MOCC-R1: Reinforcing Reasoning-Response Consistency for Multimodal Counselor Response Generation

Wenjie Zheng; Qiming Xie; Jianfei Yu; Rui Xia

cs.AISep 15, 2026

MOCC-R1: Reinforcing Reasoning-Response Consistency for Multimodal Counselor Response Generation

Authors: Wenjie Zheng, Qiming Xie, Jianfei Yu, Rui Xia

Organizations: School of Artificial Intelligence, Nanjing University of Science & Technology, Nanjing, China · School of Intelligence Science and Technology, Nanjing University, China

Abstract

Multimodal counselor response generation (MCRG) aims to generate an appropriate counselor response from multimodal dialogue histories. Progress is limited by two gaps: first, existing datasets rarely capture sustained, human-recorded counseling interactions conducted by qualified counselors; Second, existing methods do not explicitly optimize consistency between counseling reasoning and the generated response, potentially undermining the reliability of MCRG systems. Thus, we introduce MOCC, a multimodal counseling conversation corpus containing over 200 hours of interactions involving 154 credential-verified counselors. Based on MOCC, we propose MOCC-R1, a two-stage framework for optimizing reasoning-response consistency. Cold-start supervised fine-tuning trains the model to generate a structured trajectory consisting of client-state understanding, a response intent that links a counseling principle to a planned action, and the final response. Reinforcement learning (RL) then rewards grounded plan coherence and plan execution, encouraging the inferred state and plan to be supported by the dialogue context and the response to realize that plan. Experiments demonstrate the effectiveness of the proposed MOCC-R1.

Figures & tables

Figure 1: (a) A generated counseling decision chain can exhibit both a state–plan mismatch and a plan–response mismatch. (b) Reasoning-response inconsistency remains around 40% under both Structured SFT and Outcome-level GRPO.

Dataset	Source	Purpose	Modality	Language	#Dialogue	#Utterance	Avg.Turn	Duration(h)
MEDIC ( 2023 )	Role-play	Empathy	T, A, V	Chinese	771	3,443	2	11.3
StickerConv ( 2024b )	Synthetic	Empathy	T, S	English	12,931	142,093	5	/
EmpathicStories ( 2024 )	Crowdsourcing	Empathy	T, A, V	English	269	5,380	10	53.0
AvaMERG ( 2025 )	Crowdsourcing	Empathy	T, A, V	English	33,048	152,021	3	194.9
MESC ( 2025 )	TV-series	Therapy	T, A, V	English	1,019	28,762	14	18.4
M2CoSC ( 2025a )	Synthetic	Therapy	T, I	English	429	3,432	4	/

Table 1: Comparison of multimodal mental-health support dialogue datasets. T/A/V/S/I denote text/audio/video/sticker/image. All MOCC sessions are conducted by credential-verified counselors. The real-client and simulated-client designations follow source descriptions; simulated clients are portrayed by counselors, counseling graduate students, or conference participants.

Category	Total	Counselor	Client
#Sessions	482	-	-
#Dialogues	3,709	-	-
#Speakers	419	154	265
#Utterances	183,350	80,678	102,672
#Tokens	2,194,743	1,181,135	1,013,608
Avg. utts per dia	49.5	21.8	27.7

Table 2: Statistics of MOCC dataset. “utt” represents utterance, “dia” represents dialogue.

Figure 2: Distribution of client presenting problems in the MOCC dataset.

Figure 3: The overall framework of our proposed MOCC-R1.

Method	BLEU-2	Avg.B	R-L	PPL ( $\downarrow$ )	BERTScore	Dist-Avg.	Diff-IP ( $\downarrow$ )	Diff-EX ( $\downarrow$ )	Diff-ER ( $\downarrow$ )	Safety
General-Purpose MLLMs
GPT-5.1	1.18	1.67	6.81	6.51	45.70	22.43	39.33	243.00	81.15	99.23
GPT-5.5	1.46	2.00	7.90	6.08	47.92	29.40	60.02	158.22	53.48	99.75
Claude-4.5-Sonnet	1.71	2.29	8.17	5.47	48.88	29.88	31.73	238.09	33.84	99.86
Gemini-3-Pro	2.01	2.59	8.43	5.35	47.97	28.80	45.26	182.59	32.02	99.75
Grok-4.3	1.87	2.47	7.71	5.31	47.41	30.27	27.34	238.01	33.42	99.81

Table 3: Comparison of different methods on the MOCC dataset in terms of response quality metrics. Best results are highlighted in bold .

Figure 4: Comparison of structured-output methods on the MOCC dataset in terms of counseling reasoning–response consistency.

Method	BLEU-2	Avg.B	R-L	PPL ( $\downarrow$ )	BERTScore	Dist-Avg.	Diff-IP ( $\downarrow$ )	Diff-EX ( $\downarrow$ )	Diff-ER ( $\downarrow$ )	Safety
MOCC-R1 (Full)	4.79	5.15	14.20	4.24	52.56	37.62	27.61	87.41	13.16	99.76
Structured-Output Variants
Outcome-level GRPO	4.35	4.85	11.58	4.99	50.85	32.07	32.57	95.02	9.90	99.59
Structured SFT	3.68	3.98	9.94	5.33	48.32	39.80	37.64	107.92	10.53	99.65
Response-Only Variants
Response-only GRPO	4.30	4.96	13.86	4.68	51.57	32.63	30.22	90.94	11.47	99.72

Table 4: Ablation studies for MOCC-R1.

Method	Failure rate (%) $\downarrow$
Structured SFT	18.21 [17.55, 18.89]
Outcome-level GRPO	19.77 [19.08, 20.47]
MOCC-R1 (Ours)	6.03 [5.63, 6.46]

Table 5: Failure to realize an input-supported counseling plan. Brackets report Wilson 95% confidence intervals.

Figure 5: High-quality responses unsupported by the generated counseling plan. The components distinguish plans unsupported by the input from responses that do not follow an input-supported plan.

MOCC-R1 vs.	Outcome-level GRPO
MOCC-R1 vs.	Win	Tie	Loss
Fluency	53.4%	28.6%	18.0%
Helpfulness	67.2%	22.5%	10.3%
Consistency	75.7%	12.6%	11.7%

Table 6: Human evaluation.

	Case: Grounding-Plan & Plan-Execution Failures
Inputs	Presenting problem: Mental Health Conditions. Client: At the beginning of the week, I say I will only go out twice, but then they invite me and I usually go. I do not really stick to the plan. Client: At work, I am ready to finish so I can go to happy hour. I also think about wanting a beer at family events. Counselor: Are you having trouble with responsibilities at work? Client: No, my work life has been going pretty well. Counselor: How about any social problems through the alcohol? Client: No, I think my social life is actually great because of it.

Ground-Truth Response	Have you had to give up any important activities because of the alcohol use?
Structured SFT	<think> [Client State Understanding] Client continues denying concerns across multiple areas after earlier saying her job and friendships were fine. Her brief “no” suggests she remains minimally concerned while still acknowledging some impact from drinking on daily functioning. [Response Intent] Principle: facilitate awareness and insight ; Planned Action: Invite reflection on how often he drinks over different periods before moving toward treatment planning. </think> <response> So what do you feel would be helpful if we’re thinking ahead towards kind of getting into therapy? </response> GPT-5.5 Consistency Evaluation: Grounded Plan Coherence: Weak ; Plan Execution: Weak
Outcome-only GRPO	<think> [Client State Understanding] Client reports continued denial across several areas —work function seems intact based on her words alone—and she denies concerns there as part of an overall minimizing stance toward impact from drinking . [Response Intent] Principle: facilitate awareness and insight ; Planned Action: Assess whether medical issues have resulted directly from substance use. </think> <response> Have your health-related conditions occurred due to either smoking ? </response> GPT-5.5 Consistency Evaluation: Grounded Plan Coherence: Weak ; Plan Execution: Weak
MOCC-R1 (Ours)	<think> [Client State Understanding] Client directly confirms there have been no reported social difficulties related to their substance use during discussion up till now. [Response Intent] Principle: support ongoing reality testing ; Planned Action: Reflect back the absence of identified social concerns as confirmed from what they said themselves. </think> <response> And you say yourself that your social situation isn’t causing those kinds of issues either. </response> GPT-5.5 Consistency Evaluation: Grounded Plan Coherence: Full ; Plan Execution: Full

Table 7: The case study on a test case from the MOCC dataset. Red spans localize the generated content associated with GPT-5.5’s score deductions; ochre marks consistency dimensions rated Weak or None .

Panel of Figure 7 — Figure 7: Four-level rubric used by the frozen consistency judge. Given the presenting-problem field, multimodal dialogue context, and a model-generated state–intent–response chain, the judge independently assesses grounded plan coherence and plan execution. The ordinal labels Full, Substantial, Weak, and None are mapped to $1.0$ , $0.6$ , $0.3$ , and $0.0$ , respectively, for reward computation. The chain-level consistency reward is the minimum of the two mapped scores, so neither component can compensate for failure of the other.

Multimodal Generation

Multimodal generation creates coordinated text, images, audio, video, 3D, and other outputs from mixed inputs while preserving meaning and consistency.

See all 244 papers

When Less Is More: An Empirical Study of Minimal Responses in Counseling Dialogues and the Behavior of LLMs
Aug 25, 2026Zhiyang QiClinical Language Model Evaluation Language Model Generation Evaluation
The University of Tokyo
MIThinker: A Plug-and-Play Policy-Optimized Thinker For Motivational Interviewing Counseling
Jun 28, 2026Yizhe Yang, Palakorn Achananuparp, Heyan Huang +2Mental Health Counseling Emotional Support Conversation
Beijing Institute of Technology · Singapore Management University · Australian National University
MASC: A Multi-Agent Self-Calibration Framework with Latent Construct Alignment for Consistent Client Role-Playing in Psychological Counseling
Oct 6, 2026Shixin Peng, Kun Jiang, Jiaxing Zheng +2Large Language Model-Based Role-Play Simulation Multi-Agent LLM Systems
Central China Normal University, Wuhan, China · Shanghai Jiao Tong University, Shanghai, China

Aug 25, 2026J/K move · Enter open · S save

When Less Is More: An Empirical Study of Minimal Responses in Counseling Dialogues and the Behavior of LLMs

Zhiyang Qi

The University of Tokyo

Clinical Language Model Evaluation Language Model Generation Evaluation

Open paper AskPDF arXiv

In psychological counseling, effective support is not always delivered through long, information-rich responses. Minimal responses, such as backchannel cues and concise empathic statements, help convey attentive listening, express empathy, and encourage clients to continue expressing themselves. However, existing counseling dialogue systems and evaluation frameworks often favor explicit, content-rich replies, overlooking the interactional value of brief counselor utterances. This paper presents a systematic cross-lingual analysis of minimal responses across multiple counseling dialogue datasets. We develop a two-stage filtering method based on utterance length and content, followed by contextual verification using a large language model (LLM). Our analysis shows that minimal responses are common in human-collected datasets but substantially underrepresented in LLM-generated ones. We further evaluate current LLMs in manually curated dialogue contexts where human counselors used minimal responses. The results show that strong commercial LLMs are capable of generating minimal responses when explicitly instructed, but still struggle to determine when such responses are appropriate. Counseling-specific models trained on synthetic data perform particularly poorly, tending instead to produce longer and more information-rich responses. Moreover, LLM-based response-quality evaluation may undervalue minimal responses, even when they are interactionally appropriate.

Abstract

Multimodal Generation

Explore similar work

When Less Is More: An Empirical Study of Minimal Responses in Counseling Dialogues and the Behavior of LLMs

MIThinker: A Plug-and-Play Policy-Optimized Thinker For Motivational Interviewing Counseling

MASC: A Multi-Agent Self-Calibration Framework with Latent Construct Alignment for Consistent Client Role-Playing in Psychological Counseling