acceptodds
Under review as a conference paper at ICLR 2027

MindLoop: LLM-Based Closed-Loop Emotion Regulation with Wearable BCI Feedback

Abstract

Emotion regulation is essential for mental health, but access to trained professional support remains limited. Large Language Models (LLMs) offer a promising way to make emotional support more accessible through conversational interaction. However, existing studies on LLM-based emotional support mainly focus on current turn response quality, while paying limited attention to the overall regulation process. Effective emotion regulation is a structured closed-loop process requiring multi-stage guidance and timely affective feedback, raising three challenges: structuring dialogue, balancing current response quality with overall regulation outcomes, and obtaining timely affective observations. We propose MindLoop, a closed-loop policy learning approach for LLM-based emotion regulation with wearable BCI feedback. MindLoop first introduces a CBT-guided stage–strategy plan to organize multi-turn dialogue into structured regulation stages and stage-specific interventions. Based on this structure, we further develop a dual-timescale reinforcement learning method to balance overall regulation progress and current-stage intervention quality. To provide objective and timely observations of affective changes, we introduce wearable EEG/fNIRS-based affect decoding as physiological feedback for adapting regulation decisions. Simulated-interaction evaluation shows strong regulation-stage progression, successful completion, and affective outcomes. A 23-participant study further provides preliminary evidence that the simulation-trained policy supports structured real-world interaction with favorable affective changes and session ratings.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Emotion regulation, LLM-based dialogue, wearable BCI feedback

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_L5djOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

As users turn to LLMs for emotional support, this paper aims to address the gap in improving LLM-based emotion regulation. The authors propose a closed-loop framework including an LLM-based planner called MindLoop, which is trained through back-and-forth conversations with simulated users. The planner is optimized using RL and the authors' proposed policy learning approach to improve users' emotion regulation. They then evaluate the trained system in a real-world setting with human participants, using wearable BCI to capture emotional states and adapt the AI's responses accordingly.

Critical strengths

Indicating real-world implementation of MindLoop (a feasibility study) with 23 participants and using wearable BCI to capture emotional states

Critical weaknesses
  • Experiment 1: qwen is used for everything; simulating users, generating counseling responses, and judging how good those responses are. This could create a bias where the RL policy learns to produce responses that qwen itself prefers. the policy is , therefore, optimized to perform well on the same type of criteria for success.

  • Experiment 2: authors didn't show any comparison with a control group where users could use a regular LLM; Experiment 2 results don't indicate that MindLoop is more effective than an ordinary LLM or LLM without the proposed policy; same with adding EEG/BCI; so combining the two experiments makes the paper overwhelming; the scientific contribution of each should be analyzed more rigorously.

  • Move the related work section to the main text of the paper. It is important for readers to know the position of the paper relative to the literature.

Questions for authors

How do the authors ensure that the policy isn't exploiting biases in the simulated environment and LLM judges (100% success rate in simulation?)? With qwen3-30B as the user simulator, counselor, and LLM judge create a risk of reward hacking does RL policy learn to maximize rewards by learning emotion regulation?

9 comments
AuthorsOpenReview ↗

Response to Reviewer L5dj [1/6]

We sincerely thank the reviewer for the careful reading and for recognizing the feasibility study using wearable BCI during human interactions. We also appreciate the reviewer’s concerns about Qwen-specific reward bias and the lack of a standard LLM control in the human study, and the suggestion to move Related Work into the main text. We report independent CBT ratings and cross-model and human validation, explain the reward design and checkpoint results, discuss standard LLM baselines in simulation and the human no-feedback comparison, and move Related Work into the main text.

W1: Experiment 1: qwen is used for everything; simulating users, generating counseling responses, and judging how good those responses are. This could create a bias where the RL policy learns to produce responses that qwen itself prefers. the policy is , therefore, optimized to perform well on the same type of criteria for success.

Q1: How do the authors ensure that the policy is not exploiting biases in the simulated environment and LLM judges? Does the 100% simulation success rate reflect effective emotion regulation or reward hacking?

A1: Part 1. Independent CBT Quality Assessment Beyond Training Rewards

Thank you for raising this important concern. A 100% SR indicates procedural completion but does not necessarily reflect effective emotion regulation or CBT intervention quality. To investigate whether the learned policy merely exploits Qwen-specific reward preferences, we introduce independent CBT quality assessments, cross-model evaluations, and training-Judge replacement experiments.

To assess intervention quality without relying on simulator-generated affect labels, we invited a CBT expert to evaluate all 29 real-world dialogues (18 with physiological feedback and 11 without feedback) and 30 sampled simulated dialogues (15 each from ESConv and ExTES), using the original 12-item Cognitive Therapy Scale–Revised (CTS-R). In parallel, GPT-6-luna evaluated the same dialogues using an expert-designed CTS-R rubric, enabling comparison with professional assessments.

Trace SourceNExpert CTS-R (/6)LLM CTS-R (/6)
Real-world Feedback182.43 ± 0.322.46 ± 0.31
Real-world No-Feedback112.18 ± 0.402.38 ± 0.43
ESConv MindLoop152.64 ± 0.362.73 ± 0.34
ExTES MindLoop152.66 ± 0.332.81 ± 0.28
Overall592.50 ± 0.382.60 ± 0.37

Across the 59 dialogues, the LLM ratings showed strong agreement with expert assessments (Spearman ρ = 0.802, Pearson r = 0.813, ICC(A,1) = 0.784, MAE = 0.196). Based on this agreement, we further applied the expert-guided rubric to evaluate CBT intervention quality across the complete simulated test sets, independently of the simulator-generated affect labels.

DatasetMethodSR ↑EI ↑CTS-R (/6) ↑
ESConv (N=130)GLM-5.1 HP0.97770.492.63 ± 0.34
DeepSeek-V4-Flash HP0.80062.382.06 ± 0.37
MindLoop1.00078.352.83 ± 0.33
ExTES (N=200)GLM-5.1 HP0.99578.202.63 ± 0.26
DeepSeek-V4-Flash HP0.85075.862.22 ± 0.36
MindLoop1.00092.222.91 ± 0.28

MindLoop consistently achieves higher CTS-R scores than both GLM-5.1 HP and DeepSeek-V4-Flash HP on ESConv (2.83 vs. 2.63 and 2.06) and ExTES (2.91 vs. 2.63 and 2.22). Importantly, these scores are assigned by GPT-6-luna using an expert-designed rubric, independently of the Qwen-based training Judge and simulator-generated affect labels. The improvements therefore provide complementary evidence that the learned policy achieves better CBT-aligned intervention quality beyond optimizing its original training rewards.

AuthorsOpenReview ↗

Response to Reviewer L5dj [2/6]

A1 (continued): Part 2. Cross-Model Robustness and Real-World Validation

We separately tested whether the learned policy transfers beyond its original Qwen-based environment. Simulator and Executor replacement. We held the Planner fixed and replaced the Qwen3-30B User Simulator or Executor individually with DeepSeek-V4-Flash or GLM-5.1, and jointly with DeepSeek-V4.1-Flash-1017 and GLM-5.2. The joint tests used newer models because the original services were no longer available to us.

DatasetUser SimulatorExecutorAS ↑SR ↑AT ↓EI ↑PR ↑CTS-R-12 ↑
ESConvQwen3-30BQwen3-30B5.00/0.001.000/0.00017.25/0.5978.35/28.120.9846/0.122.83/0.33
ESConvDeepSeek-V4-FlashQwen3-30B5.00/0.001.000/0.00017.18/0.3863.80/10.901.000/0.0003.18/0.32
ESConvGLM-5.1Qwen3-30B5.00/0.001.000/0.00017.25/0.4466.05/15.110.969/0.1733.02/0.35
ESConvQwen3-30BGLM-5.15.00/0.001.000/0.00017.17/0.3885.22/20.651.000/0.0003.37/0.29
ESConvQwen3-30BDeepSeek-V4-Flash5.00/0.001.000/0.00017.23/0.4787.24/18.130.992/0.0883.16/0.35
ESConvDeepSeek-V4.1-FlashGLM-5.25.00/0.001.000/0.00017.38/0.6342.59/7.631.000/0.0003.68/0.28
ESConvGLM-5.2DeepSeek-V4.1-Flash5.00/0.001.000/0.00017.33/0.4960.52/16.150.908/0.2913.96/0.32
ExTESQwen3-30BQwen3-30B5.00/0.001.000/0.00017.17/0.3992.22/12.751.000/0.0002.91/0.28
ExTESDeepSeek-V4-FlashQwen3-30B5.00/0.001.000/0.00017.23/0.4363.63/9.881.000/0.0003.27/0.26
ExTESGLM-5.1Qwen3-30B5.00/0.001.000/0.00017.39/0.4967.85/10.571.000/0.0003.22/0.28
ExTESQwen3-30BGLM-5.15.00/0.001.000/0.00017.20/0.4591.64/14.590.995/0.0713.44/0.28
ExTESQwen3-30BDeepSeek-V4-Flash5.00/0.001.000/0.00017.23/0.4595.33/9.051.000/0.0003.30/0.30
ExTESDeepSeek-V4.1-FlashGLM-5.25.00/0.001.000/0.00017.22/0.5169.10/8.411.000/0.0003.73/0.26
ExTESGLM-5.2DeepSeek-V4.1-Flash5.00/0.001.000/0.00017.25/0.5973.15/11.030.990/0.1004.01/0.29

All configurations maintained 100% SR and positive mean simulated EI, despite replacing the Qwen-based User Simulator and/or Executor without retraining the Planner. With the Qwen Simulator fixed, replacing the Executor improved CTS-R in all four comparisons and EI in three. Joint replacement also yielded higher CTS-R scores, although EI decreased relative to the original Qwen-based environment. These results suggest that the learned Planner is not exclusively dependent on the original Qwen model family to produce coherent CBT interventions. The variation in EI across simulators further highlights the importance of evaluating intervention quality independently of simulator-generated affective outcomes.

Training Judge replacement. We retrained the Planner with DeepSeek-V4.1-Flash-1017 as the training Judge and evaluated it on the same test sets.

DatasetTraining JudgeAS ↑SR ↑AT ↓EI ↑PR ↑CTS-R-12 ↑
ESConvQwen3-30B5.00/0.001.000/0.00017.25/0.5978.35/28.120.9846/0.122.83/0.33
ESConvDeepSeek-V4.1-Flash5.00/0.001.000/0.00017.71/1.9379.05/19.501.000/0.0002.78/0.29
ExTESQwen3-30B5.00/0.001.000/0.00017.17/0.3992.22/12.751.000/0.0002.91/0.28
ExTESDeepSeek-V4.1-Flash5.00/0.001.000/0.00017.44/2.0190.59/12.431.000/0.0002.95/0.24

Both policies achieved 100% SR on ESConv and ExTES, with comparable EI and independently evaluated CTS-R scores. Notably, replacing the Qwen3-30B training Judge with DeepSeek-V4.1-Flash yielded similar CTS-R scores on ESConv (2.78 vs. 2.83) and ExTES (2.95 vs. 2.91). These results directly address the concern that the learned policy may exploit the scoring preferences of the original Qwen Judge, showing that comparable CBT intervention quality can be obtained when process rewards are provided by a different model family.

Real-world validation. We deployed the simulation-trained Planner without updating it. In the 18-participant physiological-feedback group, the wearable EEG/fNIRS system supported full stage coverage (AS = 5.00, SR = 1.00). Positive PANAS increased by 2.89 and negative PANAS decreased by 2.78 within sessions (both p < 0.05). An independent CBT expert rated these 18 transcripts 2.43 ± 0.32 (/6). These results support the feasibility of deploying the learned policy with real users.

AuthorsOpenReview ↗

Response to Reviewer L5dj [3/6]

A1 (continued): Part 3. Policy Design and Training Dynamics Analysis

MindLoop incorporates three design choices that reduce opportunities for reward exploitation:

  • Planner–Executor Separation: Only the Planner is optimized during RL, while the Executor remains frozen. The Planner selects actions from a constrained CBT stage–strategy space rather than generating arbitrary counseling responses, limiting direct optimization of Judge-preferred textual patterns.
  • Structured Process Rewards: Rather than relying on a single holistic LLM preference score, MindLoop combines stage-completion, strategy-appropriateness, and affective rewards. Each stage has three explicit completion criteria evaluated through binary judgments, while strategy appropriateness is scored on a discrete 0/1 scale.
  • Affective Feedback Rewards: Affective rewards are derived from segment-level post-affect observations rather than directly from the training Judge's quality assessments, providing a separate optimization signal.

Policy Checkpoint Analysis. To further examine whether policy optimization improves regulation performance rather than merely increasing training rewards, we evaluate six Planner checkpoints throughout RL training on the ESConv test set.

CheckpointTraining Reward ↑SR (%) ↑AT ↓EI ↑CTS-R (/6) ↑
u58.77 / 6.1659.23/49.3313.56/1.3234.20/25.402.58 / 0.30
u109.16 / 6.8872.31/44.9214.02/1.7740.62/32.632.58 / 0.28
u1511.00 / 6.2383.85/36.9515.37/1.3254.19/35.762.64/0.30
u2011.66 / 6.5593.85/24.1216.14/1.2559.66/31.942.69/0.26
u2513.00 / 6.61100.00/0.0017.27/0.8378.18/26.652.74/0.28
u3013.31 / 6.87100.00/0.0017.25/0.5978.35/28.122.83 ± 0.33

Across training checkpoints, SR increases from 59.23% at u5 to 100% at u25, while EI improves from 34.20 to 78.35 by u30. Since the same stage-transition constraints are applied throughout training, the increase in SR indicates that complete procedural success is not mechanically guaranteed by the constrained action space. The simultaneous improvement in EI further demonstrates that policy optimization improves simulated affective outcomes alongside stage completion.

Importantly, mean training reward increases from 8.77 to 13.31, accompanied by an improvement in independently evaluated CTS-R from 2.58 to 2.83. CTS-R scores are assigned by GPT-6-luna using the expert-designed rubric, independently of the Qwen-based training Judge. Notably, after SR reaches 100% at u25, CTS-R continues to improve from 2.74 to 2.83 at u30, while EI remains relatively stable. This indicates that additional training progress is accompanied by improvements in independently assessed CBT quality even after procedural success has saturated. Together with the cross-model and training-Judge replacement experiments, these results provide converging evidence against the observed improvements being driven solely by exploitation of Qwen-specific reward signals.

We thank the reviewer for these constructive suggestions. We have made the corresponding revisions to the manuscript and included the additional experiments in the Appendix.

We thank the reviewer for these constructive suggestions. We have made the corresponding revisions to the manuscript and included the additional experiments in the Appendix.

Reviewer_Hbs4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MindLoop, which combines a CBT-guided stage-strategy planner, dual-timescale reinforcement learning, and wearable EEG/fNIRS affect feedback for multi-turn emotion regulation. The planner is trained in simulation and deployed without further policy updates. Experiments on ESConv- and ExTES-based simulations report improved process completion and simulated affective outcomes. The participant study demonstrates end-to-end feasibility and favorable within-session self-reported changes.

Critical strengths
  1. Separating dialogue-level stage decisions from within-stage strategy decisions is well motivated. The ablations provide useful evidence that stage-specific heads and differentiated credit-assignment horizons improve simulated outcomes.

  2. The work goes beyond simulation by implementing subject-specific physiological decoding during human interaction. Cross-model simulator/executor tests and cross-block decoder evaluation also strengthen the feasibility assessment.

Critical weaknesses

The simulation metrics provide limited evidence of effective emotion regulation. Success primarily means reaching the final stage and terminating legally, while EI depends on simulator-generated affect. Cross-model tests are useful, but do not independently validate these outcomes. Moreover, training uses one seed, and rollout-level standard deviations do not establish the stability of the learned improvements.

Questions for authors
  1. Can a controlled comparison establish whether physiological feedback improves user outcomes over the same planner using text alone or self-reported affect?

  2. How does MindLoop compare with a fixed stage schedule using the same executor, intervention inventory, and dialogue budget? This would help isolate adaptive planning from structured progression.

7 comments
AuthorsOpenReview ↗

Response to Reviewer Hbs4 [1/4]

We sincerely thank the reviewer for the careful reading and for highlighting subject-specific physiological decoding during human interaction and the cross-block decoder evaluation. We also appreciate the reviewer’s concerns about the limits of simulation metrics and the stability of single-seed results, as well as the questions about physiological feedback and fixed-stage scheduling. The revision adds expert-referenced CTS-R assessments, results from five independent training runs, feedback controls in simulation and with human participants, and fixed-schedule comparisons.

Q1: The simulation metrics provide limited evidence of effective emotion regulation. Success primarily means reaching the final stage and terminating legally, while EI depends on simulator-generated affect. Cross-model tests are useful, but do not independently validate these outcomes.

A1: Thank you for identifying the need for evaluation beyond simulator-generated affect. We recognize that SR measures procedural completion, while EI reflects simulator-generated affective changes. We therefore conducted additional CTS-R intervention quality assessments.

To assess intervention quality without relying on simulator-generated affect labels, we invited a CBT expert to evaluate all 29 real-world dialogues (18 with physiological feedback and 11 without feedback) and 30 sampled simulated dialogues (15 each from ESConv and ExTES), using the original 12-item Cognitive Therapy Scale–Revised (CTS-R). In parallel, GPT-6-luna evaluated the same dialogues using an expert-designed CTS-R rubric, enabling comparison with professional assessments.

Trace SourceNExpert CTS-R (/6)LLM CTS-R (/6)
Real-world Feedback182.43 ± 0.322.46 ± 0.31
Real-world No-Feedback112.18 ± 0.402.38 ± 0.43
ESConv MindLoop152.64 ± 0.362.73 ± 0.34
ExTES MindLoop152.66 ± 0.332.81 ± 0.28
Overall592.50 ± 0.382.60 ± 0.37

Expert and LLM assessments show consistent scoring patterns across both real-world and simulated dialogues. In the human study, the physiological-feedback group achieved higher expert CTS-R scores than the No-Feedback group (2.43 vs. 2.18). MindLoop also achieved consistent expert scores on ESConv (2.64) and ExTES (2.66), with closely aligned LLM assessments. These findings provide complementary evidence of CBT quality across different interaction settings.

To further assess the reliability of LLM-based CTS-R evaluation, we conducted an agreement analysis between expert and LLM rating:

Trace SourceNSpearman ρ ↑Pearson r ↑MAE ↓ICC(A,1) ↑ [95% CI]
Real-world Feedback180.7920.8410.1300.842 [0.638, 0.936]
Real-world No-Feedback110.8250.7670.2880.701 [0.353, 0.828]
ESConv MindLoop150.7920.7940.1830.782 [0.550, 0.900]
ExTES MindLoop150.7270.6880.2220.622 [0.291, 0.746]
Overall590.8020.8130.1960.784 [0.680, 0.854]

Across the 59 dialogues, the LLM ratings showed strong agreement with expert assessments (Spearman ρ = 0.802, Pearson r = 0.813, ICC(A,1) = 0.784, MAE = 0.196). Based on this agreement, we further applied the expert-guided rubric to evaluate CBT intervention quality across the complete simulated test sets, independently of the simulator-generated affect labels.

Building on the expert–LLM agreement established above, we further employed the expert-guided GPT-6-luna Judge, validated against expert CTS-R ratings, to evaluate MindLoop and two alternative LLM-based methods under the adaptive-planning setting on the ESConv and ExTES test sets. All methods were assessed using the same expert-designed CTS-R rubric.

DatasetMethodSR ↑EI ↑CTS-R (/6) ↑
ESConvGLM-5.1 HP0.97770.492.63 ± 0.34
DeepSeek-V4-Flash HP0.80062.382.06 ± 0.37
MindLoop1.00078.352.83 ± 0.33
ExTESGLM-5.1 HP0.99578.202.63 ± 0.26
DeepSeek-V4-Flash HP0.85075.862.22 ± 0.36
MindLoop1.00092.222.91 ± 0.28

MindLoop consistently achieves higher CTS-R scores than both GLM-5.1 HP and DeepSeek-V4-Flash HP. On ESConv, MindLoop scores 2.83, compared with 2.63 and 2.06 for the two baselines, respectively. The same trend holds on ExTES (2.91 vs. 2.63 and 2.22). These improvements complement the higher simulated EI and SR results and provide additional evidence of CBT intervention quality based on dialogue content rather than simulator-generated affect.

We thank the reviewer for the valuable feedback and have revised the manuscript accordingly.

AuthorsOpenReview ↗

Response to Reviewer Hbs4 [2/4]

Q2: Moreover, training uses one seed, and rollout-level standard deviations do not establish the stability of the learned improvements.

A2: Thank you for highlighting the need to assess stability across independent training runs. We recognize that rollout-level standard deviations do not capture variation across independently trained policies. To assess the reproducibility of the learned improvements, we conducted four additional independent training runs, together with the original seed 42, yielding five independently trained policies evaluated on the same test sets.

DatasetSeedSR ↑AT ↓EI ↑PR ↑
ESConv421.0017.2578.350.9846
34071.0017.3277.570.9923
71.0018.1277.860.9692
3021.0017.0978.571.0000
5221.0017.1877.880.9923
ESConvMean ± SD1.00 ± 0.0017.39 ± 0.4278.05 ± 0.400.9877 ± 0.0117
ExTES421.0017.1792.221.0000
34071.0016.9088.981.0000
71.0018.1690.641.0000
3021.0017.1687.621.0000
5221.0016.7188.830.9950
ExTESMean ± SD1.00 ± 0.0017.22 ± 0.5689.66 ± 1.790.9990 ± 0.0022

All five independently trained policies achieve 100% SR on both datasets, with between-seed EI standard deviations of only 0.40 on ESConv and 1.79 on ExTES. Moreover, every training seed consistently outperforms the strongest prompt-based baseline, GLM-5.1 HP, in simulated EI on both datasets (ESConv: 77.57–78.57 vs. 70.49; ExTES: 87.62–92.22 vs. 78.20). These results demonstrate that the simulated improvements are reproducible across different training seeds rather than being specific to a single run.

We sincerely thank the reviewer for this valuable suggestion and have incorporated the five-seed training results into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Hbs4 [3/4]

Q3: Can a controlled comparison establish whether physiological feedback improves user outcomes over the same planner using text alone or self-reported affect?

A3: Thank you for suggesting a controlled comparison to isolate the contribution of physiological feedback. To determine whether physiological feedback improves complete interaction outcomes rather than merely changing planner actions, we conducted an affect-masking ablation in simulation and a prospective no-feedback comparison with human participants. In both settings, the same Planner and Executor were used, while explicit affective inputs were masked in the control condition, allowing the system to rely solely on textual information for intervention planning and response generation.

1. Affect-Masking Ablation in Simulation

We evaluated MindLoop on ESConv and ExTES by masking affective inputs to the Planner and Executor while keeping all other components unchanged.

DatasetMethodSR ↑AT ↓EI ↑PR ↑
ESConvMindLoop1.0017.2578.350.9846
ESConvNo-Affect1.0019.7575.020.977
ExTESMindLoop1.0017.1792.221.000
ExTESNo-Affect1.0019.5391.331.000

Masking affective inputs increased AT by 2.50 and 2.36 turns on ESConv and ExTES, respectively, while EI decreased by 3.33 and 0.89 points. SR remained at 1.00 across both conditions. These results demonstrate that affective feedback improves interaction efficiency and contributes to emotional improvement in simulation.

2. Prospective No-Feedback Comparison in Human Interaction

We additionally recruited 11 participants for a No-Feedback control group, compared with the original physiological-feedback group. Both groups used the same MindLoop Planner and Executor, wore identical EEG/fNIRS devices, and followed the same experimental protocol. Participants were blinded to their feedback conditions. The only difference was that decoded physiological affective states were withheld from the Planner and Executor in the No-Feedback group.

MetricPhys (n=18)No-Feedback (n=11)p
AS ↑5.00 ± 0.004.91 ± 0.300.3793
SR ↑1.00 ± 0.000.91 ± 0.300.3793
AT ↓18.22 ± 3.3921.80 ± 3.010.0069
EI ↑49.06 ± 36.5639.61 ± 48.370.5840
PR ↑0.78 ± 0.430.82 ± 0.401.0000
ΔPANAS Positive ↑+2.89 ± 4.59+2.27 ± 8.400.8263
ΔPANAS Negative ↓−2.78 ± 5.45−2.27 ± 5.620.8147
Expert-rated CTS-R (/6) ↑2.43 ± 0.322.18 ± 0.400.1033
Safety (/6) ↑5.39 ± 1.145.27 ± 1.010.7777

The Phys group required 3.58 fewer interaction turns on average, corresponding to a significant 16.4% reduction in AT (p = 0.0069), while achieving a 9.45-point higher mean EI. The Phys group also received higher expert-rated CTS-R scores (2.43 vs. 2.18), indicating more favorable expert assessments of CBT intervention quality. Although between-group PANAS differences were not statistically significant, both positive and negative PANAS changes were significant within the Phys group (p < 0.05), whereas neither reached significance in the No-Feedback group (Positive: p = 0.3908; Negative: p = 0.2097).

Further analysis of affective trajectories in Figure 2a shows similar stage-wise patterns across groups, with higher Stage 5 emotion improvement in the Phys group (55.39 vs. 45.00). Physiological analyses in Figures 2b–c further show significant early-to-late changes in four EEG bands in the Phys group, compared with none in the No-Feedback group. Significant changes were also observed in three fNIRS measures in the Phys group and two in the No-Feedback group. Together, these findings demonstrate improved interaction efficiency with physiological feedback, accompanied by favorable trends in affective outcomes and expert-assessed CBT quality. The physiological observations provide additional evidence of within-session changes during real-world interactions.

We thank the reviewer again for this valuable suggestion. We have revised the corresponding sections of the manuscript accordingly.

Reviewer_jFpuOpenReview ↗
Rating
2(−1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MindLoop combines a five-stage CBT-guided planner, separate stage and strategy policies with different credit-assignment horizons, and wearable EEG/fNIRS affect feedback. The planner is trained in simulation and deployed without further updates. Evaluation includes simulated dialogues, cross-model tests, and a 23-participant feasibility study.

Critical strengths
  1. Separating dialogue-level progression from within-stage strategy learning is well motivated. The ablations support this design: removing dual-timescale optimization reduces simulated EI from 78.35 to 50.73 on ESConv.

  2. Evaluation extends beyond a single simulator. Cross-model simulator/executor tests and an end-to-end human study provide useful evidence of robustness and deployment feasibility. The human results are appropriately presented as preliminary.

Critical weaknesses
  1. The additional value of physiological feedback remains unclear. The human study has no no-feedback or shuffled-feedback comparison. Appendix D.6 measures action changes under fixed-history replay, which shows sensitivity to feedback but does not establish improved regulation outcomes.

  2. The CTSR-4 assessment needs clarification. Section 3.2 says participants completed it, while the cited validation study used trained psychotherapy raters. The validity of interpreting participant ratings as evidence of CBT competence is therefore uncertain.

Questions for authors
  1. Is there any prospective outcome comparison between physiological feedback and the same planner without it? This would clarify the benefit beyond the replay analysis.

  2. Who rated CTSR-4, what instructions were provided, and was a participant-rated adaptation validated?

  3. Does SR require satisfactory completion of the stage checklists, or only reaching Stage 5 and terminating legally? A fixed-schedule control would help distinguish adaptive planning from stage-order compliance.

4 comments
AuthorsOpenReview ↗

Response to Reviewer jFpu [1/3]

Response to Reviewer jFpu [1/3]

We sincerely thank the reviewer for the careful reading and for recognizing the value of the cross-model evaluation and preliminary human feasibility study. We also appreciate the reviewer’s questions about the added value of physiological feedback, the validity of participant-rated CTSR-4, and whether the success criterion requires checklist completion. We added a simulation ablation, a human no-feedback comparison, and expert CTS-R ratings, clarified the success criterion, and compared adaptive stage timing with fixed schedules.

W1: The additional value of physiological feedback remains unclear. The human study has no no-feedback or shuffled-feedback comparison. Appendix D.6 measures action changes under fixed-history replay, which shows sensitivity to feedback but does not establish improved regulation outcomes. Is there any prospective outcome comparison between physiological feedback and the same planner without it?

Q1: Is there any prospective outcome comparison between physiological feedback and the same planner without it? This would clarify the benefit beyond the replay analysis.

A1: We appreciate the reviewer’s emphasis on assessing how physiological feedback affects outcomes across a full interaction. To address this question, we conducted an affect-masking ablation in simulation and a prospective no-feedback comparison with human participants.

1. Affect-Masking Ablation in Simulation

We evaluated MindLoop on ESConv and ExTES by masking affective inputs to the Planner and Executor while keeping all other components unchanged.

DatasetMethodSR ↑AT ↓EI ↑PR ↑
ESConvMindLoop1.0017.2578.350.9846
ESConvNo-Affect1.0019.7575.020.977
ExTESMindLoop1.0017.1792.221.000
ExTESNo-Affect1.0019.5391.331.000

Masking affective inputs increased AT by 2.50 and 2.36 turns on ESConv and ExTES, respectively, while EI decreased by 3.33 and 0.89 points. SR remained at 1.00 across both conditions. These results demonstrate that affective feedback improves interaction efficiency and contributes to emotional improvement in simulation.

2. Prospective No-Feedback Comparison in Human Interaction

We additionally recruited 11 participants for a No-Feedback control group, compared with the original physiological-feedback group. Both groups used the same MindLoop Planner and Executor, wore identical EEG/fNIRS devices, and followed the same experimental protocol. Participants were blinded to their feedback conditions. The only difference was that decoded physiological affective states were withheld from the Planner and Executor in the No-Feedback group.

MetricPhys (n=18)No-Feedback (n=11)p
AS ↑5.00 ± 0.004.91 ± 0.300.3793
SR ↑1.00 ± 0.000.91 ± 0.300.3793
AT ↓18.22 ± 3.3921.80 ± 3.010.0069
EI ↑49.06 ± 36.5639.61 ± 48.370.5840
PR ↑0.78 ± 0.430.82 ± 0.401.0000
ΔPANAS Positive ↑+2.89 ± 4.59+2.27 ± 8.400.8263
ΔPANAS Negative ↓−2.78 ± 5.45−2.27 ± 5.620.8147
Expert-rated CTS-R (/6) ↑2.43 ± 0.322.18 ± 0.400.1033
Safety (/6) ↑5.39 ± 1.145.27 ± 1.010.7777

The Phys group required 3.58 fewer interaction turns on average, corresponding to a significant 16.4% reduction in AT (p = 0.0069), while achieving a 9.45-point higher mean EI. The Phys group also received higher expert-rated CTS-R scores (2.43 vs. 2.18), indicating more favorable expert assessments of CBT intervention quality. Although between-group PANAS differences were not statistically significant, both positive and negative PANAS changes were significant within the Phys group (p < 0.05), whereas neither reached significance in the No-Feedback group (Positive: p = 0.3908; Negative: p = 0.2097).

Further analysis of affective trajectories in Figure 2a shows similar stage-wise patterns across groups, with higher Stage 5 emotion improvement in the Phys group (55.39 vs. 45.00). Physiological analyses in Figures 2b–c further show significant early-to-late changes in four EEG bands in the Phys group, compared with none in the No-Feedback group. Significant changes were also observed in three fNIRS measures in the Phys group and two in the No-Feedback group. Together, these findings demonstrate improved interaction efficiency with physiological feedback, accompanied by favorable trends in affective outcomes and expert-assessed CBT quality. The physiological observations provide additional evidence of within-session changes during real-world interactions.

We thank the reviewer again for this valuable suggestion. We have revised the corresponding sections of the manuscript accordingly.

AuthorsOpenReview ↗

Response to Reviewer jFpu [2/3]

W2: The CTSR-4 assessment needs clarification. Section 3.2 says participants completed it, while the cited validation study used trained psychotherapy raters. The validity of interpreting participant ratings as evidence of CBT competence is therefore uncertain. Who rated CTSR-4, what instructions were provided, and was a participant-rated adaptation validated?

Q2: Who rated CTSR-4, what instructions were provided, and was a participant-rated adaptation validated?

A2: Thank you for highlighting the importance of professional CBT competence assessment. In our original study, participants completed CTSR-4 using detailed scoring guidelines and item-level instructions. As CTSR-4 is intended for trained raters, the participant-rated adaptation was not independently validated for assessing professional CBT competence.

To provide a more rigorous evaluation, we introduced expert-rated CTS-R assessment and expert-guided LLM evaluation. A CBT expert evaluated all 29 real-world dialogues and 15 sampled dialogues from each of ESConv and ExTES using the standard 12-item Cognitive Therapy Scale–Revised (CTS-R), with each item rated from 0 to 6. Independently, GPT-6-luna evaluated the same dialogues across all 12 items using an expert-designed CTS-R scoring rubric.

1. Expert and LLM CTS-R Scores

Trace SourceNExpert CTS-R (/6)LLM CTS-R (/6)
Real-world Feedback182.43 ± 0.322.46 ± 0.31
Real-world No-Feedback112.18 ± 0.402.38 ± 0.43
ESConv MindLoop152.64 ± 0.362.73 ± 0.34
ExTES MindLoop152.66 ± 0.332.81 ± 0.28
Overall592.50 ± 0.382.60 ± 0.37

Expert and LLM assessments show consistent scoring patterns across both real-world and simulated dialogues. In the human study, the physiological-feedback group achieved higher expert CTS-R scores than the No-Feedback group (2.43 vs. 2.18). MindLoop also achieved consistent expert scores on ESConv (2.64) and ExTES (2.66), with closely aligned LLM assessments. These findings provide complementary evidence of CBT quality across different interaction settings.

2. Agreement Between LLM and Expert Ratings

Trace SourceNSpearman ρ ↑Pearson r ↑MAE ↓ICC(A,1) ↑ [95% CI]
Real-world Feedback180.7920.8410.1300.842 [0.638, 0.936]
Real-world No-Feedback110.8250.7670.2880.701 [0.353, 0.828]
ESConv MindLoop150.7920.7940.1830.782 [0.550, 0.900]
ExTES MindLoop150.7270.6880.2220.622 [0.291, 0.746]
Overall590.8020.8130.1960.784 [0.680, 0.854]

The LLM ratings demonstrate strong correlations and good agreement with expert assessments (Spearman ρ = 0.802, Pearson r = 0.813, ICC(A,1) = 0.784), with an MAE of 0.196 on the 0–6 scale. These results support expert-guided LLM evaluation as a reliable supplementary approach for scalable CBT quality assessment.

Building on the expert–LLM agreement established above, we further employed the expert-guided GPT-6-luna Judge, validated against expert CTS-R ratings, to evaluate MindLoop and two alternative LLM-based methods under the adaptive-planning setting on the ESConv and ExTES test sets. All methods were assessed using the same expert-designed CTS-R rubric.

DatasetMethodSR ↑EI ↑CTS-R (/6) ↑
ESConv (N=130)GLM-5.1 HP0.97770.492.63 ± 0.34
DeepSeek-V4-Flash HP0.80062.382.06 ± 0.37
MindLoop1.00078.352.83 ± 0.33
ExTES (N=200)GLM-5.1 HP0.99578.202.63 ± 0.26
DeepSeek-V4-Flash HP0.85075.862.22 ± 0.36
MindLoop1.000**92.222.91 ± 0.28

MindLoop achieved the highest CTS-R scores on both datasets, outperforming GLM-5.1 HP by 0.20 and 0.28 points, and DeepSeek-V4-Flash HP by 0.77 and 0.69 points on ESConv and ExTES, respectively. These results provide complementary evidence of improved CBT intervention quality and adherence to CBT principles, independent of simulator-generated affective outcomes.

We sincerely thank the reviewer for the constructive suggestions, which have helped us strengthen the evaluation of CBT competence. We have made the corresponding revisions to the manuscript and included the additional CTS-R evaluation experiments in the Appendix.

AuthorsOpenReview ↗

Response to Reviewer jFpu [3/3]

Q3: Does SR require satisfactory completion of the stage checklists, or only reaching Stage 5 and terminating legally? A fixed-schedule control would help distinguish adaptive planning from stage-order compliance.

A3: Thank you for highlighting the distinction between procedural success and intervention quality, and for suggesting a fixed-stage comparison. We clarify the definition of SR and introduce five fixed-schedule baselines to evaluate the contribution of adaptive stage planning.

1. Clarification of the Success Criterion

SR measures whether the system successfully reaches Stage 5 and terminates legally within the dialogue budget, without requiring all stage checklists to be satisfied. Since SR reflects procedural completion rather than the quality of CBT interventions, we additionally introduce an expert-guided LLM CTS-R Judge to independently assess CBT competence. Specifically, GPT-6-luna applies the expert-designed 12-item CTS-R rubric, with its scoring agreement against professional expert ratings validated in our response to Q2. This provides a complementary measure of intervention quality beyond procedural success and simulator-generated affective outcomes.

2. Fixed-Stage Scheduling Comparison

We designed five fixed-stage scheduling baselines (F01–F05), with predetermined stage allocations totaling 17 or 18 turns, comparable to MindLoop's average interaction length. All methods shared the same five-stage CBT structure, strategy selection mechanism, Executor, and User Simulator. The key distinction was that fixed schedules followed predetermined stage transitions, whereas MindLoop dynamically adjusted stage transitions according to the dialogue state.

ESConv

MethodS1/S2/S3/S4/S5AT ↓EI ↑PR ↑CTS-R (/6) ↑
Fixed F014/3/3/4/317.0076.50 ± 23.050.98 ± 0.122.76 ± 0.32
Fixed F024/4/3/4/318.0070.54 ± 29.420.96 ± 0.192.65 ± 0.33
Fixed F033/3/4/4/317.0079.60 ± 19.230.99 ± 0.092.59 ± 0.29
Fixed F043/3/4/4/418.0074.40 ± 17.970.98 ± 0.142.61 ± 0.35
Fixed F053/3/3/4/417.0070.89 ± 27.900.98 ± 0.122.45 ± 0.33
Fixed Average—17.4074.39 ± 3.830.98 ± 0.012.61 ± 0.11
MindLoopAdaptive17.2578.35 ± 28.120.9846 ± 0.122.83 ± 0.33

ExTES

MethodS1/S2/S3/S4/S5AT ↓EI ↑PR ↑CTS-R (/6) ↑
Fixed F014/3/3/4/317.0086.26 ± 13.921.00 ± 0.002.84 ± 0.28
Fixed F024/4/3/4/318.0086.12 ± 15.091.00 ± 0.002.75 ± 0.28
Fixed F033/3/4/4/317.0089.38 ± 11.371.00 ± 0.002.68 ± 0.25
Fixed F043/3/4/4/418.0090.77 ± 10.041.00 ± 0.002.70 ± 0.30
Fixed F053/3/3/4/417.0091.49 ± 11.801.00 ± 0.002.54 ± 0.26
Fixed Average—17.4088.80 ± 2.501.00 ± 0.002.70 ± 0.11
MindLoopAdaptive17.1792.22 ± 12.751.00 ± 0.002.91 ± 0.28

MindLoop achieved the highest CTS-R scores on both ESConv (2.83) and ExTES (2.91), outperforming the fixed-schedule averages by 0.22 and 0.21 points, respectively. Meanwhile, MindLoop improved EI by 3.96 points on ESConv and 3.42 points on ExTES compared with the fixed-schedule averages, while maintaining comparable interaction lengths. Notably, although Fixed F03 achieved slightly higher EI on ESConv (79.60 vs. 78.35), its CTS-R was considerably lower (2.59 vs. 2.83). These results demonstrate the value of adaptive stage planning in balancing emotional improvement and CBT intervention quality, beyond merely following a predefined stage sequence.

We thank the reviewer for this constructive suggestion. We have revised the manuscript to clarify the SR criterion and incorporate the fixed-stage comparisons and independent CTS-R evaluation.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer jFpu,

Thank you again for your careful reading and thoughtful feedback. We hope our response and the additional experiments have helped address your questions. As the discussion period draws to a close, we would be happy to answer any remaining questions or provide further clarification.

Best regards,

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response and Revision of the Paper

We sincerely thank the reviewers for their careful reading and constructive feedback. All three reviewers recognize the value of our real-world feasibility study using wearable EEG/fNIRS feedback. The reviewers also acknowledge the motivation for hierarchical stage–strategy planning, the effectiveness of dual-timescale reinforcement learning, and the value of component ablations and cross-model evaluations.

The main concerns involve the contribution of physiological feedback, the validity of emotion regulation and CBT quality evaluation, adaptive planning, and potential reward hacking. We have addressed these through controlled experiments, independent CBT evaluations, and robustness analyses. Below, we summarize the revisions and their locations.

  1. Independent CBT Quality Assessment. In response to Reviewers jFpu, Hbs4, and L5dj, we introduced transcript-based evaluation using the standard 12-item CTS-R. A CBT expert assessed 59 dialogues (29 human and 30 simulated). Mean expert scores were 2.43 and 2.18 for the physiological-feedback and No-Feedback groups, and 2.64 and 2.66 for ESConv and ExTES, respectively. GPT-6-luna evaluated the same transcripts using an expert-designed rubric, yielding Spearman ρ = 0.802 and ICC(A,1) = 0.784. On the full test sets, MindLoop achieved LLM-rated CTS-R scores of 2.83 and 2.91, outperforming both LLM planning baselines. (Appendix C; Lines 1239–1245 and 1263–1305, pp. 23–25.)

  2. Contribution of Physiological Feedback. In response to Reviewers jFpu, Hbs4, and L5dj, we added an affect-masking experiment in simulation. Masking affective inputs increased AT by 2.50 and 2.36 turns on ESConv and ExTES, respectively. (Appendix D.4; Lines 1496–1516, pp. 28–29.) We further recruited 11 additional participants for a prospective No-Feedback control experiment, using the same Planner, Executor, and experimental protocol as the physiological-feedback group. The physiological-feedback group required 3.58 fewer turns on average, and achieved higher mean EI and expert-rated CTS-R scores. Both positive and negative PANAS changes were significant within the physiological-feedback group (p < 0.05), but not within the No-Feedback group. (Section 4.2; Lines 363–369, p. 7; Section 5.2.1; Lines 466–485, p. 9; Appendix B.3; Lines 1137–1160, p. 22; Appendix C; Lines 1246–1261, p. 24.)

  3. Adaptive versus Fixed-Stage Scheduling. In response to Reviewers jFpu and Hbs4, we added five fixed schedules sharing MindLoop's CBT stage order, strategy selection mechanism, Executor, and User Simulator. Compared with the fixed-schedule averages, MindLoop improved EI by 3.96 and 3.42 points and CTS-R by 0.22 and 0.21 points on ESConv and ExTES, respectively. (Appendix D.3; Lines 1463–1494, p. 28.)

  4. Multi-Seed Training Stability. As Reviewer Hbs4 suggested, we evaluated five independently trained policies on ESConv and ExTES. All achieved 100% SR, with between-seed EI standard deviations of 0.40 and 1.79, respectively. (Appendix D.2; Lines 1379–1408, pp. 26–27.)

  5. Generalization Beyond Qwen. To address Reviewer L5dj's concerns about model-specific bias and reward hacking, we extended cross-model evaluations by replacing the User Simulator and Executor while keeping the Planner fixed. (Appendix D.5; Lines 1518–1553, p. 29.) We also retrained the Planner with DeepSeek-V4.1-Flash-1017 as the training Judge, obtaining comparable EI and independent CTS-R scores on both datasets, suggesting limited sensitivity to the training Judge in this comparison. (Appendix D.6; Lines 1555–1576, pp. 29–30.)

  6. Training Dynamics. In response to Reviewer L5dj, we analyzed six policy checkpoints across RL training updates 5–30 on ESConv. From update 5 to 30, mean training reward increased from 8.77 to 13.31, while independently evaluated CTS-R improved from 2.58 to 2.83. Notably, CTS-R continued to increase from 2.74 to 2.83 after SR reached 100% at update 25, suggesting improvements in CBT intervention quality beyond procedural completion. (Appendix D.1; Lines 1311–1316 and 1347–1363, pp. 25–26.)

  7. Planning Efficiency. In response to Reviewer L5dj, we added a latency benchmark on 500 fixed dialogue states. MindLoop achieved approximately 59-fold faster stage–strategy planning than Qwen3-30B HP on both datasets. (Appendix D.7; Lines 1578–1601, p. 30.)

  8. Related Work. As Reviewer L5dj suggested, we moved Related Work into the main text and clarified MindLoop's positioning relative to prior work on LLM-based emotional support and physiological affect sensing. (Section 2; Lines 96–139, pp. 2–3.)

We hope our work contributes to advancing LLM-based emotion regulation and the integration of brain–computer interfaces with intelligent systems for more effective and adaptive emotional support.

We sincerely welcome the reviewers’ further feedback and constructive suggestions to improve our study.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.