Research map / Large language models / Retrieval and knowledge
Small Models Exhibit Limited Answer Consistency in Repetition Trials of the Multiple-Choice MMLU-Redux and MedQA Benchmarks
Claudio Pinhanez, Paulo Cavalin, Cassia Sanctos, Marcelo Carpinette Grave
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a medium model (llama-3.3-70b) instead of a small one (Llama-3-8B) | Llama family · MMLU-Redux (10 repetitions) | S/T (% SURE questions): 53% -> 98% | Improved |
| raise inference temperature from 0.3 to 1.0 (ablation) | Llama-3-8B · MMLU-Redux (10 repetitions) | S/T (% SURE questions): 53% -> 11% | Hurt |
| raise inference temperature from 0.3 to 1.0 (ablation) | Llama-3-8B · MMLU-Redux (10 repetitions) | RWS (right when SURE): 0.81 -> 0.94 | Improved |
| raise inference temperature from 0.3 to 1.0 (ablation) | Llama-3-8B · MMLU-Redux (10 repetitions) | accuracy average: 0.590 -> 0.428 | Hurt |
| raise inference temperature from 0.3 to 1.0 (ablation) | medllama3-v20 · MedQA (10 repetitions) | S/T (% SURE questions): 96% -> 88% | Hurt |
| raise inference temperature from 0.3 to 1.0 (ablation) | medllama3-v20 · MedQA (10 repetitions) | RWS (right when SURE): 0.75 -> 0.76 | No clear effect |
| use instruct instead of base Llama-3-8B | Llama-3-8B · MMLU-Redux (10 repetitions) | S/T (% SURE questions): 53% -> 79% | Improved |
| use instruct instead of base Llama-3-8B | Llama-3-8B · MedQA (10 repetitions) | accuracy average: 0.512 -> 0.405 | Hurt |
| medical finetuning of Mistral-7B-Instruct into BioMistral-7B | Mistral-7B-Instruct · MedQA (10 repetitions) | accuracy average: 0.321 -> 0.371 | Improved |
| medical finetuning of llama1-7b into medalpaca-7b | llama1-7b · MedQA (10 repetitions) | accuracy average: 0.201 -> 0.342 | Improved |
| keep only SURE answers (filter out UNSURE) for Bio-Medical-Llama-3-8B | Bio-Medical-Llama-3-8B · MedQA (10 repetitions) | accuracy: 0.736 -> 0.79 | Improved |
| use granite 3.1-8b-instruct instead of 3.1-8b-base | granite 8B · MMLU-Redux (10 repetitions) | accuracy average: 0.575 -> 0.640 | Improved |
Closest papers by meaning
- Representation Consistency for Accurate and Coherent LLM Answer Aggregation (NeurIPS 2025 · similarity 0.91)
- Latent Self-Consistency for Reliable Majority-Set Selection in Short- and Long-Answer Reasoning (AAAI 2026 · similarity 0.89)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.88)
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.88)
- Evaluating language models as risk scores (NeurIPS 2024 · similarity 0.87)
- Operads for compositional reasoning in LLMs (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
