Smaller local LLMs show a minor position bias toward option ordering and partially guessing-like error variation, but accuracy remains relatively stable across permutations and repeated sampling runs¶
Claim · All claims
Evidence · 1 study · causalr2·q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? Rotating MCQ option positions slightly affects Mistral-7B and DeepSeek-R1 scores, and repeated sampling runs show slight score variation, with accuracy and fully-correct answers remaining relatively stable. → Markos Stamatakis 2026
Evidence¶
Markos Stamatakis 2026¶
Markos Stamatakis, Omkar Gavali, Joshua Berger, Christian Wartena, Anett Hoppe, and Ralph Ewerth. (2026). Are LLM-based Chatbots Good Enough to Support Computer Science Students in Multiple-Choice Exercises? https://arxiv.org/abs/2606.15919 doi:10.48550/arxiv.2606.15919
q2 · i? · causal · r2
Stability evaluation (Table 3, Sec. 4): four cyclic option permutations per question and 10 repeated runs with default sampling for the two local models. The article reports a "minor position bias for both models" and error-similarity variation hinting at partial guessing; descriptive percentages only.
"varying the order of the options slightly affects the score, indicating a minor position bias for both models."
Discussion¶
Related Claims¶
- GPT-4o and GPT-5.2 achieve higher accuracy and more fully correct answers on university IVDA multiple-choice questions than the smaller Mistral-7B and DeepSeek-R1 models — related
- LLM difficulty predictions for fraction items are less stable across repeated runs than for the overall item set — related
- Some LLMs exhibit a selection bias against selecting option D on multiple-choice pedagogy questions — related
- The benchmark procedure is stable under stochastic decoding, with a maximum accuracy standard deviation of 0.43% over 20 repeated runs — related
- Small input perturbations such as minor spelling errors can reduce state-of-the-art dialogue and assessment model accuracy by over 30% — related
- Cumulative meta-analysis shows reported GenAI productivity effects remained stable over time (stabilizing at g = 0.331) despite rapid advancement of underlying foundation models. — related
- Model and prompt choice account for only a small share of misalignment error, which concentrates in transcript-conditioned higher-order interactions — related
- Model and student error patterns on IVDA MCQs show limited question-level consistency but comparable difficulty distributions for easiest and hardest questions — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
minor-position-bias-and-partial-guessing-in-small-llms |
title |
Smaller local LLMs show a minor position bias toward option ordering and partially guessing-like error variation, but accuracy remains relatively stable across permutations and repeated sampling runs |
description |
Smaller local LLMs show a minor position bias toward option ordering and partially guessing-like error variation, but accuracy remains relatively stable across permutations and repeated sampling runs |
status |
review |
evidence_strength |
weak |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — markos-stamatakis-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.