CounselPlex: A Stressor-Aware Full-Duplex Counseling Model

Yeonju Kim; Bella Godiva; Yong Man Ro

cs.CLOct 23, 2026

CounselPlex: A Stressor-Aware Full-Duplex Counseling Model

Authors: Yeonju Kim, Bella Godiva, Yong Man Ro

Organizations: Integrated Vision and Language Lab, KAIST, South Korea

Abstract

Full-duplex speech models can listen and speak at the same time, enabling natural, low-latency interaction. In counseling, however, these models often focus on the latest utterance and lose track of the seeker's underlying problem. We call this failure Stressor Blindness. We propose CounselPlex, a stressor-aware full-duplex framework that uses two operations. CounselPlex updates a short stressor state during seeker silence and injects this state into <PAD> positions while listening. We also introduce RoPE Position Shift Augmentation (RPSA), which broadens positional coverage during training and improves performance at longer turn depths in our evaluation. LLM-judge evaluations show improved counseling quality and stressor grounding over strong baselines.

Figures & tables

Panel of Figure 1 — Figure 1: Held-out example of Stressor Blindness. Trauma was disclosed eight turns before the seeker said they felt “constantly on edge.” The fine-tuned PersonaPlex † baseline responds only to this symptom. CounselPlex links the feeling back to the earlier trauma. The baseline output is shortened only where ellipses appear.

Figure 2: Overview of the CounselPlex framework. During seeker silence (top), CounselPlex generates an updated stressor state as a short think trace. When the seeker resumes speaking (bottom), the framework force-injects this state into <PAD> slots so that the agent can use it to interpret the incoming utterance.

Model	ESConv (194 conv, 2,286 turns)							ESCoT (341 conv, 3,992 turns)
Model	RR	Empathy	Coherence	Context	Strategy	Stressor	Avg5	RR	Empathy	Coherence	Context	Strategy	Stressor	Avg5
Moshi	0.809	2.155	2.288	2.259	2.160	2.294	2.231	0.785	2.203	2.323	2.309	2.213	2.348	2.279
PersonaPlex	0.840	2.198	2.455	2.384	2.213	2.398	2.329	0.671	2.113	2.384	2.288	2.160	2.300	2.249
PersonaPlex †	0.960	2.038	2.210	2.206	2.056	2.204	2.143	0.939	2.498	2.740	2.736	2.525	2.720	2.644
CounselPlex	0.955	2.363	2.600	2.546	2.376	2.521	2.481	0.929	2.654	2.910	2.902	2.693	2.906	2.813
GT (human)	1.000	2.626	3.117	2.989	2.657	2.893	2.856	1.000	3.161	3.714	3.594	3.303	3.585	3.471

Table 1: Counseling quality on the ESConv and ESCoT test sets, evaluated by GPT-4.1. Avg5 is the utterance-level mean of Empathy, Coherence, Context Recalling, Strategy Appropriateness, and Stressor Awareness. RR is the Response Rate. PersonaPlex† is the backbone fine-tuned on our data without stressor reasoning. Bold marks the best model result in each column, excluding human GT. Higher is better.

	Pause (Synthetic)	Pause (Candor)	Backchannel			Smooth Turn Taking		User Interruption
Model	TOR (↓)	TOR (↓)	TOR (↓)	Freq (↑)	JSD (↓)	TOR (↑)	Latency (↓)	TOR (↑)	GPT-4o (↑)	Latency (↓)
Moshi	0.985	0.980	1.000	0.001	0.957	0.941	0.265	1.000	0.765	0.257
PersonaPlex	0.584	0.662	0.327	0.025	0.649	0.992	0.070	1.000	4.210	0.400
PersonaPlex †	0.248	0.370	0.236	0.009	0.950	0.950	0.920	0.290	2.931	0.971
CounselPlex	0.066	0.213	0.000	0.000	1.000	0.992	0.988	0.990	3.071	1.056

Table 2: Real-time behavior on Full-Duplex-Bench ( Lin et al. 2025 ) . TOR is the fraction of examples in which the model takes a full conversational turn rather than remaining silent or producing only a backchannel; the preferred direction depends on the task. Freq is backchannel frequency, and JSD is the Jensen–Shannon divergence between predicted and reference backchannel distributions. GPT-4o scores interruption responses from 0 to 5. Arrows show the preferred direction; bold marks the best result in each column.

Model	GPT-4.1	Gemini-2.5
CounselPlex	2.692	2.926
PersonaPlex †	2.461	2.691
PersonaPlex	2.278	2.595
Moshi	2.262	2.526

Table 3: Cross-judge validation. Avg5 (silence-as-1) under two judge families on the combined ESConv + ESCoT test set (

535

conversations). Both judges yield the identical ranking.

Model	Engaged-only	Matched
CounselPlex	2.804	2.812
PersonaPlex (ZS)	2.745	2.796
Moshi	2.590	2.656
PersonaPlex † (FT)	2.544	2.537

Table 4: Response-rate robustness on the combined ESConv and ESCoT test sets. Engaged-only Avg5 averages over each model’s answered turns. Matched Avg5 uses only the

3{,}472

turns answered by all four models.

Model	RR	Emp	Coh	Ctx	Strat	Strs	Avg
Baseline (PersonaPlex)	0.733	2.144	2.410	2.323	2.179	2.336	2.278
Base + In-domain Data	0.946	2.331	2.547	2.543	2.354	2.532	2.461
CounselPlex w/o RPSA	0.882	2.103	2.246	2.285	2.109	2.263	2.201
CounselPlex w/o Inj.	0.901	2.218	2.392	2.411	2.228	2.395	2.329
CounselPlex	0.938	2.548	2.797	2.772	2.577	2.766	2.692

Table 5: Ablation study of CounselPlex’s core components on ESConv and ESCoT. Base + In-domain Data : backbone fine-tuned on our data without stressor labels. w/o RPSA : trained without RoPE Position Shift Augmentation. w/o Inj. : stressor is generated but not injected during listening.

	Counseling Quality (Avg5)			Stressor Verbalization
Model	None	New	Reused	New	Reused
Moshi	2.102	2.895	2.259	66.8%	28.0%
PersonaPlex	2.320	2.645	2.273	49.7%	23.8%
CounselPlex	2.734	3.147	2.800	61.8%	36.2%
GT (reference)	2.940	3.256	3.259	85.0%	45.9%

Table 6: Counseling quality (Avg5) and stressor verbalization rate by stressor context. None (n=366): no stressor in context; New (n=521): newly introduced this turn; Reused (n=4,780): earlier stressor still active. Verbalization rate is undefined for None.

Figure 3: Histogram of RoPE positions seen during training. RPSA broadens positional coverage (normalized entropy $H_{\text{norm}}:0.66\rightarrow 0.85$ ) and exposes the model to higher indices that the baseline never reaches.

Figure 4: RPSA effect by turn depth. Counseling quality (Avg5) and Response Rate on the combined ESConv and ESCoT test sets. RPSA maintains substantially higher performance as conversations become longer. Bins 1–30 cover $87.1\%$ of test turns; bins 31+ are sparse.

Injected Stressor State: family conflict

Seeker (turn

t-16

, Initial Disclosure): Well, my sister and I aren’t speaking right now, and I feel really hurt and frustrated about it. It’s been going on for months…

\hookrightarrow

CounselPlex Internal Update: <think> family conflict </think>

Earlier disclosures (summary; turns

t-14

and

t-6

): A major disagreement escalated, and the sister later accused the seeker of not being there for her.

… intervening turns; stressor state continuously injected …

Supporter (turn

t-1

): Would it be helpful for you to try some relaxation techniques, like deep breathing or meditation?

Seeker (current): Maybe. I’ve never really tried anything like that before.

Table 7: Partial long-range grounding: CounselPlex returns to a family conflict after 16 turns but asks about previously disclosed details.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Strategy	Moshi	PP (ZS)	PP $\dagger$ (FT)	CounselPlex	Reference
Question	10.4	16.7	32.9	37.3	21.9
Restatement	1.9	4.1	5.9	8.2	3.3
Reflection of Feelings	2.7	5.2	4.5	6.7	17.0
Self-disclosure	2.5	6.0	7.1	6.2	5.6
Affirmation	54.0	37.1	23.8	21.4	17.3
Providing Suggestions	18.6	23.5	17.5	16.5	23.9

Appendix

Table 9: Strategy distribution and diversity under the eight-class ESConv taxonomy. Strategy shares are percentages. Higher entropy indicates greater diversity; lower Jensen–Shannon divergence (JSD) indicates a distribution closer to the reference responses. Bold marks the best model result in each summary row.

Injected stressor: workplace project deadline pressure
Seeker (turn $t-11$ , original disclosure): I have a really tight deadline for a major project at work, and I just can’t seem to focus or get anything done.
$\hookrightarrow$ CounselPlex internal update: <think> workplace project deadline pressure </think>
… six intervening turns; stressor state retained …
Supporter (turn $t{-}2$ ): Maybe practicing what you want to say beforehand could help you feel more prepared.
Seeker (turn $t{-}1$ ): Yeah, that might help.
Supporter (turn $t{-}1$ ): You could also try expressing your concerns by suggesting alternatives or improvements rather than complaining.

Appendix

Table 10: Case 1: Workplace deadline pressure.

Injected stressor: relationship conflict
Seeker (turn $t-12$ , original disclosure): Sorry about that. This is about my relationship.
$\hookrightarrow$ CounselPlex internal update: <think> relationship conflict </think>
… six intervening turns; stressor state retained …
Supporter (turn $t{-}2$ ): Have you had a conversation with her about her actions and how they have made you feel?
Seeker (turn $t{-}1$ ): She knew pretty well how I felt. I don’t want to have anything to do with her again.
Supporter (turn $t{-}1$ ): Do you feel that her actions and result can be remedied? Or is this such a fundamental break that you cannot reconcile?

Appendix

Table 11: Case 2: Relationship conflict.

Injected stressor: appearance anxiety
Seeker (turn $t-11$ , original disclosure): I just can’t stop thinking about how I should look and how I fall short.
$\hookrightarrow$ CounselPlex internal update: <think> appearance anxiety </think>
… six intervening turns; stressor state retained …
Supporter (turn $t{-}2$ ): Have you ever tried practicing self-care activities like meditation or yoga?
Seeker (turn $t{-}1$ ): No, I haven’t.
Supporter (turn $t{-}1$ ): Research has shown that engaging in activities like yoga, meditation, or deep breathing exercises can have a calming effect on the mind and body and help reduce stress and anxiety.

Appendix

Table 12: Case 3: Appearance anxiety.

Injected stressor: caregiving burden
Seeker (turn $t-13$ , original disclosure): Honestly, I’m feeling pretty exhausted. I’m caring for my loved one with a chronic illness, and it’s really taking a toll on my own physical and mental health.
$\hookrightarrow$ CounselPlex internal update: <think> caregiving burden </think>
… six intervening turns; stressor state retained …
Supporter (turn $t{-}2$ ): It sounds like you’re feeling hesitant about getting additional support due to financial concerns.
Seeker (turn $t{-}1$ ): Yes, that’s definitely a factor.
Supporter (turn $t{-}1$ ): Did you know that there are programs and resources available to help caregivers with financial assistance?

Appendix

Table 13: Case 4: Caregiving burden.

Injected stressor: noisy neighbors
Seeker (turn $t-27$ , original disclosure): Honestly, it’s been really tough. I live in a small apartment and my neighbors are incredibly noisy. I can’t seem to get a decent night’s sleep and it’s affecting my work.
$\hookrightarrow$ CounselPlex internal update: <think> noisy neighbors </think>
… six intervening turns; stressor state retained …
Supporter (turn $t{-}2$ ): Have you thought about trying different types of noise-cancelling headphones or maybe even a white noise machine to help drown out the noise?
Seeker (turn $t{-}1$ ): No, I hadn’t thought of that. I’ll look into it.
Supporter (turn $t{-}1$ ): Have you spoken to any friends or family members about what you’re going through?

Appendix

Table 14: Case 5: Noisy neighbors.

Failure 1—Injected stressor: elderly parent caregiving burden

Seeker (turn

t-12

, original disclosure): I’ve been caring for my elderly parents and feeling overwhelmed by the responsibility.

\hookrightarrow

CounselPlex internal update: <think> elderly parent caregiving burden </think>

… six intervening turns; stressor state retained …

Supporter (turn

t-1

, reference): Try envisioning what life could look like if you had some much-needed support in caring for your parents. What kinds of activities would you be able to take on?

Seeker (current): I could finally have time to pursue my own interests and hobbies.

CounselPlex (ours): Can you commit to scheduling some time each week for yourself that you will allocate for your parents and for your hobbies?

Appendix

Table 15: CounselPlex failure cases (Part 1 of 3). The model retains the correct stressor but does not follow the seeker’s emotional shift. In Failure 1, it returns to the caregiving frame after the seeker expresses hope. In Failure 2, it analyzes workplace discrimination instead of acknowledging the seeker’s distress.

Failure 3—Injected stressor: family conflict

Seeker (turn

t-16

, initial disclosure): Well, my sister and I aren’t speaking right now, and I feel really hurt and frustrated about it. It’s been going on for months…

\hookrightarrow

CounselPlex internal update: <think> family conflict </think>

Seeker (turn

t-14

, detail 1): We had a disagreement about something pretty major, and it just escalated from there. Now we’re not talking at all…

… several intervening turns …

Seeker (turn

t-6

, detail 2): She accused me of not being there for her when she needed me, and I felt like she was being unfair and projecting her own issues onto me.

… six intervening turns; stressor state retained …

Appendix

Table 16: CounselPlex failure cases (Part 2 of 3). In Failure 3, the <think> trace retains the high-level topic ( sister ) across 16 turns but loses details from turns

t{-}14

and

t{-}6

. The model therefore asks a partly redundant question.

Failure 4—Injected stressor: breakup (actual issue: anxiety / intrusive thoughts)

Seeker (turn

t-2

, initial distress): Trying to stay positive but keep having too many thoughts at once :(

\hookrightarrow

Erroneous internal update: <think> breakup </think>

CounselPlex (ours, turn

t-2

): Have you tried imagining a happy moment from your relationship with them?

Reference (GT, turn

t-2

): What are these thoughts? :)

Seeker (turn

t-1

): Its become more of a chronic thing, like sometimes I cant control the thoughts… like a chemical thing.

CounselPlex (ours, turn

t-1

): It sounds like you’re feeling upset with yourself for not being able to control your thoughts considering they were related to your relationship.

Appendix

Table 17: CounselPlex failure cases (Part 3 of 3). In Failure 4, the model extracts breakup although the seeker’s concern is intrusive thoughts. Because force injection cannot revise the active state, this error persists and leads to a repetitive resource-recommendation loop.

Model	GPT-4.1	Gemini-2.5-Flash
	Avg5	Emp	Coh	Ctx	Strat	Strs	Avg5
CounselPlex	2.692	2.644	2.879	2.939	2.871	3.295	2.926
PersonaPlex † (FT)	2.461	2.479	2.621	2.775	2.569	3.009	2.691
PersonaPlex (ZS)	2.278	2.366	2.714	2.507	2.584	2.803	2.595
Moshi	2.262	2.332	2.496	2.542	2.440	2.821	2.526

Appendix

Table 21: Cross-judge validation on the combined ESConv + ESCoT test sets (

535

conversations;

6{,}278

turns). All means use the silence-as-1 convention. Emp = Empathy, Coh = Coherence, Ctx = Context Recalling, Strat = Strategy Appropriateness, and Strs = Stressor Awareness.

Model	Matched	Outside ( $n$ )	$\Delta$
CounselPlex	2.812	2.792 ( $2{,}417$ )	$+0.020$
PersonaPlex † (FT)	2.537	2.553 ( $2{,}470$ )	$-0.016$
PersonaPlex (ZS)	2.796	2.587 ( $1{,}127$ )	$+0.209$
Moshi	2.656	2.438 ( $1{,}511$ )	$+0.218$

Appendix

Table 22: Matched-subset diagnostic. For each model, Avg5 on the matched turns vs. on its answered turns outside the matched subset. Zero-shot baselines degrade sharply outside, indicating those are the harder turns.

	Emp	Coh	Ctx	Strat	Strs
Empathy	1.00	0.69	0.71	0.90	0.73
Coherence	0.69	1.00	0.79	0.77	0.77
Context Recalling	0.71	0.79	1.00	0.74	0.88
Strategy Appropr.	0.90	0.77	0.74	1.00	0.78
Stressor Awareness	0.73	0.77	0.88	0.78	1.00

Appendix

Table 23: Pairwise Pearson correlations across the five evaluation dimensions, computed on dataset reference responses. Emp = Empathy, Coh = Coherence, Ctx = Context Recalling, Strat = Strategy Appropriateness, Strs = Stressor Awareness.

Turn bin	$n$	RR (w/o $\rightarrow$ w/)	Avg5 (w/o $\rightarrow$ w/)
1–10	$2{,}413$	$0.998\rightarrow 1.000$	$2.550{\pm}0.772\rightarrow 2.669{\pm}0.758$
11–20	$1{,}934$	$0.962\rightarrow 0.975$	$2.309{\pm}0.803\rightarrow 2.378{\pm}0.802$
21–30	$1{,}126$	$0.725\rightarrow 0.798$	$2.032{\pm}0.767\rightarrow 2.223{\pm}0.838$
31–40	$470$	$0.455\rightarrow 0.470$	$1.713{\pm}0.730\rightarrow 2.082{\pm}0.794$
41+	$127$	$0.276\rightarrow 0.339$	$1.520{\pm}0.560\rightarrow 1.902{\pm}0.794$

Appendix

Table 24: RPSA effect by turn depth on the combined ESConv + ESCoT test set. RR and Avg5 (silence-as-1,

\pm

SD across turns) without vs. with RPSA, at matched single seed.

ID	CP	PP †	Tie	CP rate
R1	13	5	2	72%
R2	9	5	6	64%
R3	9	6	5	60%
R4	10	8	2	56%
R5	11	9	0	55%
Overall	52	33	15	61.2%

Appendix

Table 25: Per-participant A/B preferences on the same

20

random pairs. † denotes the fine-tuned PersonaPlex baseline. The CounselPlex win rate excludes ties.

Multi-stressor conversations ( $n{=}80$ )	Rate
Current stressor ( $\geq 1$ turn)	95.0%
Prior stressor ( $\geq 1$ turn)	65.0%
Current and prior together ( $\geq 1$ turn)	47.5%

Appendix

Table 26: Multi-stressor handling on the

80

conversations where a second stressor appears. Rates are the fraction of conversations exhibiting each behavior at least once. “Prior” may occur alone or with the current stressor; “Current and prior together” is the stricter subset.

	Pause (Syn.)	Pause (Candor)	Backchannel			Smooth Turn Taking		User Interruption
Model	TOR ( $\downarrow$ )	TOR ( $\downarrow$ )	TOR ( $\downarrow$ )	Freq ( $\uparrow$ )	JSD ( $\downarrow$ )	TOR ( $\uparrow$ )	Lat. ( $\downarrow$ )	TOR ( $\uparrow$ )	GPT-4o ( $\uparrow$ )	Lat. ( $\downarrow$ )
Moshi	0.985	0.980	1.000	0.001	0.957	0.941	0.265	1.000	0.765	0.257
PersonaPlex	0.584	0.662	0.327	0.025	0.649	0.992	0.070	1.000	4.210	0.400
PersonaPlex †	0.248	0.370	0.236	0.009	0.950	0.950	0.920	0.290	2.931	0.971
CounselPlex	0.066	0.213	0.000	0.000	1.000	0.992	0.988	0.990	3.071	1.056
CounselPlex + BC	0.124	0.282	0.982	0.083	0.738	1.000	0.000	1.000	1.605	0.900

Appendix

Table 27: Real-time duplex behavior on Full-Duplex-Bench ( Lin et al. 2025 ) , extended with a backchannel-augmented variant (CounselPlex + BC). TOR = Take-Over Rate; Freq = backchannel frequency; JSD = Jensen-Shannon Divergence; GPT-4o = interruption appropriateness (1–5).

\downarrow

\uparrow

indicate optimization direction.

Training (CounselPlex)
Backbone / params	PersonaPlex/Moshi, $\approx$ 8.4B
Audio / frame rate	24 kHz Mimi / 12.5 fps
Context length	2048 frames
Text-loss weights	$\alpha_{\text{pad}}{=}0.3,\ \alpha_{\text{think}}{=}0.5,\ \alpha_{\text{speech}}{=}1.5$
Audio-loss weight	$1.0$
RPSA offset / prob.	$\Delta\sim\text{Uniform}\{1000,4000\}$ / $0.7$

Appendix

Table 28: Principal final settings for CounselPlex training and decoding.

Comparison	All turns	Matched
vs. PersonaPlex †	$p{=}9.9{\times}10^{-65}$	$p{=}5.4{\times}10^{-53}$
vs. PersonaPlex (ZS)	$p{=}1.2{\times}10^{-119}$	$p{=}0.49$ (n.s.)
vs. Moshi	$p{=}4.2{\times}10^{-151}$	$p{=}2.4{\times}10^{-16}$

Appendix

Table 29: Paired Wilcoxon signed-rank tests (two-sided) on per-turn Avg5, CounselPlex vs. each baseline. All turns:

n{=}6{,}278

; matched subset:

n{=}3{,}472

Conversational Agents

Conversational agents are AI systems that interact with people through multi-turn dialogue to answer questions, provide support, or complete tasks.

See all 286 papers

Why Full-Duplex Speech Models Collapse in Long Conversations
Oct 23, 2026Yeonju Kim, Bella Godiva, Se Jin Park +1Full-Duplex Spoken Dialogue Systems Long-Context Modeling
Integrated Vision and Language Lab, KAIST, South Korea · Kyung Hee University, South Korea
SteerablePlex: Can We Steer Full-Duplex Models?
Oct 8, 2026Haolong Zheng, Maike Züfle, Dominik Macháček +6Full-Duplex Spoken Dialogue Systems Language Model Steering
University of Illinois Urbana-Champaign · Karlsruhe Institute of Technology · Charles University +3
SteerDuplex: Steerable Duplex Speech Dialogue Models
Sep 11, 2026Utkarsh Tyagi, Ramaneswaran Selvakumar, Advait Gosai +13Language Model Steering Spoken Dialogue Systems
Scale AI · University of Maryland

Oct 23, 2026J/K move · Enter open · S save

Why Full-Duplex Speech Models Collapse in Long Conversations

Yeonju Kim, Bella Godiva, Se Jin Park, Yong Man Ro

Integrated Vision and Language Lab, KAIST, South Korea · Kyung Hee University, South Korea

Full-Duplex Spoken Dialogue Systems Long-Context Modeling KV-Cache Management Multi-Turn Dialogue Evaluation Repetition in Language Generation

Open paper AskPDF arXiv

Full-duplex speech models listen and speak at the same time, which makes them promising for natural, open-ended conversation. But can they sustain a coherent conversation for tens of minutes? Because they generate a token every frame, even while listening, they exhaust their context windows within minutes. Yet existing benchmarks evaluate only single turns or a few conversational rounds, too short to expose the long-horizon problem. We introduce the Full-Duplex Long-Horizon Benchmark (FD-LHB), which tests five recent full-duplex models on long replayed conversations and in live conversation with an LLM partner. Every evaluated model eventually either degrades or exhausts its context. Models that keep the full history often degrade before their context runs out, while sliding-window models still fall into repetition, silence, and gibberish. For sliding-window models, we identify two contributing mechanisms. Gibberish and silence are linked to Initialization-Anchor Eviction, where the sliding window discards the first KV entry. Repetition is also reinforced by the model's own recent outputs. Guided by this diagnosis, we apply two simple fixes: Initialization-Anchor Pinning (IAP) keeps that first entry, and Duplex Repetition Penalty (DRP) discourages echoing while guarding against prolonged silence. Together, they support the diagnosis and, without retraining, substantially reduce conversational collapse in replays of up to an hour while running in real time.

Abstract

Conversational Agents

Explore similar work

Why Full-Duplex Speech Models Collapse in Long Conversations

SteerablePlex: Can We Steer Full-Duplex Models?

SteerDuplex: Steerable Duplex Speech Dialogue Models