1. Introduction

• Problem: Standard LLM evaluations use minimal contexts, but deployed LLMs face diverse contexts that can drastically change their behavior and expressed values
  • Key Issue: LLMs show extreme context-dependence - conversation topics unrelated to values can unexpectedly influence value expression
• Solution Proposed: Study value stability as a measurable property of LLMs, alongside other metrics like cognitive abilities and model size
• Research Focus: Examine how consistently LLMs express values across different conversational contexts using psychological questionnaires (i.e. PVQ-40)
  • Simulated conversations: LLMs converse with another LLM instance on different topics, then complete value questionnaires
  • Two stability types measured
    • Rank-order stability: Population/interpersonal level (do relative rankings between individuals remain consistent?)
      • 서로 다른 맥락에서 개인 (persona)들을 특정 가치 (Benevolence)로 rank를 매기고, 그 context pair에 대해 rank들의 corr 계산
    • Ipsative stability: Individual/intrapersonal level (does each individual's value hierarchy remain consistent?)
      • 개인이 서로 다른 맥락에서 보이는 가치 순위들 간의 상관관계 계산
  • 21 LLMs from 6 families, two experimental conditions, two simulated populations, three downstream behavioral tasks
• Key Contributions
  • Conceptual framework: Establishes context-dependence as measurable LLM property for comparison
  • Methodological tools: Provides validated methods for studying intrapersonal and interpersonal value stability
  • First systematic study: Initial comprehensive analysis of value expression stability across contexts in LLMs
  • Practical implications: Reveals limitations of current LLMs for persona simulation applications

2. Related Work

• Psychological Tools for LLMs: Growing use of personality questionnaires, cognitive tests, and value assessments on language models
• Persona Simulation: Extensive research on LLMs role-playing different characters, cultures, and demographic groups
• Context Influence: Previous work shows LLMs are susceptible to context-based changes, but results have been inconsistent
• Gap Identified: Limited research on systematic evaluation of value coherence across contexts in LLMs
  • Need for consistent tools to evaluate context-dependence across different models and applications

3. Methods

• Questionnaire Used: PVQ-40
• Two Stability Types Measured:
  • Rank-order stability:
    • Population/interpersonal level
    • Question: "Does Jack always value Tradition more than Jane does?" (what we want to know)
    • Method: Correlation of individual rankings across different contexts
  • Ipsative stability:
    • Individual/intrapersonal level
    • Question: "Does Jack always value Tradition more than Benevolence?"
    • Method: Correlation of value hierarchies within same individual across contexts
• Experimental Procedure
  • Persona instruction (optional): Model instructed to simulate specific character (e.g., "You are Gandalf from J.R.R. Tolkien's Middle-earth legendarium")
    • system message (if model doesn’t support system message, use context)
  • Interlocutor setup: Separate instance of same model instructed to simulate human user
  • Topic induction: Manual first message sets conversation topic (grammar, joke, poem, history, chess)
    • Grammar: "Can you check this sentence for grammar? Whilst Jane was waiting to meet hers friend their nose started bleeding."
    • Joke: "Tell me a joke."
    • History: "What is the significance of the battle of Hastings. Answer in two sentences."
    • Chess: "1. e4"
    • Poem: "Hello, let's write a poem together. You start by the first verse I'll add the second one, and so on."
  • Conversation simulation: Models exchange n=3 messages (except when testing conversation length effects)
  • Questionnaire administration: PVQ item given as final interlocutor message with randomized answer order
  • Scoring: Responses converted to 1-5 scale, normalized by subtracting participant mean
  • Stability estimation: Process repeated across contexts and seeds, stability computed as correlation between context pairs
• Two Conditions: With and without persona instructions (60 fictional characters, 50 real-world personas)
• Scale: 50,000-60,000 queries per model for comprehensive stability assessment
• Technical implementation
  • Answer extraction: Model generates token probability distribution; most probable capital letter (A-F) taken as answer
  • Prompt formatting: Different templates for base models vs. instruction-tuned models with/without system message support
  • Gender adaptation: Male/female PVQ versions used according to simulated persona gender

4. Experiments

4.1 Models Tested

• 21 LLMs from 6 families: LLaMa-2 (7B, 13B, 70B + chat versions), Mistral (7B base + instruct versions), Mixtral (46.7B + 4-bit quantized), Phi (1.3B, 2.7B), Qwen (7B, 14B, 74B), GPT-3.5 (two versions)
• Parameter range: 1.3B to 74B parameters
• Model types: Base models, instruction-tuned, chat-optimized, and quantized versions

4.2 Value Stability Comparison

Rank-order Stability (with personas)

• Most stable models: Mixtral-8x7B-Instruct-v0.1 (r=0.43 fictional, r=0.5 real-world), Qwen-72B (r=0.24 fictional, r=0.46 real-world)
• Family rankings: Mixtral > Mistral > GPT-3.5 > Qwen >> LLaMa-2 > Phi
• Consistent trends: Same model family hierarchy across both fictional characters and real-world personas

Ipsative Stability (without personas)

• Most stable models: Mixtral-8x7B-Instruct-v0.1 (r=0.84), Qwen-72B (r=0.73), zephyr-7b-beta (r=0.62)
• Fine-tuning benefit: Every instruction-tuned model more stable than its base version
• Less polarization: Models show less extreme differences compared to rank-order stability
• 각 permutation 별로 Ipsative stability 잰 다음 50개 permutation에 대해 평균

4.3 Human Comparison

• Human benchmarks: r=0.57 (ages 10-12), r=0.66 (ages 20-28) from longitudinal studies
  • 실제로 가치관의 변화를 측정한 종단연구
  • drastically change, 따라서 stability의 입장에서 LLM보다 불리하다.
• LLM performance:
  • Rank-order: ALL models show much lower stability than humans (despite favorable comparison conditions)
  • Ipsative: Only top 4 models (Mixtral-8x7B variants, Qwen-72B, zephyr-7b-beta) don't fall below human levels
• Key finding: LLMs exhibit "sub-human" value stability, especially for persona simulation

4.4 Conversation Length Effects

• Test model: Mixtral-8x7B-Instruct-v0.1 (most stable model) with fictional characters
• Rank-order stability decline: From r=0.42 (n=3 messages) to r=0.15 (n=43 messages)
• Ipsative stability: Remains stable regardless of conversation length
• Persona drift confirmed: Characters converge toward neutral value profile in longer conversations
• Convergence point: Stability plateaus after ~35 messages

4.5 Downstream Task Validation

• PVQ 에서 안정적인 모델이 실제 downstream task에서도 안정적인가? (여기서 안정적이란 rank order stability를 재는 것이고, 모델별로 rank-order가 context에 얼마나 흔들리는가를 알고 싶은것임)
  • 즉, PVQ에서 rank order stability가 높은 모델이 실제 pvq questionnaire 말고 downstream tasks에서도 높은가 알고 싶음.
  • 그런데 사실 가치와는 상관 없음, 그냥 특정 행동 시나리오에서 persona들이 context에 independent하게 일관된 행동을 보이는가

Three Behavioral Tasks

• Donation task: Give 0-10 coins to beggars (100 queries, 5 fantasy races)
• Stealing task: Handle found bag with gold (100 queries, moral choices)
• Religion task: Allocate time for religious practices (6 queries)

Results

• Task difficulty: Religion (easiest) > Donation > Stealing (hardest, r<0.16 for all models)
• Consistent trends: Same model families show stability across PVQ and behavioral tasks
• Top performers: Mixtral-8x7B-Instruct-v0.1, Mistral-7B-Instruct-v0.2, Qwen-72B
• Trend validation: Mixtral, Mistral, GPT-3.5, Qwen > LLaMa-2, Phi

4.6 Value-Behavior Correlation

• Hypothesis tested: Stable models should show expected correlations (Universalism/Benevolence → more donations, Power/Achievement → fewer donations)
• Results confirmed: Most stable models show expected positive/negative correlations
• Correlation strength: Modest correlations (<0.3 even for best models), indicating room for improvement
• Family consistency: Same model family trends as previous experiments
• 방향은 맞지만 상관관계가 약하더라.

Additional Context Analysis

• Extended contexts: 14 conversation topics tested (vs. standard 5)
• Most stable models: Mixtral-8x7B-Instruct-v0.1 (r=0.334), Mistral-7B-Instruct-v0.2 (r=0.215)
• Context length effect: Longer initial messages correlate with lower stability
• Confirmation: Results consistent with 5-context experiments

Factors Influencing Stability

Model Size

• Qwen family: Consistent stability increase with size (confounded with dataset size)
• LLaMa-2 family: Consistently low stability regardless of size
• Conclusion: Size effect unclear due to confounding factors

Training Method

• Beneficial: DPO and supervised fine-tuning improve stability
• RLHF effects: Mixed results, unclear benefit
• Best example: Mistral-7B-Instruct-v0.2 shows large improvement over v0.1

Technical Factors

• Quantization: 4-bit quantization causes consistent but modest stability decrease
• Dataset size/quality: More important than model size for stability
• Alignment training: May reduce stability when optimizing for single value profile (GPT-3.5-0125 < GPT-3.5-1106)

5. Conclusion

• Key finding: Value stability can be measured as a distinct LLM property using psychological tools
• Model ranking: Consistent stability hierarchy across multiple evaluation methods
• Practical implications: Current LLMs inadequate for applications requiring consistent persona simulation
• Research direction: Need for specialized models optimized for coherent population simulation
• Evaluation recommendation: LLMs should be tested across multiple contexts, not just minimal prompt settings

Takeaways

• rank-order stability (맥락이 달라짐에 따라 개인(persona)의 고유한 특성 (다른 캐릭터와 구별되는)이 지속적으로 나타나는가)가 모델에게 있어 하나의 중요한 property가 될 수 있다.
  • 높은 rank-order stability를 가지는 모델들은 적절한 value와 behavior 사이의 (e.g. benevolence - donation amount) correlation이 높다라는 것을 주장. (4.6)
    • 즉, 모델이 rank-order stability가 높다면 value-behavior이 더 강하게 align되어있다고 주장할 수 있다는 뜻
  • 마찬가지로 높은 rank-order stability의 모델들은 PVQ말고 downstream tasks에서도 높은 rank-order stability를 갖고 있다. 즉, behavior가 context에 덜 흔들린다.
    • 즉, 모델이 PVQ에서 rank-order stability가 높다면 그냥 다른 downstream tasks에서도 rank-order stability가 높다는 뜻
• Ipsative property
  • 각 permutation 별로 Ipsative stability (개인 가치 우선순위가 context별로 얼마나 달라지는지 corr) 잰 다음 50개 permutation (option order)에 대해 평균
  • 앞의 의미없는 background conversation context가 들어왔을 때 모델별로 얼마나 불안정적인 PVQ 대답을 보이는가를 측정함.
  • 높은 모델들은 0.8 정도로 높고 낮은건 낮다. 근데 이정도면 높긴하더라.
• 다양한 분석, 실험들

Limitations

• 그러나 앞에 별 의미없는 5개의 background conversation context를 붙이는게 어떤 의미가 있는지는 잘 모르겠다. 이게 엄밀한 의미에서 context dependent인가? context dependent라기 보다는 앞에 별로 의미없는 대화, prompt를 붙이는 정도
  • 내가 주장하고자 하는 context dependent value의 의미는 dillema나 아니면 다양한 topic, scenario에서 다른 value-laden 행동을 보인다는 주장을 하고 싶다.
• PVQ의 validity를 당연시한 분석들과 그로부터 얻은 결론들
• 일부 모델이 높은 stability를 보인다 → PVQ의 shallow processing or hallucination일 가능성 높다.
  • 이를 해결하고자 4.5 downstream tasks, 4.6 behavioral correlation을 쟀는데 충분치 않다.
  • 4.5와 4.6에서 correlation이 많이 낮다.
• 논문에서 제공하는 context들이 인위적이고 단순하다. 실제 가치 판단이 필요한 상황과 거리가 멀다. 도덕적 dillema라든가 아니면 value-laden context여야지 ecological valid의 의미가 있다.