How Prompt Style Affects AI Accuracy and Speed in Competitive Programming

Saad B. Alarifi; Mohammed A. Alam

cs.LGOct 21, 2026

How Prompt Style Affects AI Accuracy and Speed in Competitive Programming

Authors: Saad B. Alarifi, Mohammed A. Alam

Organizations: Riyadh Schools Riyadh, Saudi Arabia · Department of Computer Science Northwestern University Evanston, IL, USA

Abstract

Large language models are increasingly used in competitive programming, and their outputs are sensitive to how a prompt is phrased. Prior work has compared prompt strategies for accuracy on competitive programming, but not for runtime speed, and has not situated the results against human performance. In an exploratory study of one prompt in each of nine styles on GPT-5.5 across three accuracy tasks from the International Olympiad in Informatics (IOI) and two NP-hard speed tasks, we found that no single prompt was best: the strongest prompt depended on the task type and on whether results were summarized by the mean or the median, suggesting that, at least for these tasks, prompt style should be matched to the objective rather than treated as universal. Many differences in mean scores came from a few outlier runs, showing that evaluating prompting by a single metric can obscure how much its effectiveness varies. Compared descriptively with human IOI contestants, who worked under different conditions, the AI was neither strictly better nor strictly worse, exceeding the average contestant on one task but falling short on the other two. By extending prompt comparison from accuracy to runtime speed and grounding it against human performance, this study gives a broader picture of how prompt style affects competitive programming performance.

Figures & tables

Setting	Maximum Clique Size	Number Partitioning
Test input	Random graph on $N$ vertices; each edge included independently with probability $N^{-1/N}$	$N$ random integers between 1 and $2^{48}$
Test generation	Deterministic from a fixed seed	Deterministic from a fixed seed
Baseline solution	MaxCliqueDyn exact algorithm [ 18 ]	Our own exact meet-in-the-middle solution
Baseline time limit	10 s per test case	10 s per test case
Ceiling (largest $N$ solved by baseline)	220	49
AI solution time limit	1 s per test case	1 s per test case

TABLE I: Speed evaluation settings.

	IOI contestants	AI solver (this study)
Source of scores	Official IOI results [ 23 , 24 , 25 ]	oj.uz, official IOI scoring [ 17 ]
Scores averaged per task	Prisoner Challenge (2022): 353 contestants, 30 gold medalists Message (2024): 366 contestants, 34 gold medalists Migrations (2025): 334 contestants, 28 gold medalists	81 runs (9 prompt styles $\times$ 9 runs)
Time	5 hours per contest day, shared across three tasks	One response per run
Submissions per task	Up to 50	1 per run
Feedback before resubmitting	Subtask results after each submission	None
Task score	Sum over subtasks of the best result across all submissions	Score of the single submission

TABLE II: Scores and conditions for human IOI contestants and the AI solver on the accuracy tasks.

Prompt	Prisoner Challenge	Message	Migrations	Average
Prompt 1	62.78	30.63	23.68	39.03
Prompt 2	68.89	29.61	22.70	40.40
Prompt 3	63.17	29.94	18.68	37.26
Prompt 4	75.00	30.45	34.16	46.54
Prompt 5	67.44	28.96	26.55	40.98
Prompt 6	79.44	36.26	26.88	47.53

TABLE III: AI accuracy scores by prompt style, scores are out of 100. Mean over 9 iterations per prompt–task pair. Shading compares prompts within each column (green = best, red = worst); bold = column best.

Prompt	Prisoner Challenge	Message	Migrations	Average
Prompt 1	65.00	31.40	20.00	38.80
Prompt 2	65.00	30.33	20.00	38.44
Prompt 3	65.00	29.32	20.00	38.11
Prompt 4	65.00	31.40	28.10	41.50
Prompt 5	65.00	29.32	20.01	38.11
Prompt 6	80.00	31.40	20.00	43.80

TABLE IV: AI accuracy scores by prompt style, scores are out of 100. Median over 9 iterations per prompt–task pair. Shading compares prompts within each column (green = best, red = worst); bold = column best.

Fig. 1: Mean accuracy by prompt style, scores are out of 100 and averaged across the three IOI tasks (Prisoner Challenge, Message, Migrations).

† Confounded with grading order; see Section V-C .
Task	Type	H	$p$	Holm $p$	$\varepsilon^{2}$
Prisoner Challenge	Accuracy	19.58	0.012	0.048	0.24
Message	Accuracy	17.16	0.029	0.086	0.21
Migrations	Accuracy	7.07	0.529	1.000	0.09
Maximum Clique Size	Speed	2.35	0.968	1.000	0.03
Number Partitioning †	Speed	20.77	0.008	0.039	0.26

TABLE V: Kruskal-Wallis H test of whether prompt style affected scores on each task (9 prompt styles

\times

9 runs, 8 degrees of freedom), with Holm-adjusted

p

-values and effect size

\varepsilon^{2}

. Shading marks significance at the 0.05 level after Holm adjustment (green = significant, red = not significant); bold = significant before adjustment.

Prompt	Maximum Clique Size	Number Partitioning	Average
Prompt 1	65.35	80.27	72.81
Prompt 2	72.32	79.59	75.96
Prompt 3	69.65	87.30	78.47
Prompt 4	58.74	78.68	68.71
Prompt 5	61.16	78.91	70.04
Prompt 6	62.02	84.58	73.30

TABLE VI: AI speed scores by prompt style, normalized to 100. Maximum completed

N

within the 1 s time limit, normalized per task; mean over 9 iterations. Shading compares prompts within each column; bold = column best.

Prompt	Maximum Clique Size	Number Partitioning	Average
Prompt 1	67.27	91.84	79.56
Prompt 2	80.00	87.76	83.88
Prompt 3	67.27	87.76	77.52
Prompt 4	58.18	87.76	72.97
Prompt 5	67.27	87.76	77.52
Prompt 6	58.18	87.76	72.97

TABLE VII: AI speed scores by prompt style, normalized to 100. Maximum completed

N

within the 1 s time limit, normalized per task; median over 9 iterations. Shading compares prompts within each column; bold = column best.

Fig. 2: Mean speed performance by prompt style, normalized to 100 per task and averaged across the two NP-hard optimization tasks (Maximum Clique Size, Number Partitioning).

Fig. 3: AI performance relative to human contestants on the IOI accuracy tasks, scores are out of 100. Bars show the mean score for all contestants, the mean score of gold medalists, and the AI solver (averaged across all nine prompt styles). Contestant and gold-medalist scores were obtained from the official IOI statistics [ 23 , 24 , 25 ] .

LLM Prompting

Large language model prompting studies how instructions, examples, context, and formats steer model behavior at inference time.

See all 429 papers

Mar 29, 2026

Model Capability Dominates: Inference-Time Optimization Lessons from AIMO 3

Majority voting over multiple LLM attempts improves mathematical reasoning, but correlated errors limit the effective sample size. A natural fix is to assign different reasoning strategies to different voters. The approach, Diverse Prompt Mixer, is tested on the AIMO 3 competition: 3 models, 23+ experiments, 50 IMO-level problems, one H100 80 GB, 5-hour limit. Every prompt-level intervention fails. High-temperature sampling already decorrelates errors; weaker strategies reduce accuracy more than they reduce correlation. Across an 8-point capability gap at equal N=8 and every optimization tested, model capability dominates. The gap between the best majority-vote score (42/50) and pass@20 (~45.5) is selection loss, not prompt loss. A verifier-based selector could close it. Prompt engineering cannot.

Natapong Nitarach

LLM Reasoning LLM Prompting Language Model Ensembles Test-Time Scaling Self-Consistency Decoding

Ask PDF

Oct 10, 2025

LiveOIBench: Can Large Language Models Outperform Human Contestants in Informatics Olympiads?

Competitive programming problems are increasingly used to evaluate the coding capabilities of large language models (LLMs) due to their complexity and ease of verification. Yet, current coding benchmarks face limitations such as a lack of exceptionally challenging problems, insufficient test case coverage, and reliance on online platform APIs that limit accessibility. To address these issues, we introduce LiveOIBench, a large-scale competitive programming benchmark featuring 403 expert-curated problems, averaging 60 official test cases each, drawn from 72 contests across 14 Informatics Olympiads held between 2023 and 2025. LiveOIBench has four key features: (1) expert-designed tasks with detailed subtask rubrics and extensive test cases; (2) direct comparison to elite human contestants; (3) continuous updates to reduce contamination risk; and (4) a fully offline, reproducible evaluation system. Benchmarking 34 popular general-purpose and reasoning LLMs, we find that GPT-5 achieves an 81.76th percentile, still falling short of top human contestants, while among the open-weight models, GPT-OSS-120B reaches only the 60th percentile. Reasoning-trace analyses indicate that robust reasoning models prioritize precise problem analysis over excessive exploration. Finally, analyses across release dates, task familiarity, and code similarity find minimal evidence of data contamination in our benchmark. Our leaderboard, code, and data are available at: https://liveoibench.github.io/.

Kaijian Zou, Aaron Xiong, Yunxiang Zhang +6

University of Michigan, Ann Arbor.

LLM Evaluation Competitive Programming Code Generation Evaluation Code Generation Benchmarks

Ask PDF

Apr 3, 2026

GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning

Competitive programming remains one of the last few human strongholds in coding against AI. The best AI system to date still underperforms the best humans competitive programming: the most recent best result, Google's Gemini3 Deep Think, attained 8th place even not being evaluated under live competition conditions. In this work, we introduce GrandCode, a multi-agent RL system designed for competitive programming. The capability of GrandCode is attributed to two key factors: (1) It orchestrates a variety of agentic modules (hypothesis proposal, solver, test generator, summarization, etc) and jointly improves them through post-training and online test-time RL; (2) We introduce Agentic GRPO specifically designed for multi-stage agent rollouts with delayed rewards and the severe off-policy drift that is prevalent in agentic RL. GrandCode is the first AI system that consistently beats all human participants in live contests of competitive programming: in the most recent three Codeforces live competitions, i.e., Round1087 (Mar 21, 2026), Round~~1088 (Mar 28, 2026), and Round~~1089 (Mar 29, 2026), GrandCode placed first in all of them, beating all human participants, including legendary grandmasters. GrandCode shows that AI systems have reached a point where they surpass the strongest human programmers on the most competitive coding tasks.

Ornith Team, Xiaoya Li, Guoyin Wang +3

DeepReinforce Team

RL for Code Generation Agentic RL Group Relative Policy Optimization Competitive Programming Multi-Agent Reinforcement Learning

Ask PDF

Abstract

LLM Prompting

Explore similar work

Model Capability Dominates: Inference-Time Optimization Lessons from AIMO 3

LiveOIBench: Can Large Language Models Outperform Human Contestants in Informatics Olympiads?

GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning