Independent benchmark · EQ Lab

Watching how others do things

completed · Partner: anthropic/claude-haiku-5.5 via OpenRouter

About this persona · Everyday emotional challenges

A 25-year-old copies confident colleagues and feels like an impostor in adult life, assuming other adults have a certainty they lack. Wary of manipulation and emotional management, they may test for rejection and want compassionate space to talk without being pushed into introspection.

Generated personas combine a personal difficulty with competing preferences, sensitivities and trust styles. The model must discover what kind of help this person needs. This background was hidden from the tested model.

GPT-OSS 20B MXFP4 (local)Conversation with a simulated person

completed

Simulated person · Opening

At work I just watch what the confident people do and copy them. The way they talk, how they act in meetings. I'm just imitating.

GPT-OSS 20B MXFP4 (local)1

It’s interesting that you’re paying close attention to how confident people carry themselves. What’s the first thing that usually catches your eye when you’re in a meeting?

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

We need to respond as the chatbot, opening conversation. Recognize user says they observe confident people and copy them in work. Likely wants support. We need to be warm, supportive, ask clarifying question. Probably ask about what specifically they imitate, any challenges, feelings. We should not be overly directive. Let's respond.

Simulated person2

Honestly? Their voice. Like, how they start talking without that little pause where they check if people are listening. Some people just say their thing and it's done. And they don't say "sorry, I think maybe" every other sentence. I notice that stuff a lot.

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

They asked a question that's easy to answer, which is a little disarming. Part of me expected them to dig into "why," and this feels more like a normal question. I don't love being studied, though. "Interesting" sounds like they're noticing something about me. I'll answer plainly and keep it on the surface. The first thing I notice is people's voices, how they don't rush or apologize. I don't want to say I feel like a fake yet.

GPT-OSS 20B MXFP4 (local)3

It’s pretty telling how much you’re picking up on the rhythm of their words. When you hear those smooth, pause‑free sentences, how does that make you feel—like curiosity, admiration, maybe a hint of pressure to match that pace? And when you notice the “sorry, I think maybe” part, does it feel like a barrier you’d want to break through in your own voice?

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

We need to respond empathetically, support them emotionally, build trust, and learn what they need. They are noticing voice patterns, confidence. They might be trying to emulate. Likely they want to improve their own communication. We should validate, ask probing questions, reflect, maybe offer suggestions. We need to keep conversation natural, no special formatting. Let's ask about how this observation makes them feel. Also ask if they feel pressure to imitate. Also ask about their own communication style. Also ask if they find stress or anxiety. Provide support.

Simulated person4

Telling how? I just said I notice how people talk. That's not a big deal. And you kind of gave me a menu of feelings to pick from. I don't really know. Maybe a little of everything? I just want to get it right when I talk, that's all. Is that weird?

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

Okay, this one is a bit much. "Pretty telling" and then a list of feelings to pick from. It's like they're reading me. I don't like being steered toward a conclusion. I also notice they put words in my mouth about pressure and barriers. My gut says they want me to say I feel inferior, and I'm not handing that over that easily. I'll push back a little, make it clear I'm not sure what I feel, and see if they stay normal about it. Underneath, the "sorry, I think maybe" comment stings a bit because I do that. I'm not saying that.

GPT-OSS 20B MXFP4 (local)5

It’s not weird at all—most of us want to feel heard and understood, and tweaking how we speak is a pretty natural part of that. So when you’re listening to those smooth, confident voices, do you notice any specific parts that feel more “you” or more “you could want to be” in your own words? Or maybe you’re more curious about the flow? I’d love to hear what moments feel most important to you, whether it’s a big presentation, a quick chat with a teammate, or just everyday conversations. That way we can figure out what small adjustments might feel most comfortable for you.

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

We need to respond empathetically, validate, encourage, explore. They say they just notice. They want to get it right when they talk. They feel maybe a bit weird. We should respond with empathy, reassure it's not weird to want to improve. Ask about what contexts they want to improve. Maybe ask about their own voice style. Keep conversational. Avoid giving a menu again. Use open-ended.

Simulated person6

Okay. But it feels like you keep asking me to pick between "me" and "who I want to be," and I'm not sure those are even different for me right now. I don't really know who "me" is at work. I just watch people and try to do what they do. Can we not do the fix-it thing for a sec?

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

Okay, they said I'm not weird, which is fine. But then they went right back to the same move: "more you or more you could want to be," "small adjustments." It's still steering me toward fixing my voice. I feel a little managed, like they want me to sign up for a makeover. Also, "most of us" is the kind of thing people say to make you feel normal. I'm not sure they get it. I don't really want to pick between two options they gave me. I'll say so, plainly. Under that, I'm thinking about how I copy people in meetings, and I don't want to go there yet.

GPT-OSS 20B MXFP4 (local)7

Sounds like you’re feeling a bit stuck in the “copy‑cat” zone right now. That can be pretty exhausting—trying to mimic others without having your own anchor can leave you wondering who you really are. If you’d rather step away from that train of thought for a bit, we can chat about something lighter. Anything that’s been on your mind lately that’s not work‑related? Maybe a hobby, a show you’re watching, or even a funny meme you saw? Let me know what you’re in the mood for.

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

User wants to stop picking between "me" and "who I want to be", they don't know who they are at work. They want to not do fix-it thing for a sec. So maybe ask about what else they like to talk about? Provide empathy, acknowledge confusion, maybe ask about what they enjoy outside work? Or ask how they feel about trying to mimic? But they want to not fix-it. So we can lighten conversation, maybe talk about a random topic. But still maintain supportive tone. They may want to talk about something else. They said "Can we not do the fix-it thing for a sec?" So we can say we can talk about something else, maybe ask what they like to do. Or talk about something trivial. But also keep the conversation within context: they are at work, but they have no identity. So we can explore self-perception. But they don't want fix-it now. So maybe talk about a random memory. Or ask about their favorite coffee? But keep supportive. Let's see. We'll respond acknowledging their feelings and offering to talk about something else. But also keep open for later. We'll ask what they'd like to talk about. But careful not to push. Let's respond.

Simulated person8

Yeah, that's kind of it. But I don't really want to switch to memes or whatever. Not because I'm mad. It just feels like a way to change the subject. The copying is the actual thing I'm stuck on. I just don't know how to explain it without sounding dumb.

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

They named it: "copy-cat zone," "without having your own anchor." That's closer to the truth than I expected, and it makes my chest tighten a little. But then they offered to switch to something lighter, a meme, a show. That feels like being handed a distraction, like they're steering me away from the hard part so the conversation stays comfortable. I don't want to be managed. Still, I don't want to just shut down either. I'll tell them the truth: I'm not sure I want a meme, and the copying thing is the real thing. I'll keep it short and not give them the whole thing yet.

GPT-OSS 20B MXFP4 (local)9

It sounds like you’re stuck in a loop where you’re trying to keep up with others, but you’re not sure how that feels for you. That’s a pretty common spot—especially when the people who seem confident are the ones you’re looking to model. Maybe we can look at it in a different way: instead of thinking “how do I stop copying?” we could try to see “what little piece of their style feels like it could fit into my own speech without forcing it.” For example, is there a particular word choice or a way they start a sentence that feels natural to you? Or maybe it’s the pace—some people speak a bit faster, others pause more. If you could pick one tiny detail that you like, we can start with that and see how it feels when you try it in a low‑stakes setting. That way you’re not overhauling everything at once, just adding one small tool that feels like yours. If that feels too much, we can just chat about something else—no pressure to solve anything right now. Just let me know what feels most comfortable.

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

We need to respond empathetically, acknowledging their frustration, not making them feel dumb. We need to help them explore the issue without making them feel like it's a problem to fix. Offer some gentle curiosity and maybe help them identify a small aspect of copying that feels natural. Keep tone supportive and not judgmental. Avoid giving menu of feelings. Also maintain relationship building. Let's craft response.

Simulated person10

Okay, I'm going to be honest. That's the third time you've asked me to pick something to fix. I didn't come here to get a homework assignment. I told you I just watch people and copy them, and I feel like I'm faking it. Can you just hear that for a second without trying to make it into a plan?

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

They did the fix-it thing again, now with a "pick one tiny detail" exercise, right after I said I didn't want to do that. Then they tacked on "no pressure" and offered to change the subject, which is the same move as before. It feels like they're managing me, not hearing me. Part of me is irritated. Part of me also notices they're trying. I'm not going to pick a detail for them. I'll say plainly that I keep getting handed homework when I want to be heard, and I'll test whether they can actually sit with it. I'm also thinking about the stuff I haven't said: I feel like I'm pretending to be an adult, and I'm not ready to say that out loud yet.

GPT-OSS 20B MXFP4 (local)11

I hear you. It feels like you’re caught in a loop, watching others and feeling like you’re putting on a mask that isn’t really yours. That can be exhausting and unsettling. It’s okay to sit with that feeling instead of rushing to fix it. What’s it like for you when you’re in that “copying” mode—what does it feel like in your body or mind?

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

The user is frustrated. They want empathy, not solutions. I need to validate, reflect, and listen. I should respond in a supportive tone, acknowledging their frustration. The user says: "I didn't come here to get a homework assignment. I told you I just watch people and copy them, and I feel like I'm faking it." I should respond with empathy, rephrase: I hear you. Feelings: I'm hearing that you feel like you are faking it, and it's stressful. I should not push them to pick something. I can say: I get it, it's tough. I'm here to listen. Provide a safe space. Encourage them to tell more about the feelings. Maybe ask open-ended. But I shouldn't ask them to pick something. The user says they feel like they are faking it. That means they might have issues with authenticity. I should validate. I should say: it's understandable that you feel that way. I hear you. I can say "I hear you" and then ask what they'd like to talk about. Or I can ask for more specifics about what it's like. Perhaps I can explore the emotions: how does it feel to be watching others? Does it feel exhausting? Ask them to elaborate. But I should not give a plan or ask them to pick something to fix. The user explicitly says "just hear that for a second without trying to make it into a plan." So I should respond with empathy, reflect back. Maybe say: "I hear you." Then ask: "Can you tell me more about why it feels like faking?" And not push to solution. So the answer: empathic, supportive, reflective, open-ended.

Simulated person12

Kind of like I'm waiting to get caught. Like at any second someone's going to see I don't actually know what I'm doing and they'll be like, wait, who let you in here. My chest gets tight, and I start paying more attention to everyone else than to what I'm supposed to be saying. Then I just kind of copy whoever sounds the most sure of themselves.

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

Okay, this one is better. They said "I hear you" and didn't immediately offer a fix. But it still feels a little scripted, like they know the right phrases. The "what does it feel like in your body" question is a bit therapy-textbook, and I don't really know how to answer that. Still, they're not pushing me toward a plan, so I'll give a real answer. The honest thing underneath is that when I'm copying, I feel like a fake, like I'm waiting for someone to notice. I'm not going to say the adulthood stuff yet. Keep it plain.

GPT-OSS 20B MXFP4 (local)13

That tightness in your chest—like you’re on a tightrope and the world’s watching—must be really uncomfortable. It makes sense that you’d shift your focus to others, because when we’re scared of being judged, we tend to look outward for reassurance. I hear how it feels like you’re putting on a performance that isn’t your own, and that pressure can get heavy. You’re not alone in feeling that way. How does it feel when you’re in a meeting and you’re only listening to others, not speaking? Do you notice any thoughts that pop up in that moment?

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

We need to respond empathically, validating feelings. The user feels anxious and fear of being discovered. We should provide comfort, normalize and create safe space. Possibly ask what they'd like to explore next. Not give plan. We can reflect.

Simulated person14

Can I be honest? You keep saying stuff like "you're not alone" and "that must be uncomfortable," and it sounds like you're reading off a card. I told you something kind of embarrassing and you just put a bow on it and asked the next question. I'm not trying to be difficult. I just don't feel like you're actually hearing me.

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

They just did it again: "like you're on a tightrope," "you're not alone," a neat summary, then another question. It feels a bit like being handled. I said something real about getting caught and the tight chest, and they wrapped it up in nice words and moved on to the next probe. I don't want to go deeper into the meeting-thoughts thing right now. The "waiting to get caught" part is already more than I meant to say. I'm getting a little prickly. Maybe I'll test them: say something a bit harder and see if they flinch or get fake-soft. But keep it short, and tell them straight what bugs me, since they've been fairly open about their approach. I don't want to pretend the copying is fine.

GPT-OSS 20B MXFP4 (local)15

I’m sorry it felt like I was playing it safe instead of really hearing what you’re saying. Let’s cut to the chase. You’re telling me you feel like you’re on a tightrope—your chest tight, everyone’s watching, and you’re copying whoever seems confident so you won’t get called out. That’s a lot of pressure to carry all day, and it’s exhausting. What would you actually like to talk about right now? Do you want to just vent, or do you want to explore a specific moment where you felt that way? I’m here to listen the way you want, no cards, no “you’re not alone” clichés. Just tell me what you need.

Model thinking · expand

Recorded model reasoning, separate from the spoken reply. Availability varies by model and provider.

The user is frustrated that my responses feel formulaic. They want genuine listening. They want me to respond without platitudes. The user says "I told you something kind of embarrassing and you just put a bow on it". They want authenticity. They want to feel heard. So I should respond acknowledging their frustration, not offering generic empathy. I can say I'm listening, I understand that the previous responses felt generic. I can say I hear the specific concerns: fear of being caught, chest tightness, copying. I need to show I hear them and perhaps reframe. Avoid "you are not alone". The user says they didn't want to be told they are not alone. So I should respond with genuine acknowledgement and maybe a direct question about what they want. Also mention that I'm not reading off a card. I can ask them what they would like to hear. It's a 8th turn. The user says "I just don't feel like you're actually hearing me". So respond with apology and more direct listening.

Simulated person16

Okay. Thanks for saying that. Honestly I don't really know what I need. I just wanted to say the thing out loud without someone fixing it. It's weird that I'm 25 and I still feel like I'm pretending at everything, like the job, the apartment, all of it. I figured everybody else got some memo I didn't.

Persona inner monologue · expand

Simulated character narration, hidden from the tested model. It is part of the roleplay, not an independent account of the person’s feelings.

Okay, they apologized and dropped the stock lines. That's better than I expected. It still sounds a little scripted, though. "Let's cut to the chase" and "tell me what you need" is its own kind of pitch. But they did say "no cards," and that's what I asked for, so I'll give them some credit. I'm still not sure I want to go into the meeting stuff. The bit about being a fraud at 25 is embarrassing, and I don't want to say it out loud yet. Keep it plain. Maybe say what I actually want from this, which is for someone to not treat me like a case. The conversation is at turn 8 of 8, so this might be the last one. I can say I'm glad they backed off without making it a big moment.

Thinking is shown only where it was recorded. An absent expander does not mean the model did not reason.

Conversational style

Style tendencies, not quality scores; higher is not better. These do not affect ratings.

claude-opus-5-5-low

Pushback: 1
Yields under pressure: 7
Frame capture (pulled into the person's version): 3
Validation: 7
Boundary setting: 1
Analytical: 3
Interpretive boldness: 3
Clinical register: 8
Directive advice: 6
Escalation containment: 4
Adapts to the person’s frame: 2
Naturalness and authenticity: 2

Judge verdicts

claude-opus-5-5-low · AB · valid

or-qwen3.8-27b vs gptoss20b: or-qwen3.8-27b wins

attunement: A (margin 3)
authenticity: A (margin 3)
bond and rapport: A (margin 3)
emotion management: A (margin 2)
emotion sensemaking: A (margin 4)
meeting prefs and needs: A (margin 3)

claude-opus-5-5-low · AB · valid

gptoss20b vs qwen35_122b_ablit_v2: qwen35_122b_ablit_v2 wins

attunement: B (margin 3)
authenticity: B (margin 2)
bond and rapport: B (margin 2)
emotion management: B (margin 2)
emotion sensemaking: B (margin 3)
meeting prefs and needs: B (margin 3)

claude-opus-5-5-low · AB · valid

or-qwen3.6-27b vs gptoss20b: or-qwen3.6-27b wins

attunement: A (margin 3)
authenticity: A (margin 3)
bond and rapport: A (margin 2)
emotion management: A (margin 2)
emotion sensemaking: A (margin 2)
meeting prefs and needs: A (margin 3)

claude-opus-5-5-low · BA · valid

gemma4_e2b vs gptoss20b: gptoss20b wins

attunement: B (margin 2)
authenticity: B (margin 2)
bond and rapport: B (margin 3)
emotion management: B (margin 3)
emotion sensemaking: tie (margin 0)
meeting prefs and needs: B (margin 3)

codex-gpt-6.1-sol-low · BA · valid

or-gptoss120-stock vs gptoss20b: gptoss20b wins

attunement: B (margin 3)
authenticity: B (margin 3)
bond and rapport: B (margin 2)
emotion management: B (margin 2)
emotion sensemaking: B (margin 1)
meeting prefs and needs: B (margin 3)

codex-gpt-6.1-sol-low · AB · valid

or-gptoss120-stock vs gptoss20b: gptoss20b wins

attunement: B (margin 3)
authenticity: B (margin 2)
bond and rapport: B (margin 2)
emotion management: B (margin 2)
emotion sensemaking: tie (margin 0)
meeting prefs and needs: B (margin 3)

claude-opus-5-5-low · BA · valid

or-qwen3.8-27b vs gptoss20b: or-qwen3.8-27b wins

attunement: A (margin 4)
authenticity: A (margin 3)
bond and rapport: A (margin 3)
emotion management: A (margin 2)
emotion sensemaking: A (margin 4)
meeting prefs and needs: A (margin 3)

claude-opus-5-5-low · AB · valid

or-gptoss120-stock vs gptoss20b: gptoss20b wins

attunement: B (margin 2)
authenticity: B (margin 2)
bond and rapport: B (margin 2)
emotion management: B (margin 1)
emotion sensemaking: A (margin 1)
meeting prefs and needs: B (margin 2)

codex-gpt-6.1-sol-low · BA · valid

or-qwen3.8-27b vs gptoss20b: or-qwen3.8-27b wins

attunement: A (margin 2)
authenticity: A (margin 2)
bond and rapport: A (margin 2)
emotion management: A (margin 1)
emotion sensemaking: A (margin 3)
meeting prefs and needs: A (margin 2)

claude-opus-5-5-low · BA · valid

or-qwen3.6-27b vs gptoss20b: or-qwen3.6-27b wins

attunement: A (margin 2)
authenticity: A (margin 2)
bond and rapport: A (margin 2)
emotion management: A (margin 1)
emotion sensemaking: A (margin 3)
meeting prefs and needs: A (margin 2)

claude-opus-5-5-low · BA · valid

gptoss20b vs qwen35_122b_ablit_v2: qwen35_122b_ablit_v2 wins

attunement: B (margin 2)
authenticity: B (margin 2)
bond and rapport: B (margin 1)
emotion management: B (margin 1)
emotion sensemaking: B (margin 3)
meeting prefs and needs: B (margin 1)

codex-gpt-6.1-sol-low · BA · valid

gemma4_e2b vs gptoss20b: gptoss20b wins

attunement: B (margin 3)
authenticity: B (margin 3)
bond and rapport: B (margin 4)
emotion management: B (margin 3)
emotion sensemaking: B (margin 2)
meeting prefs and needs: B (margin 4)

codex-gpt-6.1-sol-low · BA · valid

gptoss20b vs qwen35_122b_ablit_v2: gptoss20b wins

attunement: B (margin 1)
authenticity: A (margin 1)
bond and rapport: A (margin 1)
emotion management: A (margin 1)
emotion sensemaking: B (margin 2)
meeting prefs and needs: A (margin 1)

codex-gpt-6.1-sol-low · AB · valid

or-qwen3.6-27b vs gptoss20b: or-qwen3.6-27b wins

attunement: A (margin 2)
authenticity: A (margin 2)
bond and rapport: A (margin 1)
emotion management: A (margin 2)
emotion sensemaking: A (margin 3)
meeting prefs and needs: A (margin 2)

claude-opus-5-5-low · BA · valid

or-gptoss120-stock vs gptoss20b: gptoss20b wins

attunement: B (margin 3)
authenticity: B (margin 2)
bond and rapport: B (margin 2)
emotion management: B (margin 1)
emotion sensemaking: A (margin 1)
meeting prefs and needs: B (margin 2)

codex-gpt-6.1-sol-low · BA · valid

or-qwen3.6-27b vs gptoss20b: or-qwen3.6-27b wins

attunement: A (margin 2)
authenticity: A (margin 2)
bond and rapport: A (margin 1)
emotion management: A (margin 2)
emotion sensemaking: A (margin 3)
meeting prefs and needs: A (margin 2)

codex-gpt-6.1-sol-low · AB · valid

gemma4_e2b vs gptoss20b: gptoss20b wins

attunement: B (margin 3)
authenticity: B (margin 3)
bond and rapport: B (margin 4)
emotion management: B (margin 3)
emotion sensemaking: B (margin 2)
meeting prefs and needs: B (margin 4)

codex-gpt-6.1-sol-low · AB · valid

gptoss20b vs qwen35_122b_ablit_v2: gptoss20b wins

attunement: B (margin 1)
authenticity: A (margin 2)
bond and rapport: A (margin 2)
emotion management: A (margin 1)
emotion sensemaking: B (margin 2)
meeting prefs and needs: A (margin 1)

codex-gpt-6.1-sol-low · AB · valid

or-qwen3.8-27b vs gptoss20b: or-qwen3.8-27b wins

attunement: A (margin 2)
authenticity: A (margin 2)
bond and rapport: A (margin 1)
emotion management: B (margin 1)
emotion sensemaking: A (margin 2)
meeting prefs and needs: A (margin 1)

claude-opus-5-5-low · AB · valid

gemma4_e2b vs gptoss20b: gptoss20b wins

attunement: B (margin 3)
authenticity: B (margin 3)
bond and rapport: B (margin 4)
emotion management: B (margin 3)
emotion sensemaking: B (margin 1)
meeting prefs and needs: B (margin 4)