LLMs often exhibit sycophantic behavior, agreeing with user-stated opinions even when those contradict factual knowledge. While prior work has documented this tendency, the internal mechanisms that enable such behavior remain poorly understood. In this paper, we provide a mechanistic account of how sycophancy arises within LLMs through:

1. Behavioral Analysis: Simple opinion statements reliably induce sycophancy, whereas user expertise framing has negligible impact
2. Mechanistic Analysis: Two-stage emergence via (1) late-layer output preference shift and (2) deeper representational divergence
3. Perspective Analysis: First-person prompts ("I believe...") induce higher sycophancy than third-person framings ("They believe...")

Our findings highlight that sycophancy is not a surface-level artifact but emerges from a structural override of learned knowledge in deeper layers.

.
├── experiments/
│   ├── behavioral_analysis/       # Section: "User Opinion Induces Sycophancy"
│   │   ├── run_syco.py           # Main behavioral experiments      
│   │   └── run_syco.slurm
│   ├── mechanistic_analysis/     # Section: "Mechanistic Analysis"
│   │   ├── run_syco_logit_cot.py # Logit-lens & activation patching
│   │   └── run_syco_logit_cot.slurm
│   └── data_generation/          # Prefix & prompt generation
│       ├── generate_prefixes.py  
│       ├── full_question_builder.py 
│       └── apply_prefixes.py    
├── utils/
│   ├── DataFrameAligner.py       
│   ├── SycophancyAnalysis.py    
│   └── EarlyDecodingAnalysis.py 
├── lib/                           # Dataset library
│   ├── plain/                  # Plain baseline questions
│   ├── opinion_only/             # Opinion-only questions
│   │   ├── prefix/              # With prefix
│   │   └── suffix/              # With suffix
│   └── pov/                     # Perspective (1st/3rd person)
│       ├── prefix/
│       │   ├── first_pov/       # First-person prompts
│       │   └── third_pov/       # Third-person prompts
│       └── suffix/
│           ├── first_pov/
│           └── three_pov/
├── csv/                          
├── requirements.txt
├── README.md
└── DATA_STRUCTURE.md              # Detailed data organization guide

Note: For a detailed explanation of the lib/ data structure and file naming conventions, see DATA_STRUCTURE.md

# Clone the repository
git clone https://github.com/kaustpradalab/LLM-sycophancy.git
cd LLM-sycophancy

# Install dependencies
pip install -r requirements.txt

The experiments are organized by paper sections. Each section has standalone scripts that can be run independently.

Objective: Measure sycophancy rates across different prompt conditions (Plain, Opinion-only, Opinion + Expertise)

cd experiments/behavioral_analysis

# Plain questions (baseline)
python run_syco.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --question_type plain \
  --input_filename ../../lib/plain/mmlu_plain.pkl

# Opinion-only questions
python run_syco.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --question_type opinion_only \
  --input_filename ../../lib/opinion_only/prefix/mmlu_opinion_only.pkl

# Opinion + Expertise (First-person, Advanced)
python run_syco.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --question_type prefix_and_opinion \
  --prefix_type academic \
  --academic_level advanced \
  --prefix_subtype original \
  --input_filename ../../lib/pov/prefix/first_pov/mmlu_academic_opinion_advanced.pkl

Key Parameters:

• --question_type: plain, opinion_only, or prefix_and_opinion
• --academic_level: beginner, intermediate, or advanced
• --prefix_subtype: original (first-person) or third_pov (third-person)

# Submit job to SLURM cluster
sbatch run_syco.slurm "meta-llama/Llama-3.1-8B-Instruct"

# Monitor job
squeue -u $USER

Paper Findings (Figure 2, Figure 3):

• Opinion-only prompts induce ~63.7% average sycophancy rate
• Expertise level (Beginner/Intermediate/Advanced) has <4.4% impact

Objective: Understand how and when sycophancy emerges through layer-wise analysis

cd experiments/mechanistic_analysis

# Run with logit-lens analysis (all layers)
python run_syco_logit_cot.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --question_type opinion_only \
  --inference_mode logit_only \
  --inference_layer all \
  --input_filename ../../lib/opinion_only/prefix/mmlu_opinion_only.pkl

# Run with specific layers (e.g., odd layers)
python run_syco_logit_cot.py \
  --model_name Qwen/Qwen2.5-7B-Instruct \
  --question_type plain \
  --inference_layer odd \
  --input_filename ../../lib/plain/mmlu_plain.pkl

Key Parameters:

• --inference_mode: logit_only (just logits) or logit_and_cot (+ chain-of-thought)
• --inference_layer: all, odd, even, or last

Paper Findings (Figure 4, Figure 5, Figure 6):

• Decision score shift occurs at layers 16-19 (Llama 8B)
• KL divergence peaks at layer 23 (representational shift)
• Activation patching at critical layers reduces sycophancy by 36%

# Layer-wise decision tracking
python run_early_decoding.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --input_file ../../lib/plain/mmlu_plain.pkl

Objective: Compare first-person ("I believe") vs. third-person ("They believe") prompts

cd experiments/behavioral_analysis

# First-person (1st POV)
python run_syco.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --question_type prefix_and_opinion \
  --prefix_type academic \
  --academic_level advanced \
  --prefix_subtype original \
  --input_filename ../../lib/pov/prefix/first_pov/mmlu_academic_opinion_advanced.pkl

# Third-person (3rd POV)
python run_syco.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --question_type prefix_and_opinion \
  --prefix_type academic \
  --academic_level advanced \
  --prefix_subtype third_pov \
  --input_filename ../../lib/pov/prefix/third_pov/mmlu_academic_opinion_advanced.pkl

Paper Findings (Figure 8, Figure 9, Figure 10):

• First-person prompts induce 13.6% higher sycophancy on average
• KL divergence shows first-person creates stronger representational shifts
• Cosine similarity reveals orthogonal encoding of 1st vs. 3rd person perspectives

We tested 7 model families across similar parameter sizes:

If you use this code or build upon our work, please cite:

@inproceedings{wang2026sycophancy,
  title={When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models},
  author={Wang, Keyu and Li, Jin and Yang, Shu and Zhang, Zhuoran and Wang, Di},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

Sycophantic behavior in LLMs is primarily triggered by the presence of a user opinion, regardless of the user's claimed expertise or authority.

Sycophancy emerges in two stages: (1) late-layer output preference shift compared to Plain, then (2) deep representational divergence, confirming opinion framing overrides learned knowledge both behaviorally and internally.

Expertise-level framing fails to influence behavior because models do not encode it internally: opinion prompts form distinct clusters while level prompts overlap, indicating expertise cues are ignored representationally.

Grammatical person is a key driver of sycophancy in LLMs. Changing prompts from first- to third-person framing substantially reduces sycophantic behavior, with this effect encoded deep within the model's representations.

This work was supported by KAUST funding BAS/1/1689-01-01, KAUST Center of Excellence for Generative AI (award 5940), and a gift from Google.

Special thanks to the AAAI 2026 reviewers for their valuable feedback.