p-storm/STL-MCQA-results

STL Prompting for Zero-Shot MCQA

Results of zero-shot Multiple-Choice Question Answering (MCQA) experiments for the paper:

Dang, Q. P., Tran-Truong, P. T., Vu, D. L., Nguyen, L. S. T., Vo, Q. T. N., & Quan, T. (2026). Enhancing large language model performance for automatic zero-shot multiple-choice question answering via single-token logit prompting. Computers and Education: Artificial Intelligence. DOI: 10.1016/j.caeai.2026.100578

Source code: github.com/LeoTTTPhat/STL-Prompting-For-ZeroShot-MCQA

This dataset contains the raw per-question predictions, logits, probabilities, and resource-usage measurements produced by evaluating three prompting strategies (STL, CoT, LTL/baseline) across several English and Vietnamese MCQA benchmarks and open-weight 7B LLMs.

---

What is inside

Strategy	Description	Output format
STL (Single-Token Logit)	Each option is scored independently with a yes/no prompt; the yes logits are normalized with softmax. Mitigates the Multiple-Choice Symbol Binding (MCSB) problem.	yes_no (CSV)
CoT (Chain-of-Thought)	Free-form reasoning followed by an answer letter.	A_B_C_D (CSV) + reasoning_full (JSONL)
LTL (baseline / Logit baseline)	Standard single-prompt logit extraction over options A–D.	A_B_C_D (CSV)

Benchmarks

Benchmark	Language	Models covered
ARC-Challenge	English	Mistral-7B-Instruct-v0.1, Llama-2-7b-chat-hf
ARC-C_RAG	English (RAG-augmented)	Mistral-7B-Instruct-v0.1
ViCS-MCQ	Vietnamese	Qwen2.5-7B-Instruct, Vistral-7B-Chat
VMLU	Vietnamese	Qwen2.5-7B-Instruct
VSMRC	Vietnamese	Qwen2.5-7B-Instruct
ViCS-MCQ + VSMRC (combined)	Vietnamese	Qwen2.5-7B-Instruct, Vistral-7B-Chat

Models

• Mistral-7B-Instruct-v0.1
• Llama-2-7b-chat-hf (daryl149/llama-2-7b-chat-hf)
• Qwen2.5-7B-Instruct (Qwen/Qwen2.5-7B-Instruct)
• Vistral-7B-Chat (VietAI/Vistral-7B-Chat)

---

File structure

.
├── CoT_<BENCHMARK>_<MODEL>_A_B_C_D.csv              # CoT predictions (English / per-benchmark)
├── CoT_<BENCHMARK>_<MODEL>_A_B_C_D_reasoning_full.jsonl  # CoT free-form reasoning text
├── LTL_<BENCHMARK>_<MODEL>_A_B_C_D.csv              # Baseline logit predictions
├── STL_<BENCHMARK>_<MODEL>_yes_no.csv               # STL yes/no logit predictions
├── Summary_<NAME>_LTL_STL_CoT.csv                   # Aggregate accuracy / latency per method
└── ViCS+VSMRC/                                      # Combined Vietnamese benchmark
    ├── CoT_<MODEL>_A_B_C_D.csv
    ├── CoT_<MODEL>_A_B_C_D_reasoning_full.jsonl
    ├── LTL_<MODEL>_A_B_C_D.csv
    ├── STL_<MODEL>_yes_no.csv
    └── Summary_<NAME>_LTL_STL_CoT.csv

Naming convention: <METHOD>_<BENCHMARK>_<MODEL>_<FORMAT>.csv

• METHOD ∈ {CoT, LTL, STL}
• BENCHMARK ∈ {ARC-Challenge, ARC-C_RAG, ViCS-MCQ, VMLU, VSMRC} (the combined set in ViCS+VSMRC/ omits the benchmark token)
• MODEL ∈ {Mistral-7B-Instruct-v0.1, llama-2-7b-chat-hf, Qwen2.5-7B-Instruct, Vistral-7B-Chat}
• FORMAT ∈ {A_B_C_D, yes_no}

---

Column schemas

CoT — A_B_C_D (top-level benchmarks)

Column	Meaning
method	cot
question_id	Index of the question in the benchmark
question	Question text (with options)
truth_label	Ground-truth answer letter
prediction	Predicted answer letter (parsed)
input_tokens	Prompt tokens consumed
output_tokens	Generated tokens
latency_seconds	Wall-clock inference time
target_tokens	Target token mode (A_B_C_D)
reasoning	Model reasoning text
logit_A … logit_D	Raw logits for each option letter
prob_A … prob_D	Normalized probability for each option letter

CoT — A_B_C_D (inside ViCS+VSMRC/ — extended)

Same as above plus the following leading/extra columns: parse_method (how the answer letter was extracted), parse_failed (bool), and a trailing dataset_name (either ViCS-MCQ or VSMRC).

LTL (baseline) — A_B_C_D

Column	Meaning
method	baseline
question_id, question, truth_label, prediction	As above
input_tokens, output_tokens, target_tokens	Token counts / mode
logit_A … logit_D	Raw logits per option
prob_A … prob_D	Normalized probabilities per option
latency_seconds	Inference time
memory_usage_mb, peak_memory_mb, gpu_memory_usage_mb	RAM / VRAM usage
avg_cpu_usage_percent, max_cpu_usage_percent	CPU utilization
avg_gpu_usage_percent, max_gpu_usage_percent	GPU utilization
dataset_name	(only in ViCS+VSMRC/ files)

STL — yes_no

Column	Meaning
method	stl
question_id, question, truth_label, prediction	As above
input_tokens, output_tokens, target_tokens	Token counts (yes_no)
raw_yes_logit_prompt_A … raw_no_logit_prompt_D	Raw yes/no logits per option-specific prompt
normalized_yes_logit_prompt_A … normalized_no_logit_prompt_D	Normalized logits per option prompt
prob_A … prob_D	Final softmax probability per option
latency_seconds	Inference time (k prompts per question)
memory_usage_mb, peak_memory_mb, gpu_memory_usage_mb	RAM / VRAM usage
avg_cpu_usage_percent, max_cpu_usage_percent	CPU utilization
avg_gpu_usage_percent, max_gpu_usage_percent	GPU utilization
dataset_name	(only in ViCS+VSMRC/ files)

CoT — reasoning_full.jsonl

One JSON object per question:

{
  "question_id": 0,
  "truth_label": "D",
  "prediction": "D",
  "parse_method": "pattern",
  "reasoning_full": "Reasoning:\n...Answer: D"
}

Summary — Summary_<NAME>_LTL_STL_CoT.csv

Column	Meaning
method	LTL / STL / CoT
dataset	Benchmark name
model	Hugging Face model id
accuracy	Accuracy (0–1)
accuracy_percent	Accuracy (%)
average_latency_seconds	Mean latency per question
num_samples	Number of questions evaluated

---

Loading the data

import pandas as pd

# CoT predictions (single benchmark)
df = pd.read_csv("CoT_ViCS-MCQ_Qwen2.5-7B-Instruct_A_B_C_D.csv")

# STL yes/no predictions
df_stl = pd.read_csv("STL_VSMRC_Qwen2.5-7B-Instruct_yes_no.csv")

# CoT reasoning
import json
with open("CoT_ViCS-MCQ_Qwen2.5-7B-Instruct_A_B_C_D_reasoning_full.jsonl") as f:
    reasoning = [json.loads(line) for line in f]

# Summary
summary = pd.read_csv("Summary_ViCS-MCQ_Qwen2.5-7B-Instruct_LTL_STL_CoT.csv")

---

Citation

@article{DangTranTruong2026enhancing,
  title   = {Enhancing large language model performance for automatic zero-shot multiple-choice question answering via single-token logit prompting},
  author  = {Dang, Q. P. and Tran-Truong, P. T. and Vu, D. L. and Nguyen, L. S. T. and Vo, Q. T. N. and Quan, T.},
  journal = {Computers and Education: Artificial Intelligence},
  year    = {2026},
  doi     = {10.1016/j.caeai.2026.100578}
}

License

The result files are released under the MIT license (same as the source repository). Please cite the paper above when using this dataset.