-Instruct | SFT (Best-of-4 Datasets) | 18.82 | +10.80 | 0% | 58% |
| Open-Source | Qwen2.5-7B-Instruct | CoT (Reasoning-Augmented) | 10.60 | +2.58 | 0% | 78% |
| Open-Source | Qwen2.5-7B-Instruct | Self-Critique (Single-Pass) | 15.32 | +7.30 | 0% | 66% |
| Open-Source | Qwen2.5-7B-Instruct | Self-Refine (Iterative) | 13.30 | +5.28 | 2% | 80% |
| Open-Source | Qwen2.5-7B-Instruct | CLS (Self-CWM) | 36.74 | +28.72 | 0% | 66% |
| Open-Source | Qwen2.5-7B-Instruct | CLS (Gemini-CWM) | 39.32 | +31.30 | 4% | 34% |
| Open-Source | Qwen2.5-7B-Instruct | CLS (Distilled-CWM) | 42.18 | +34.16 | 2% | 32% |
| Open-Source | Llama-3-8B-Instruct | Vanilla (Direct Prompting) | 35.34 | โ | 2% | 42% |
| Open-Source | Llama-3-8B-Instruct | SFT (Best-of-4 Datasets) | 25.26 | -10.08 | 4% | 52% |
| Open-Source | Llama-3-8B-Instruct | Self-Refine (Iterative) | 40.10 | +4.76 | 4% | 36% |
| Open-Source | Llama-3-8B-Instruct | CLS (Distilled-CWM) | 58.48 | +23.14 | 8% | 14% |
| Frontier | Llama-4-Scout | Vanilla / CLS (Distilled) | 28.78 / 61.26 | +32.48 | 2% / 12% | 34% / 12% |
| Frontier | GPT-4o | Vanilla / CLS (Distilled) | 60.07 / 72.50 | +12.43 | 22% / 34% | 16% / 4% |
| Frontier | Grok-4 Fast | Vanilla / CLS (Distilled) | 63.18 / 70.14 | +6.96 | 4% / 14% | 4% / 4% |
| Domain Models | EmoLLM-v3.0 | Standalone Release | 16.22 | โ | 2% | 76% |
| Domain Models | SoulChat2.0-Llama3-8B | Standalone Release | 46.36 | โ | 6% | 26% |
Ablation Study¶
The authors dissect the contributions of structured user-state representation (SUSR), inner thought prediction, hidden need grounding, and the cognitive world model on both open-source backbones (using Gemini-CWM as the reference predictor):
| Backbone | Ablation Variant | Avg Score | Absolute Gain (\(\Delta\)) | Drop vs. Full CLS |
|---|---|---|---|---|
| Qwen2.5-7B-Instruct | Vanilla (w/o CLS) | 8.02 | โ | -31.30 |
| Qwen2.5-7B-Instruct | w/o Look-ahead (SUSR-only) | 23.94 | +15.92 | -15.38 |
| Qwen2.5-7B-Instruct | w/o Inner Thought | 32.00 | +23.98 | -7.32 |
| Qwen2.5-7B-Instruct | w/o Hidden Need | 34.62 | +26.60 | -4.70 |
| Qwen2.5-7B-Instruct | w/o CWM (Assistant-side critique) | 37.46 | +29.44 | -1.86 |
| Qwen2.5-7B-Instruct | Full CLS (Gemini-CWM) | 39.32 | +31.30 | 0.00 |
| Llama-3-8B-Instruct | Vanilla (w/o CLS) | 35.34 | โ | -19.54 |
| Llama-3-8B-Instruct | w/o Look-ahead (SUSR-only) | 46.92 | +11.58 | -7.96 |
| Llama-3-8B-Instruct | w/o Inner Thought | 54.70 | +19.36 | -0.18 |
| Llama-3-8B-Instruct | w/o Hidden Need | 53.14 | +17.80 | -1.74 |
| Llama-3-8B-Instruct | w/o CWM (Assistant-side critique) | 53.66 | +18.32 | -1.22 |
| Llama-3-8B-Instruct | Full CLS (Gemini-CWM) | 54.88 | +19.54 | 0.00 |
Key Findings¶
- Drastic Reduction in Severe Failures: Across backbones, single-pass generation suffers catastrophic failure rates (e.g., 92% on Qwen2.5-7B and 42% on Llama-3-8B). CLS slashes failures to 32% and 14%, respectively. This confirms that emotional support collapses not from a lack of vocabulary, but from high-risk drafts triggering irreversible user defensiveness.
- Task-Specific SFT Fails Multi-Turn Support: Fine-tuning backbones on established datasets (AUGESC, SerForEmo, ESConv, ExTES) often yields negative returns under SAGE (e.g., drops from 35.34 to 8.97โ25.26 on Llama-3-8B). SFT models learn surface empathetic templates and premature advice giving, which actively clash with dynamic user needs.
- Targeted Harm Reduction, Not Uniform Boosting: Paired counterfactual rollouts (Table 21) demonstrate that selective revision provides asymmetric utility: on the most severe 25% draft-negative subset, revision produces an average score gain of +5.375 and a 68.8% pairwise improvement rate, confirming its function as a safety net.
- Distilled CWM Surpasses Proprietary Teachers: Qwen2.5-7B fine-tuned on simulation traces outperforms Gemini 2.5 Flash as the CWM (42.18 vs. 39.32 on Qwen; 58.48 vs. 54.88 on Llama), demonstrating that task-specialized distillation sharpens the model's sensitivity to subtle user resistance.
Highlights & Insights¶
- User-Side Simulation Paradigm Shift: By shifting evaluation from "how well did the assistant follow instructions?" to "how will the user interpret this utterance in their private thoughts?", CLS addresses the pragmatic core of interpersonal communication.
- Principled Decoupling of Temporal Scales: Explicitly partitioning user state into static cognitive schemas and transient emotional snapshots shields the system from context hallucinations where single turns wipe out global user persona.
- High-Value Selective Intervention: With an average revision triggering rate of ~35%, CLS avoids unnecessary re-generation on innocuous turns while concentrating test-time compute precisely where interpersonal risk is acute.
Limitations & Future Work¶
- Inference-Time Latency Overhead: The full pipeline requires user-state parsing, draft generation, CWM prediction, evaluation, and selective revision, accumulating ~15.9 seconds per turn. Although caching schemes (ProfileCached) reduce this to ~13.1s, further acceleration is required for interactive real-time deployment.
- Reliance on Simulated Proxies: Primary validation relies on the SAGE Sentient Agent. Although corroborated by human evaluations and a 30-case user pilot, comprehensive clinical studies involving real help-seekers in crisis remain essential before real-world deployment.
- Risk of Excessive Agreeableness: Over-optimizing to avoid user resistance could inadvertently encourage sycophancy or validate maladaptive behaviors, requiring future safeguards to balance autonomy support with constructive psychological boundaries.
Related Work & Insights¶
- vs. Strategy-Guided Emotional Support Systems (e.g., PAL, MISC, EmoDynamiX): Traditional methods focus on predicting assistant action strategies (e.g., empathy, questioning, advice), which operate at too coarse a granularity to prevent pragmatic misinterpretation; CLS models the fine-grained psychological reaction at the receiving end.
- vs. Self-Refinement Frameworks (e.g., Self-Refine, CRITIC, Reflexion): Standard self-refinement evaluates drafts against internal rubrics from the generator's perspective; CLS grounds critique in an explicit world model simulating the user's Theory of Mind.
Rating¶
- Novelty: โญโญโญโญโญ First framework to formalize emotional support generation as candidate evaluation under user-side interpretation uncertainty via an operational cognitive world model.
- Experimental Thoroughness: โญโญโญโญโญ Comprehensive evaluation across SAGE multi-turn simulation, multi-judge LLM profiling, third-party human ratings, profile-conditioned evaluation, first-person user pilot, and ethical value-alignment audits.
- Writing Quality: โญโญโญโญโญ Clear conceptual formulation, rigorous ablation design, and rich qualitative case studies.
- Value: โญโญโญโญโญ Provides an effective inference-time architectural blueprint for combating preachy advice and conversational breakdown in sensitive human-AI interactions.