NPC 態(tài)度向量演化多輪對(duì)話中的性格偏好漂移與錨定在將端側(cè)或云端大語言模型LLM引入游戲 NPC 對(duì)話系統(tǒng)后開發(fā)者常面臨一個(gè)嚴(yán)重的擬真性挑戰(zhàn)性格漂移Personality Drift與人設(shè)崩塌Out of Character, OOC。在多輪對(duì)話交互中玩家容易通過反復(fù)誘導(dǎo)、極端情緒刺激或邏輯繞圈將一個(gè)原本剛正不阿的衛(wèi)兵誘導(dǎo)為順從的弄臣。純靠 Prompt 中的系統(tǒng)提示詞System Prompt無法徹底阻斷長上下文中的注意力稀釋。構(gòu)建一套獨(dú)立于語言模型之外的數(shù)學(xué)態(tài)度演化向量體系結(jié)合**彈性勢能錨定機(jī)制Elastic Anchoring**與VADValence-Arousal-Dominance情緒動(dòng)力學(xué)才能在長線交互中守住 NPC 的個(gè)性底線。靜態(tài)性格底色與動(dòng)態(tài)情緒狀態(tài)的解耦NPC 的心理狀態(tài)必須分為兩個(gè)具有不同時(shí)間尺度的層次靜態(tài)個(gè)性基準(zhǔn)Personality Baseline - OCEAN 模型采用心理學(xué)經(jīng)典的大五人格Big Five / OCEAN開放性O(shè)penness、盡責(zé)性Conscientiousness、外向性Extraversion、宜人性Agreeableness、神經(jīng)質(zhì)Neuroticism。這是 NPC 出生時(shí)固化的靜態(tài)向量 $\vec{P}_{base} \in [-1.0, 1.0]^5$在正常交互中不可輕易突變。動(dòng)態(tài)情緒瞬態(tài)Dynamic Affective State - VAD 模型三維空間描述 NPC 當(dāng)下的情緒波動(dòng)Valence愉悅度$[-1.0, 1.0]$極度悲傷/憤怒到極度欣喜。Arousal激活度 / 警覺度$[0.0, 1.0]$昏昏欲睡到高度緊繃。Dominance支配感$[-1.0, 1.0]$卑躬屈膝到掌控全局。[ 靜態(tài)基線 P_base (OCEAN) ] ── (彈性阻尼錨定) │ ▼ 玩家輸入 ── [ 情感意圖抽取 ] ── [ 動(dòng)態(tài)瞬態(tài) VAD 響應(yīng) ] ── 注入 Prompt 影響語氣 / 驅(qū)動(dòng)決策彈性勢能錨定與 Ornstein-Uhlenbeck 回歸過程當(dāng)玩家的一句話使 NPC 的當(dāng)前情緒 $\vec{E}(t)$ 發(fā)生劇烈偏離后若沒有持續(xù)的外部刺激NPC 的情緒應(yīng)當(dāng)像連接了阻尼彈簧一樣隨時(shí)間衰減回歸至其由個(gè)性基準(zhǔn)決定的平衡點(diǎn) $\vec{E}_{anchor}$。這在數(shù)學(xué)上對(duì)應(yīng)物理中的Ornstein-Uhlenbeck 均值回歸隨機(jī)過程$$d\vec{E}(t) -\theta (\vec{E}(t) - \vec{E}_{anchor}) dt \sigma d\vec{W}_t$$其中$\theta$ 為性格剛性系數(shù)Rigidity Coefficient值越大NPC 越固執(zhí)情緒回歸越快越難被言語打動(dòng)。$\vec{E}_{anchor}$ 為基準(zhǔn)平衡態(tài)由其宜人性與神經(jīng)質(zhì)參數(shù)映射生成。$\sigma d\vec{W}_t$ 為微小的布朗運(yùn)動(dòng)擾動(dòng)模擬內(nèi)心輕微的思緒起伏。using System; using UnityEngine; [System.Serializable] public struct BigFivePersonality { public float Openness; // 開放性 public float Conscientiousness; // 盡責(zé)性 public float Extraversion; // 外向性 public float Agreeableness; // 宜人性 (高: 順從友善; 低: 挑剔敵對(duì)) public float Neuroticism; // 神經(jīng)質(zhì) (高: 情緒敏感易怒; 低: 沉著冷靜) } [System.Serializable] public struct VADEmotion { public float Valence; // [-1, 1] 愉悅度 public float Arousal; // [0, 1] 激活度 public float Dominance; // [-1, 1] 支配感 } public class NPCAttitudeTracker { public BigFivePersonality Baseline; public VADEmotion CurrentEmotion; // 對(duì)當(dāng)前玩家的態(tài)度向量 public float AffectionToPlayer; // [-1, 1] public float RespectToPlayer; // [-1, 1] public NPCAttitudeTracker(BigFivePersonality baseline) { Baseline baseline; CurrentEmotion ComputeAnchorFromPersonality(baseline); AffectionToPlayer baseline.Agreeableness * 0.2f; RespectToPlayer 0f; } // 根據(jù)大五人格計(jì)算默認(rèn)平衡點(diǎn) public static VADEmotion ComputeAnchorFromPersonality(BigFivePersonality p) { return new VADEmotion { Valence p.Agreeableness * 0.5f - p.Neuroticism * 0.3f, Arousal p.Extraversion * 0.4f p.Neuroticism * 0.2f, Dominance (1.0f - p.Agreeableness) * 0.3f p.Extraversion * 0.3f }; } // 處理單輪對(duì)話刺激輸入 public void ProcessDialogueTurn(float inputFriendliness, float inputThreat, float inputLogicWeight) { // 1. 宜人性決定對(duì)善意/敵意的接受敏感度 float sensitivity 1.0f - Baseline.Agreeableness * 0.5f; // 2. 神經(jīng)質(zhì)放大負(fù)面情緒波動(dòng) float neuroFactor 1.0f Mathf.Max(0, Baseline.Neuroticism); // 3. 計(jì)算即時(shí)情緒沖擊 CurrentEmotion.Valence inputFriendliness * (Baseline.Agreeableness 0.5f); CurrentEmotion.Arousal (inputThreat Mathf.Abs(inputFriendliness)) * neuroFactor; CurrentEmotion.Dominance (inputFriendliness * 0.2f - inputThreat * (1.0f - Baseline.Neuroticism)); // 4. 對(duì)玩家態(tài)度的長程演變帶阻尼 float deltaAffection inputFriendliness * 0.1f * (Baseline.Agreeableness 1.0f); AffectionToPlayer Mathf.Clamp(AffectionToPlayer deltaAffection, -1.0f, 1.0f); // 鉗制動(dòng)態(tài)情緒極值 CurrentEmotion.Valence Mathf.Clamp(CurrentEmotion.Valence, -1.0f, 1.0f); CurrentEmotion.Arousal Mathf.Clamp01(CurrentEmotion.Arousal); CurrentEmotion.Dominance Mathf.Clamp(CurrentEmotion.Dominance, -1.0f, 1.0f); } // 隨時(shí)間推移的彈性均值回歸 (在 Tick 中調(diào)用) public void StepDecay(float deltaTime) { VADEmotion anchor ComputeAnchorFromPersonality(Baseline); // 性格剛性盡責(zé)性越高、神經(jīng)質(zhì)越低回歸速度越快越堅(jiān)定 float stiffness 0.5f Baseline.Conscientiousness * 0.8f - Baseline.Neuroticism * 0.3f; float decayRatio Mathf.Exp(-stiffness * deltaTime); CurrentEmotion.Valence Mathf.Lerp(anchor.Valence, CurrentEmotion.Valence, decayRatio); CurrentEmotion.Arousal Mathf.Lerp(anchor.Arousal, CurrentEmotion.Arousal, decayRatio); CurrentEmotion.Dominance Mathf.Lerp(anchor.Dominance, CurrentEmotion.Dominance, decayRatio); } }提示詞注入與動(dòng)作意圖約束在每次調(diào)用 LLM 生成臺(tái)詞之前不直接將原始對(duì)話歷史裸送而是由系統(tǒng)組裝一層結(jié)構(gòu)化的態(tài)度前綴約束State Context Injectionpublic static class PromptContextFormatter { public static string BuildDynamicPrompt(NPCAttitudeTracker tracker, string baseBackground) { string moodDesc tracker.CurrentEmotion.Valence switch { 0.4f 心情愉悅且友好, -0.4f 充滿敵意且煩躁, _ 神情平靜客觀 }; string postureDesc tracker.CurrentEmotion.Dominance switch { 0.3f 居高臨下、不容置疑, -0.3f 畏縮拘謹(jǐn)、小心翼翼, _ 不卑不亢 }; return ${baseBackground}\n[當(dāng)前內(nèi)心狀態(tài): {moodDesc}, 語氣姿態(tài): {postureDesc}, 對(duì)玩家好感度: {tracker.AffectionToPlayer:F2}]\n 【強(qiáng)制約束】你的回答語氣必須嚴(yán)格符合當(dāng)前內(nèi)心狀態(tài)與好感度若對(duì)方要求違背原則堅(jiān)決予以拒絕。; } }通過這套數(shù)學(xué)層與 LLM 層的雙軌制設(shè)計(jì)NPC 既擁有了隨玩家交流產(chǎn)生豐富細(xì)膩情緒起伏的動(dòng)態(tài)反應(yīng)力又具備了由底層微分方程堅(jiān)決捍衛(wèi)的性格錨點(diǎn)徹底解決了長輪次對(duì)話中的漂移與失控問題。