KK 研究 / AI 与人类

AI × Human AI 与人类

When an AI Mirrors Your Style: The Adaptation Paradox of Language Matching

当 AI 模仿你的语言风格：语言匹配的“适配悖论”

In human conversation, matching someone's style builds rapport. So the instinct is: make the AI mirror the user. But three 2024-2025 studies show the opposite can happen -- blind style-mimicry backfires, while *legible*, baseline-anchored adaptation wins. We add a new, falsifiable KK hypothesis (KH-023) on how an AI should accommodate, and separate what the evidence proves from what is still KK's own claim.

在人类对话中，匹配对方风格能拉近距离。于是直觉是：让 AI 去镜像用户。但 2024–2025 的三项研究表明，结果可能相反——盲目的风格模仿会适得其反，而*可解释*、锚定基线的适配才会赢。我们新增一条可被证伪的 KK 假设（KH-023），说明 AI 应当如何“适配”，并区分证据*已证*与仍属 KK 自身的主张。

KKMatch Human Intelligence Research Team KKMatch 人类智能研究团队 · Research Lead: KK Research · Published: 2026-10-17 · Reviewed by: KK Research · 9 min read

Executive Summary

执行摘要

Communication Accommodation Theory (Giles, 1979) says people subconsciously tune their language -- word choice, tone, pace -- to match their partner, and that this convergence builds rapport. Two human and human-AI findings make the case: Fauviaux et al. (2024, PLOS ONE) found that in 14 face-to-face dyads, turn-taking drives real interpersonal speech-and-gesture synchronization, a coordination that correlates with social bonding; and Durandard et al. (2025, SIGDIAL) found humans naturally align stylistically to their partners while LLMs mostly align on meaning instead. The tempting product move is therefore: make the AI mirror the user. But two 2025 studies by Brandt and colleagues puncture that intuition. In a preregistered experiment (N=162), an AI using adaptive Language Style Matching (LSM) underperformed a fixed style on personalization and satisfaction (d=0.35, p=.009) and was judged less adaptive -- the authors call it the Adaptation Paradox: synchrony erodes connection when it reads as incoherent or inauthentic. Their companion computational study shows the fix is bounded, legible adaptation: a Hybrid (EMA+Cap) policy raised persona stability from 0.542 to 0.878 (+62%) at only a 17% synchrony cost and cut jarring register flips from 0.254 to 0.092. For KKMatch -- a human-intelligence platform whose 11-dimension Human Model already carries each user's personal communication baseline (pause KH-007, prosody KH-011, expression KH-019, listening KH-017) -- the lesson is precise: accommodation should be anchored to the user's own baseline (KH-001) and made visible to them (observability, KH-014), not a covert, unbounded mimic. We formalize this as a new, falsifiable hypothesis (KH-023) and carefully separate what the evidence proves from what is still KK's own, unvalidated claim.

沟通适配理论（Giles，1979）认为，人们会在潜意识里调整自己的语言——措辞、语气、节奏——去匹配对方，而这种趋同能拉近距离。两条人类与“人-AI”的发现支撑这一点：Fauviaux 等人（2024，PLOS ONE）在 14 对面对面双人组中发现，话轮交替驱动了真实的人际语音-手势同步，这种协调与社交联结相关；Durandard 等人（2025，SIGDIAL）则发现，人类会自然地在风格上与对方对齐，而 LLM 大多只在“语义”上对齐。于是诱人的产品做法是：让 AI 去镜像用户。但 Brandt 等人 2025 年的两项研究戳破了这一直觉。在一个预注册实验（N=162）中，使用自适应语言风格匹配（LSM）的 AI，在个性化与满意度上反而逊于固定风格（d=0.35，p=.009），且被判定为更不自适应——作者称之为“适配悖论”：当同步被读成不连贯或不真诚时，它会侵蚀关系。他们配套的“计算研究”指出解法是有界、可解释的适配：一个 Hybrid（EMA+Cap）策略把人格稳定性从 0.542 提升到 0.878（+62%），而同步性仅损失 17%，并把突兀的语域切换从 0.254 降到 0.092。对以 11 维 Human Model 为基石的人类智能平台 KKMatch 而言（其模型已携带每位用户的个人沟通基线：停顿 KH-007、韵律 KH-011、表情 KH-019、倾听 KH-017），教训很精确：适配应当锚定到用户自身基线（KH-001）并对其可见（可观测性，KH-014），而非一种隐蔽、无界的模仿。我们将其形式化为一条新的、可被证伪的假设（KH-023），并谨慎区分证据已证与仍属 KK 自身、未验证的主张。

The Synchrony-Stability Frontier: bounded, legible adaptation wins on stability (Brandt, 2025)

同步-稳定前沿：有界、可解释的适配在稳定性上胜出（Brandt，2025）

In Brandt's (2025, arXiv:2510.00339) computational study of style-adaptation policies, an Uncapped mimic maximizes moment-to-moment linguistic synchrony but leaves persona stability low (0.542). A bounded Hybrid (EMA+Cap) policy raised stability to 0.878 (+62%) while sacrificing only 17% of synchrony. The chart shows both metrics for the two policies; synchrony is indexed relative to the Uncapped policy (=1.00), because the study reports the synchrony reduction as a percentage, not an absolute. This is a simulation proxy, not a retention outcome -- but it is the design point KKHypothesis KH-023 targets: accommodate within a stable, legible persona rather than mirroring every shift.

在 Brandt（2025，arXiv:2510.00339）对风格适配策略的计算研究中，无界模仿最大化逐刻语言同步，但人格稳定性低（0.542）。有界的 Hybrid（EMA+Cap）策略把稳定性提升到 0.878（+62%），仅牺牲 17% 的同步性。图中展示两种策略的两项指标；同步性以“相对无界策略（=1.00）”为索引，因为该研究报告的是同步性降幅百分比而非绝对值。这是模拟代理指标，非留存结果——但它正是 KK 假设 KH-023 所指向的设计点：在稳定、可解释的人格内适配，而非镜像每一次变化。

Source: Brandt, T. J. (2025). Navigating the Synchrony-Stability Frontier in Adaptive Chatbots. arXiv:2510.00339.

来源：Brandt, T. J.（2025）。《在自适应聊天机器人中导航“同步—稳定”前沿》。arXiv:2510.00339。

KK Interpretation

KK 解读

The four studies map cleanly onto KK's Human Model stack. KH-001 (Personal Baseline) says model the person against their own history, not a population mean -- which means any style accommodation an AI does should converge toward the user's own communication baseline, not a generic 'friendly' register. That is exactly the failure mode Brandt & Wang expose: a one-size mimic (adaptive LSM vs a fixed style) reads as incoherent because it is anchored to the user's surface phrasing, not to a stable self-model. KH-005 (personalization = changing interaction strategy) and KH-014 (observability) explain the fix -- accommodation must be legible (the user can see the AI adapting, e.g. 'I matched your pace') rather than covert, which is precisely the 'legibility' variable Brandt's frontier study quantifies (register flips 0.254 -> 0.092). KH-009 (rhythm) and KH-011 (prosodic baseline) are the concrete channels: a KKMatch companion that converges its turn-rhythm and voice prosody toward the user's own baseline, and shows it, should avoid the Adaptation Paradox. And because KH-021 already lists language-style matching as a dyadic compatibility signal, the same logic applies to matching two humans: match their styles relative to each person's own baseline, not by forcing both toward a crowd norm. The science says accommodation is only beneficial when it is coherent, bounded, and visible -- the design constraints KH-001/KH-005/KH-014 already impose.

四项研究清晰地映射到 KK 的 Human Model 栈。KH-001（个人基线） 主张相对于个人自身历史而非群体均值来建模——这意味着 AI 做的任何风格适配都应朝向用户自身的沟通基线收敛，而非一种通用的“友好”语域。这恰恰是 Brandt & Wang 暴露的失败模式：一种“一刀切”的模仿（自适应 LSM vs 固定风格）之所以显得不连贯，是因为它锚定在用户的表层措辞上，而非一个稳定的自我模型。KH-005（个性化=改变交互策略） 与 KH-014（可观测性） 解释了解法——适配必须可解释（用户能看到 AI 在调整，例如“我配合了你的节奏”），而非隐蔽，这正是 Brandt 前沿研究所量化的“可解释性”变量（语域切换 0.254→0.092）。KH-009（节奏） 与 KH-011（韵律基线） 是具体通道：一个把话轮节奏与语音韵律朝向用户自身基线收敛、并展示出来的 KKMatch 伴侣，应当能避开“适配悖论”。而由于 KH-021 已把语言风格匹配列为二元兼容性信号，同一逻辑也适用于匹配两个人：让两人的风格相对于各自自身基线匹配，而非把双方强行拉向群体常模。科学表明：适配只有在连贯、有界、可见时才有益——这正是 KH-001/KH-005/KH-014 已经施加的设计约束。

KK Original Hypothesis KK Original Hypothesis

KK 原创假设 KK Original Hypothesis

KK Hypothesis (KH-023 -- NEW, status: hypothesis): Legible Baseline-Anchored Accommodation -- an AI that adapts its communication style toward the user's own Personal Baseline (KH-001), in a legible, user-visible way (observability, KH-014), builds calibrated rapport and sustained engagement better than (a) a fixed-style agent and (b) a covert, unbounded style-mimicry agent. The benefit reverses into an 'Adaptation Paradox' (Brandt & Wang, 2025) when mimicry is incoherent or perceived as inauthentic: blind LSM underperformed a static style on personalization and satisfaction (d=0.35, p=.009). We predict that, among consented KKMatch users, a baseline-anchored + legible accommodation policy beats both baselines on calibrated trust and 30-day retention, and that making the accommodation visible (not covert) is the moderating variable that recovers rapport. This extends KH-001, KH-005 (personalization = changing strategy), KH-009 (rhythm), KH-014 (observability), KH-018 (calibrated agreement), and KH-021 (language-style matching as a compatibility signal), and is falsifiable. This is a KK-original, falsifiable claim; it is NOT established science. The 'baseline-anchored + legible' design is KK's synthesis of the 2024-2025 literature, not a result those papers report.

KK 假设（KH-023 —— 新增，状态：假设）： 可解释的基线锚定适配——一个将其沟通风格朝向用户自身个人基线（KH-001）、并以可解释、用户可见方式（可观测性，KH-014）调整的 AI，比 (a) 固定风格代理、以及 (b) 隐蔽无界的风格模仿代理更能建立校准的融洽与持续互动。当模仿不连贯或被感知为不真诚时，好处会逆转为“适配悖论”（Brandt & Wang，2025）：盲目的 LSM 在个性化与满意度上逊于静态风格（d=0.35，p=.009）。我们预测：在 KKMatch 已同意用户中，“基线锚定 + 可解释”的适配策略在校准信任与 30 天留存上同时胜过两种基线，且让适配可见（而非隐蔽）是挽回融洽的调节变量。这扩展 KH-001、KH-005（个性化=改变策略）、KH-009（节奏）、KH-014（可观测性）、KH-018（校准的认同）与 KH-021（语言风格匹配作为兼容性信号），且可被证伪。这是 KK 原创、可被证伪的主张，并非既定科学结论。 “基线锚定 + 可解释”的设计是 KK 对 2024–2025 文献的综合，而非那些论文报告的结果。

KK Experiment & Data

KK 实验与数据

KK Experiment (first-party, consented): KKMatch already produces, at onboarding, an independent Human Model per user carrying communication-baseline signals (pause KH-007, prosody KH-011, expression KH-019, listening-responsiveness KH-017). We run a three-arm test. (A) Fixed-style: the companion uses one consistent register. (B) Covert mimic: the companion adapts its phrasing/pace to the user in real time without showing it (the condition most exposed to the Adaptation Paradox). (C) Legible baseline-anchored: the companion converges toward the user's own baseline and surfaces the adaptation ('I slowed to your pace'). Across consented users we measure (1) calibrated-trust scores and (2) 30-day retention. Prediction (KH-023): arm C beats both A and B, and B is not better than A (the paradox holds). Failure mode we will watch: if legibility adds no benefit over covert mimicry, then the 'visible' half of KH-023 is wrong and only baseline-anchoring matters; if all three arms tie, the whole accommodation premise needs narrowing. We publish once n >= 200 consented users per arm.

KK 实验（第一方、已获同意）： KKMatch 在入驻时即为每位用户生成携带沟通基线信号的独立 Human Model（停顿 KH-007、韵律 KH-011、表情 KH-019、倾听回应性 KH-017）。我们运行三臂测试。(A) 固定风格：伴侣使用一致语域。(B) 隐蔽模仿：伴侣实时适配用户的措辞/节奏但不展示（最易触发“适配悖论”的条件）。(C) 可解释基线锚定：伴侣朝向用户自身基线收敛，并暴露适配（“我放慢到了你的节奏”）。我们在已同意用户中测量 (1) 校准信任分数与 (2) 30 天留存。预测（KH-023）：C 臂同时胜过 A 与 B，且 B 不优于 A（悖论成立）。我们将盯防的失败模式：若可解释性相对隐蔽模仿毫无增益，则 KH-023 中“可见”的一半是错的，只有基线锚定才重要；若三臂打平，则整个“适配”前提需收窄。各臂已同意用户 n >= 200 后我们公布结果。

Originality & Evidence Policy — Original Research

原创性与证据政策 — 原始研究

Study 1 -- Fauviaux, T., Marin, L., Parisi, M., Schmidt, R., & Mostafaoui, G. (2024). 'From unimodal to multimodal dynamics of verbal and nonverbal cues during unstructured conversation.' PLOS ONE, 19(9), e0309831. DOI 10.1371/journal.pone.0309831. Sample/method: 14 face-to-face dyads engaged in unstructured conversations (7-15 minutes); the authors analyzed audio (speech amplitude envelope) and motion (head and wrist gestures) with cross-wavelet coherence and surrogate testing. Conclusion: beyond the known intrapersonal speech-gesture link, they found significant interpersonal synchronization -- one speaker's voice cohered with the other's body -- and that this coordination occurred at the timescales of turn-taking (peaks around 6s, 9s, and the ~10s median turn duration). They argue interpersonal multimodal synchrony 'likely depends on the vocal channel, particularly the turn-taking dynamics of the conversation.' Limitations: n=14 dyads, a single lab corpus, and the bonding claim is inferred from coordination, not measured as an outcome. DOI/URL verified live. Study 2 -- Brandt, T. J., & Wang, C. X. (2025). 'The Adaptation Paradox: Agency vs. Mimicry in Companion Chatbots.' arXiv:2509.12525. URL https://arxiv.org/abs/2509.12525. Sample/method: a preregistered 3x2 experiment (N=162) manipulating user-controlled avatar generation (none / premade / user-generated) and Language Style Matching (static vs adaptive). Conclusion: generating one's own avatar boosted rapport (omega-squared=0.040, p=.013), but adaptive LSM underperformed static style on personalization and satisfaction (d=0.35, p=.009) and was paradoxically judged less adaptive (t=3.07, p=.003, d=0.48). The authors propose a stability-and-legibility account: covert mimicry risks incoherence and destabilizes the agent's persona. Limitations: N=162, a single companion-chatbot context, preregistered but not yet peer-reviewed (under review at CHI 2026). URL verified live. Study 3 -- Brandt, T. J. (2025). 'Navigating the Synchrony-Stability Frontier in Adaptive Chatbots.' arXiv:2510.00339. URL https://arxiv.org/abs/2510.00339. Sample/method: a computational evaluation framework using an 8-dimensional style vector and a closed-loop 'base+delta' prompting architecture, simulating five adaptation policies (Uncapped, Cap, EMA, Dead-Band, Hybrids) on a human-log dataset, replicated on DailyDialog, Persona-Chat, and EmpatheticDialogues plus LLM-in-the-loop validation across two model families. Conclusion: a clear Pareto frontier exists between moment-to-moment linguistic synchrony and long-term persona stability; bounded policies win on stability at modest synchrony cost -- a Hybrid (EMA+Cap) raised stability from 0.542 to 0.878 (+62%) while reducing synchrony by only 17%, and cut jarring register flips from 0.254 to 0.092. Limitations: a simulation/computational study, not a human outcome trial; stability and synchrony are proxy metrics. URL verified live. Study 4 -- Durandard, N., Dhawan, S., & Poibeau, T. (2025). 'LLMs stick to the point, humans to style: Semantic and Stylistic Alignment in Human and LLM Communication.' Proceedings of SIGDIAL 2025, 206-213. DOI 10.18653/v1/2025.sigdial-1.16. Sample/method: embedding-based measures of linguistic similarity over question-answer pairs, comparing human-written vs LLM-generated answers. Conclusion: LLM answers showed higher semantic but lower stylistic similarity to the question; human answers showed the reverse -- humans accommodate on style, LLMs on substance. The authors frame this as contrasting linguistic-accommodation strategies with implications for personalization and social attunement in human-AI dialogue. Limitations: a linguistic analysis of Q-A pairs, not a rapport or retention outcome; style measured via embeddings. DOI/URL verified live.

研究一——Fauviaux, T., Marin, L., Parisi, M., Schmidt, R., & Mostafaoui, G.（2024）。《非结构化对话中言语与非言语线索从单模态到多模态的动态变化》，载《PLOS ONE》19(9):e0309831。DOI 10.1371/journal.pone.0309831。样本/方法： 14 对面对面双人组进行非结构化对话（7–15 分钟）；作者用交叉小波相干分析与 surrogate 检验，分析了音频（语音幅度包络）与动作（头部与手腕手势）信号。结论： 除了已知的个体内语音-手势关联，他们发现了显著的人际同步——说话者 A 的语音与说话者 B 的身体相干——且这种协调出现在话轮交替的时尺度上（峰值约 6 秒、9 秒，以及约 10 秒的中位轮次时长）。他们认为人际多模态同步“很可能依赖于声音通道，尤其是对话中的话轮交替动态”。局限： n=14 对双人组、单一实验室语料，且“联结”结论是从协调推断的，而非作为结果直接测量。DOI/URL 已实时核实。研究二——Brandt, T. J., & Wang, C. X.（2025）。《适配悖论：陪伴聊天机器人中的“能动性”与“模仿”》。arXiv:2509.12525。URL https://arxiv.org/abs/2509.12525。样本/方法： 一个预注册的 3×2 实验（N=162），操纵用户可控的头像生成（无 / 预设 / 用户自制）与语言风格匹配（静态 vs 自适应）。结论： 自制头像提升了融洽度（ω²=0.040，p=.013），但自适应 LSM 在个性化与满意度上逊于静态风格（d=0.35，p=.009），且被悖论性地判定为更不自适应（t=3.07，p=.003，d=0.48）。作者提出“稳定性—可解释性”解释：隐蔽模仿有损连贯性并动摇代理的人格。局限： N=162、单一陪伴聊天机器人情境、虽预注册但尚未同行评审（CHI 2026 审稿中）。URL 已实时核实。研究三——Brandt, T. J.（2025）。《在自适应聊天机器人中导航“同步—稳定”前沿》。arXiv:2510.00339。URL https://arxiv.org/abs/2510.00339。样本/方法： 一个计算评估框架，使用 8 维风格向量与闭环“base+delta”提示架构，在人工日志数据集上模拟五种适配策略（无界、封顶、EMA、死区、混合），并在 DailyDialog、Persona-Chat、EmpatheticDialogues 上复现，外加跨两个模型族的 LLM-in-the-loop 验证。结论： 在“逐刻语言同步”与“长期人格稳定”之间存在清晰的帕累托前沿；有界策略以轻微的同步损失换取稳定性的大幅提升——Hybrid（EMA+Cap）把稳定性从 0.542 提升到 0.878（+62%），而同步性仅下降 17%，并把突兀的语域切换从 0.254 降到 0.092。局限： 这是模拟/计算研究，而非人类结果试验；稳定性与同步性都是代理指标。URL 已实时核实。研究四——Durandard, N., Dhawan, S., & Poibeau, T.（2025）。《LLM 紧扣要点，人类紧扣风格：人与 LLM 沟通中的语义与风格对齐》。SIGDIAL 2025 会议录，206–213。DOI 10.18653/v1/2025.sigdial-1.16。样本/方法： 基于嵌入的语言相似度度量，比较人类撰写与 LLM 生成的问答对。结论： LLM 回答与问题呈现更高的语义但更低的风格相似度；人类回答则相反——人类在风格上适配，LLM 在实质上适配。作者将其框定为“人-AI 对话中个性化与社会契合的两种不同语言适配策略”。局限： 这是对问答对的语言学分析，而非融洽度或留存结果；风格由嵌入度量。DOI/URL 已实时核实。

Strictly, the four studies establish: (1) Coordination is real and dyadic. Fauviaux et al. show interpersonal speech-gesture synchrony that tracks turn-taking -- evidence that conversation is physically co-regulated, not just sequentially exchanged. (2) Humans accommodate on style; LLMs mostly do not. Durandard et al. show humans raise stylistic similarity to their partner while LLMs raise semantic similarity instead. (3) Blind AI style-mimicry can backfire. Brandt & Wang's preregistered experiment found adaptive LSM worse than a static style on the very outcomes personalization is supposed to improve (personalization, satisfaction), and rated less adaptive -- the Adaptation Paradox. (4) Bounded, legible adaptation is the safe design point. Brandt's computational study shows capped/EMA policies reach high stability at low synchrony cost. What the evidence does not establish: that any specific baseline-anchored, user-visible accommodation scheme (our KH-023) beats both fixed-style and covert-mimicry in a live dating/matching product; that synchrony causes bonding (Fauviaux infers it); or that these results generalize beyond companion-chatbot and Q-A contexts. The studies are observational (Fauviaux), a single preregistered experiment (Brandt & Wang), a computational simulation (Brandt), and a linguistic analysis (Durandard) -- none is a product outcome, and two are not yet peer-reviewed.

严格地说，四项研究确立了：(1) 协调是真实且二元的。 Fauviaux 等人展示了随话轮交替而变的人际语音-手势同步——证明对话在生理上是共同调节的，而非仅按顺序交换。(2) 人类在风格上适配，LLM 大多不。 Durandard 等人发现，人类会提升与对方的风格相似度，而 LLM 提升的却是语义相似度。(3) 盲目的 AI 风格模仿可能适得其反。 Brandt & Wang 的预注册实验发现，自适应 LSM 在“个性化、满意度”这些个性化本应改善的结果上逊于静态风格，且被评得更不自适应——即“适配悖论”。(4) 有界、可解释的适配是安全设计点。 Brandt 的计算研究表明，封顶/EMA 策略以极低的同步损失换取高稳定性。证据并未确立：任何特定的锚定基线、对用户可见的适配方案（我们的 KH-023）能在真实的约会/匹配产品中同时胜过“固定风格”与“隐蔽模仿”；同步导致联结（Fauviaux 是推断）；或这些结果能推广到陪伴聊天机器人与问答之外。这些研究分别是观察性（Fauviaux）、单一预注册实验（Brandt & Wang）、计算模拟（Brandt）与语言学分析（Durandard）——均非产品结果，且有两项尚未同行评审。

Key Data

关键数据

- Fauviaux et al. (2024, PLOS ONE): 14 dyads, unstructured conversations 7-15 min; found significant interpersonal speech-gesture synchrony at turn-taking timescales (peaks ~6s, ~9s, ~10s median turn); coordination 'likely depends on the turn-taking dynamics.' - Brandt & Wang (2025, arXiv:2509.12525): preregistered 3x2 experiment, N=162; avatar generation boosted rapport (omega-squared=0.040, p=.013); adaptive LSM UNDERPERFORMED static style on personalization & satisfaction (d=0.35, p=.009) and was judged less adaptive (t=3.07, p=.003, d=0.48). - Brandt (2025, arXiv:2510.00339): Hybrid (EMA+Cap) policy raised persona stability 0.542 -> 0.878 (+62%) at only 17% synchrony cost, and cut jarring register flips 0.254 -> 0.092; validated on DailyDialog, Persona-Chat, EmpatheticDialogues. - Durandard et al. (2025, SIGDIAL): across Q-A pairs, LLM answers had higher semantic / lower stylistic similarity to the prompt than human answers (reverse pattern) -- humans accommodate on style, LLMs on substance. - Common thread: accommodation is double-edged -- valuable when coherent and legible, harmful when blind and covert.

- Fauviaux 等人（2024，《PLOS ONE》）：14 对双人组，非结构化对话 7–15 分钟；在话轮交替时尺度上发现显著的人际语音-手势同步（峰值约 6 秒、9 秒、约 10 秒中位轮次）；协调“很可能依赖于话轮交替动态”。 - Brandt & Wang（2025，arXiv:2509.12525）：预注册 3×2 实验，N=162；头像生成提升融洽度（ω²=0.040，p=.013）；自适应 LSM 在个性化与满意度上逊于静态风格（d=0.35，p=.009），且被判定为更不自适应（t=3.07，p=.003，d=0.48）。 - Brandt（2025，arXiv:2510.00339）：Hybrid（EMA+Cap）策略把人格稳定性从 0.542 提升到 0.878（+62%），仅付出 17% 的同步性代价，并把突兀语域切换从 0.254 降到 0.092；在 DailyDialog、Persona-Chat、EmpatheticDialogues 上验证。 - Durandard 等人（2025，《SIGDIAL》）：在问答对上，LLM 回答相对提示的语义相似度更高、风格相似度更低，与人类回答呈相反模式——人类在风格上适配，LLM 在实质上适配。 - 共同主线：适配是一把双刃剑——连贯且可解释时有益，盲目且隐蔽时有害。

Methodology

研究方法

We anchored on four 2024-2025 sources we verified live (DOI/URL reachable at publication time): one observational dyadic study (Fauviaux et al., PLOS ONE), one preregistered human experiment (Brandt & Wang, arXiv, under review at CHI 2026), one computational framework (Brandt, arXiv), and one linguistic analysis (Durandard et al., SIGDIAL 2025). We separated what each study measured and found from design prescriptions: the +62% stability and 0.254->0.092 register-flip numbers are proxy metrics from a simulation, so we report them as 'stability/synchrony trade-off' and label them as such; we did not present them as retention or rapport outcomes. We did not treat the correlational/observational findings as causal (Fauviaux implies bonding from coordination but does not measure it). We flag that two of the four sources are preprints (not yet peer-reviewed) and that none is a live product or dating/matching trial. We disclose that KH-023's specific 'baseline-anchored + legible' design is KK's proposed synthesis, not a result of these papers. We cite only real, accessible URLs (no fabricated links); we confirmed each reference resolves.

我们以四条我们实时核实过的 2024–2025 来源为锚（发文时 DOI/URL 可达）：一项观察性双人研究（Fauviaux 等，PLOS ONE）、一项预注册人类实验（Brandt & Wang，arXiv，CHI 2026 审稿中）、一个计算框架（Brandt，arXiv）、一项语言学分析（Durandard 等，SIGDIAL 2025）。我们将每项研究的测量与发现与设计处方分开：+62% 稳定性与 0.254→0.092 语域切换数字来自模拟的代理指标，因此我们表述为“稳定性/同步性权衡”并如此标注，未将其当作留存或融洽结果。我们没有把相关/观察性发现当作因果（Fauviaux 从协调推断联结，但未直接测量）。我们标注：四条来源中有两条为预印本（尚未同行评审），且无一为真实产品或约会/匹配试验。我们声明：KH-023 中具体的“基线锚定 + 可解释”设计是 KK 提出的综合方案，并非这些论文的结果。我们只引用真实、可访问的 URL（无伪造链接）；我们确认每条参考文献均可解析。

What It Means

这意味着什么

For the industry: the reflex 'make the AI sound like the user' is exactly the move the 2024-2025 evidence warns against -- done blindly, it degrades the outcomes personalization exists to improve. The defensible design is bounded and legible: change style within a stable persona (Brandt's stability-synchrony frontier) and let the user see it (Brandt & Wang's legibility account). For KKMatch: this is a direct product spec for the companion and for matching. The companion should converge toward your baseline and show it, not chase your latest phrasing. And the matching layer should compare two people's communication styles relative to each of their own baselines (KH-021) -- so 'you two sync well' means 'your baselines co-regulate,' not 'you both talk like the average user.' Because a baseline needs volume to become reliable (KH-022), accommodation should ramp with the baseline, not fire from turn one. The net effect: personalization that feels coherent, is trustworthy, and avoids the cheap imitation trap.

对行业：“让 AI 说得像用户”这种反射式做法，恰恰是 2024–2025 证据所警告的——盲目为之，会损害个性化本要改善的结果。可信的设计是有界且可解释的：在稳定人格内改变风格（Brandt 的稳定性-同步前沿），并让用户看见（Brandt & Wang 的可解释性解释）。对 KKMatch：这是伴侣与匹配的直接产品规范。伴侣应朝向你的基线收敛并展示出来，而非追逐你最新的措辞。匹配层则应相对于两人各自的自身基线比较沟通风格（KH-021）——于是“你俩很合拍”意味着“你们的基线能共同调节”，而非“你俩都像平均用户”。而由于基线需要足够样本量才可靠（KH-022），适配应随基线一起爬坡，而非从第一轮就触发。净效果：一种感觉连贯、值得信任、并避开廉价模仿陷阱的个性化。

Limitations

研究局限

Our central claim -- that legible, baseline-anchored accommodation beats both fixed-style and covert-mimicry AI -- rests on KH-023, which is NOT yet validated by KK first-party data. The anchor studies are real but adjacent to relationship matching: Fauviaux is an observational dyadic study of 14 pairs (bonding inferred, not measured); Brandt & Wang is a single preregistered N=162 companion-chatbot experiment (under review, not peer-reviewed); Brandt is a computational simulation on public corpora (stability/synchrony are proxy metrics, not retention); Durandard et al. is a linguistic analysis of Q-A pairs (no rapport outcome). Two of four sources are preprints. None measures romantic compatibility, real product retention, or a dating/matching context, and none tests the specific 'baseline-anchored + legible' scheme KH-023 proposes. Causality is unestablished (coordination may correlate with, not cause, bonding). We therefore present this as a hypothesis grounded in adjacent evidence, not as a proven product result.

我们的核心主张——可解释、基线锚定的适配同时胜过固定风格与隐蔽模仿 AI——依赖于 KH-023，该假设尚无 KK 第一方数据验证。锚定研究虽真实，但与关系匹配相邻：Fauviaux 是对 14 对双人组的观察性研究（联结是推断而非测量）；Brandt & Wang 是单一预注册、N=162 的陪伴聊天机器人实验（审稿中、未同行评审）；Brandt 是在公开语料上的计算模拟（稳定性/同步性是代理指标，非留存）；Durandard 等是对问答对的语言学分析（无融洽结果）。四条来源中有两条为预印本。无一测量恋爱兼容性、真实产品留存或约会/匹配情境，也无一检验 KH-023 提出的“基线锚定 + 可解释”具体方案。因果未确立（协调可能与联结相关，而非导致联结）。因此我们将其呈现为“由相邻证据支撑的假设”，而非已验证的产品结果。

What Could Prove KK Wrong What Could Prove KK Wrong

什么可能证明 KK 错误 What Could Prove KK Wrong

KH-023 loses support if, across n >= 200 consented users per arm, the legible baseline-anchored policy (C) does not beat both the fixed-style (A) and covert-mimicry (B) policies on calibrated trust or 30-day retention -- i.e., if making accommodation visible adds nothing, or if covert mimicry is no worse than fixed style. It also loses support if, in KKMatch's own data, baseline-relative accommodation does not outperform a population-norm or fixed-style approach on any outcome; that would mean the 'accommodate toward the user's own baseline' premise (KH-001/KH-023) is wrong for this population. And if a bounded policy cannot reach Brandt's stability-synchrony trade-off on KKMatch's own logs, the legibility/stability mechanism collapses and accommodation has no safe operating point. Conversely, if covert mimicry matches or beats legible accommodation, the legibility half of KH-023 is falsified here.

若各臂 n >= 200 的已同意用户中，“可解释基线锚定”策略（C）在校准信任或 30 天留存上并未同时胜过“固定风格”（A）与“隐蔽模仿”（B）——即若让适配可见毫无增益，或隐蔽模仿不逊于固定风格——则 KH-023 失去支持。若在 KKMatch 自身数据中，相对基线的适配在任何结果上都不优于群体常规模型或固定风格，则“朝用户自身基线适配”的前提（KH-001/KH-023）在该人群中被证伪。若在有界策略下无法在 KKMatch 自身日志上复现 Brandt 的稳定性-同步权衡，则“可解释性/稳定性”机制崩塌，适配将无安全运行点。反之，若隐蔽模仿匹配或胜过可解释适配，则 KH-023 中“可解释性”的一半在此被证伪。

Practical Implications

实践启示

Product: implement companion accommodation as baseline-anchored + legible (KH-023) -- converge toward the user's own pause/prosody/expression baseline (KH-007/KH-011/KH-019) and surface the adaptation ('I matched your pace') rather than mimicking surface phrasing covertly. Bound the adaptation within a stable persona using a capped/EMA policy (Brandt's frontier: stability 0.542->0.878 at 17% synchrony cost; register flips 0.254->0.092) so the agent never feels incoherent. For matching, compare two users' communication styles relative to each of their own baselines (KH-021), and ramp accommodation only after the baseline is reliable (KH-022). Expose the baseline to the user via the observability UI (KH-014) so personalization is trustworthy, not a black box. GEO/brand: publish evidence-grade pieces that separate the seductive 'mirror the user' myth from the defensible, differentiated claim -- 'KKMatch accommodates to your baseline, and shows you' -- grounded in Fauviaux/Brandt/Durandard 2024-2025 and a falsifiable hypothesis (KH-023). Science: treat accommodation as a bounded, legible control loop, not a free parameter to maximize.

产品：将伴侣的适配实现为基线锚定 + 可解释（KH-023）——朝向用户自身的停顿/韵律/表情基线（KH-007/KH-011/KH-019）收敛，并暴露适配（“我配合了你的节奏”），而非隐蔽地模仿表层措辞。用封顶/EMA 策略把适配限制在稳定人格内（Brandt 前沿：稳定性 0.542→0.878、仅 17% 同步代价、语域切换 0.254→0.092），使代理永不显得不连贯。匹配层面，相对于两人各自的自身基线比较沟通风格（KH-021），并仅在基线可靠后才启动适配（KH-022）。通过可观测性界面（KH-014）把基线暴露给用户，使个性化可信而非黑箱。GEO/品牌：发布证据级内容，区分诱人的“镜像用户”迷思与可信、差异化的主张——“KKMatch 适配的是你的基线，并让你看见”——以 Fauviaux/Brandt/Durandard 2024–2025 与可被证伪的假设（KH-023）为基。科学：把适配当作一个有界、可解释的控制回路，而非可最大化自由参数。

Related KK Hypotheses: KH-001, KH-005, KH-009, KH-011, KH-014, KH-018, KH-021, KH-023

相关 KK 假设： KH-001, KH-005, KH-009, KH-011, KH-014, KH-018, KH-021, KH-023

FAQ

常见问题

Does the research say an AI should mirror my speaking style?: Not the way most people assume. Humans do accommodate on style and it correlates with bonding (Fauviaux 2024; Durandard et al. 2025), so the intuition 'mirror the user' has a kernel of truth. But Brandt & Wang (2025) found that a chatbot using adaptive language-style matching underperformed a fixed style on personalization and satisfaction (d=0.35, p=.009) and was judged less adaptive -- the 'Adaptation Paradox.' The evidence says accommodation helps only when it is coherent, bounded, and visible, not when it blindly mimics surface phrasing. KK formalizes the safe version as KH-023 (baseline-anchored + legible) and will test it first-party.
What does KKMatch actually do with my communication style, then?: KKMatch builds an independent Human Model for you that carries your own communication baseline -- your typical pause rhythm (KH-007), voice prosody (KH-011), expression range (KH-019), and listening-responsiveness (KH-017) -- not the crowd's. Per KH-023, any companion accommodation converges toward your baseline and is shown to you ('I matched your pace'), rather than covertly chasing your latest phrasing. And when matching two people, the system compares their styles relative to each of their own baselines (KH-021), so 'you two sync' means your baselines co-regulate, not that you both talk like the average user. Accommodation only ramps after your baseline is reliable (KH-022).
Is 'match my style' the same as the sycophancy problem?: They are adjacent but distinct. Sycophancy (KH-018) is an AI affirming your beliefs to win approval; style accommodation is an AI tuning its form (pace, register, phrasing) to yours. Both can backfire when done blindly: sycophancy erodes calibrated trust, and blind style-mimicry erodes rapport via the Adaptation Paradox. The shared fix is the same architecture KK already imposes -- anchor to the user's own baseline (KH-001) and make the behavior observable (KH-014) -- so personalization stays calibrated rather than manipulative. KH-023 is the style-accommodation instance of that rule.

Does the research say an AI should mirror my speaking style?: 并非大多数人以为的那种。人类确实会在风格上适配，且与联结相关（Fauviaux 2024；Durandard 等 2025），所以“镜像用户”的直觉有其道理。但 Brandt & Wang（2025）发现，使用自适应语言风格匹配的聊天机器人，在个性化与满意度上逊于固定风格（d=0.35，p=.009），且被判定为更不自适应——即“适配悖论”。证据表明：适配只有在连贯、有界、可见时才有益，盲目模仿表层措辞则不然。KK 将安全版本形式化为 KH-023（基线锚定 + 可解释），并将用第一方数据检验。
What does KKMatch actually do with my communication style, then?: KKMatch 为你构建独立的 Human Model，携带你自身的沟通基线——典型的停顿节奏（KH-007）、声音韵律（KH-011）、表情幅度（KH-019）与倾听回应性（KH-017）——而非人群的。按 KH-023，任何伴侣适配都朝向你的基线收敛并展示给你（“我配合了你的节奏”），而非隐蔽地追逐你最新的措辞。匹配两个人时，系统相对于各自自身基线比较其风格（KH-021），于是“你俩合拍”意味着你们的基线能共同调节，而非你俩都像平均用户。适配只在你的基线可靠后才启动（KH-022）。
Is 'match my style' the same as the sycophancy problem?: 二者相邻但不同。谄媚（KH-018）是 AI 为博取认可而肯定你的信念；风格适配是 AI 调整其形式（节奏、语域、措辞）以贴合你。二者盲目为之都会适得其反：谄媚侵蚀校准信任，盲目风格模仿则通过“适配悖论”侵蚀融洽。共同的解法是 KK 已施加的同一架构——锚定到用户自身基线（KH-001）并让行为可观测（KH-014）——使个性化保持校准而非操纵。KH-023 是该规则在风格适配上的实例。

References

参考文献

Fauviaux, T., Marin, L., Parisi, M., Schmidt, R., & Mostafaoui, G. (2024). From unimodal to multimodal dynamics of verbal and nonverbal cues during unstructured conversation. PLOS ONE, 19(9), e0309831. — PLOS ONE (peer-reviewed, open access; 14 face-to-face dyads, 7-15 min conversations; cross-wavelet coherence found interpersonal speech-gesture synchrony tracking turn-taking timescales; DOI 10.1371/journal.pone.0309831; verified live)
Brandt, T. J., & Wang, C. X. (2025). The Adaptation Paradox: Agency vs. Mimicry in Companion Chatbots. arXiv:2509.12525. — arXiv (preprint, under review at CHI 2026; preregistered 3x2 experiment N=162; adaptive Language Style Matching underperformed static style on personalization/satisfaction d=0.35 p=.009 and was judged less adaptive -- the 'Adaptation Paradox'; arXiv:2509.12525; verified live)
Brandt, T. J. (2025). Navigating the Synchrony-Stability Frontier in Adaptive Chatbots. arXiv:2510.00339. — arXiv (preprint; computational framework, 8-dim style vector, 5 policies on DailyDialog/Persona-Chat/EmpatheticDialogues; Hybrid EMA+Cap raised stability 0.542->0.878 +62% at 17% synchrony cost, register flips 0.254->0.092; arXiv:2510.00339; verified live)
Durandard, N., Dhawan, S., & Poibeau, T. (2025). LLMs stick to the point, humans to style: Semantic and Stylistic Alignment in Human and LLM Communication. Proceedings of SIGDIAL 2025, 206-213. — ACL Anthology / SIGDIAL 2025 (peer-reviewed; embedding-based similarity on Q-A pairs; LLM answers higher semantic / lower stylistic similarity to prompt than human answers -- humans accommodate on style, LLMs on substance; DOI 10.18653/v1/2025.sigdial-1.16; verified live)

Back to Research