Arena-style rankings of language models are a widely used evaluation framework in online leaderboards, research papers, and model evaluation in production. These rankings aim to realistically assess model performance using pairwise comparisons of anonymous models on user prompts, aggregating votes via the Bradley-Terry model to produce the final ranking. However, we find that these leaderboards are highly vulnerable to several manipulation strategies that unscrupulous providers could use to boost a model's rank: strategic voting (individual votes that boost a model's rank, including votes on irrelevant models), strategic prompting (choosing prompts that favor a model), and strategic nomination (boosting the rank of a model by inserting irrelevant models). We demonstrate these effects on public leaderboard data, analyze the circumstances that cause these vulnerabilities, and describe approaches to safeguard against each attack. Notably, these issues hold for any preference-based evaluation using the Bradley-Terry model that allows open prompting, voting, or model submission. These manipulations boost the rank of nearly every tested model; each attack can boost a model by up to 15-20 places in a ranking and, combined, they can boost a model by up to 45 places.
Figures & tables
Figure 1: Top: Highest model rank (dark bars) and Elo (light bars) achieved under greedy strategic voting, capped at 10K votes ( ∼10% of total votes). Numbers above bars indicate the best absolute rank that a model achieves. Models are boosted up to 20 places in the rankings; 23 models (42% of the dataset) move over 10 places up, and eight models move into the top 5. Bottom: Highest model rank & Elo achieved under heuristic strategic voting, under same conditions. Models are boosted up to 15 places in the rankings; 21 models (38% of the dataset) move over 10 places up, and 15 models move into the top 5.
Figure 2: Change in model rank and Elo achieved when adding strategic nominations (10 copies of the most favorable model). Change in rank is the left-hand bar of each double-bar; change in Elo is the right-hand (lighter) bar. Numbers above bars indicate the best absolute rank achieved by the model (1=best). For example, the model originally in 19th place (light blue) jumps 15 places to 4th place.
Figure 3: Change in model rank and Elo achieved by adding strategic prompts, capped at 10K prompts ( ∼10% of total battles). Models are boosted up to 15 places in the ranking.
Figure 4: Change in rank and Elo when performing strategic nominations followed by strategic prompting. Models are boosted up to 45 places in the rankings, with 9 models moving over 10 places upwards.
Voting (greedy) also see [ 8 ]
Voting (heuristic)
Prompting
Nomination
Prompting + nomination
Max rank boost
20 places ↑
15 places ↑
15 places ↑
15 places ↑
45 places ↑
Must identify models?
All models
Only target model A
–
–
–
Cost of Elo calculations ( m models, n votes; n≫m )
O(n)
0
≤O(m) ; depends on desired # of prompt subsets
O(m)
O(m)
Possible defenses
Block bots & prompting w/o voting
Block bots & prompting w/o voting
Stratify/cluster prompts
Harmonic Borda count
Harmonic Borda count
Stratify/cluster prompts
Table 1: Comparison of attack strategies and their properties.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Boost in rank achievable under strategic nomination when using the standard Borda count (blue) vs. the harmonically weighted Borda count (orange). The harmonic version reduces the maximum boost from 15 places to 8 places.
β
Mean best boost
Max boost
Kendall τ with regular BT
0
2.74
15
1
0.25
2.74
14
0.993
0.50
2.63
10
0.974
0.75
2.81
8
0.946
1.00
4.39
12
0.852
Appendix
Table 2: Effect of harmonic Borda count on the highest boosts that strategic nomination can achieve (mean highest boost per model, and highest boost across models), as well as Kendall τ distance from original ranking, at different β weightings.
Figure 6: Number of votes under the greedy strategy (top) and heuristic strategy (bottom) needed to achieve greater improvements in leaderboard rank. Point labels indicate the number of models that can achieve that rank improvement with under 10K strategic votes. Small improvements to rank (2-3 places) are relatively easy to achieve, requiring under 2% of total votes (2K of 100K total) to be manipulated and achievable by a relatively large subset of models. Larger improvements (15+ places) are achievable by a smaller subset of models with more manipulated votes. While the maximum rank boost achieved is higher for the greedy strategy, more models are able to achieve smaller boosts under the heuristic strategy (and typically with fewer votes).
Figure 7: Counterfactual of model ranks with and without the family B of models that moves the target model furthest up when included. We plot each model’s rank when family B is excluded from the Elo calculation, versus its rank (among models not in B ) when family B is included in the Elo dataset, i.e., the original dataset.
Figure 8: Counterfactual of model ranks with and without the family B of models that moves the target model furthest down when included. We plot each model’s rank when family B is excluded from the Elo calculation, versus its rank (among models not in B ) when family B is included in the Elo dataset, i.e., the original dataset.
Figure 9: Average rank change achieved as more strategic prompts are added.
Figure 10: Model detection precision at different fractions of total outputs guessed for open-weight models (square markers, dashed lines) and proprietary GPT models (circle markers, solid lines). Random baseline shown at 0.25 (base rate).
Figure 11: Precision@30 (15% of outputs) as a function of tokens scored for open-weight models. Many models reach near-optimal precision with as few as 30 tokens, suggesting substantial efficiency gains are possible by truncating log-probability calculations early.
Feature
Coefficient
elo density vs above
0.354919
winrate variance
0.079747
# total battles
-0.064431
elo gap vs model above
0.028544
winrate vs model above
-0.027852
model Elo
0.026598
Appendix
Table 3: Regression coefficients for the best Δ rank achieved from strategic voting (RMSE=2.75).
Feature
Coefficient
# battles in best prompt lang
-0.337490
elo gap vs model above
-0.126945
elo density vs above
0.111206
winrate variance
0.069611
family size
-0.066190
winrate vs model above
0.040990
Appendix
Table 4: Regression coefficients for the best Δ rank achieved from strategic prompting (RMSE=2.67).
Feature
Coefficient
elo density vs above
0.231206
elo gap vs model above
-0.200973
winrate variance
0.179882
model Elo
0.157270
# battles in best prompt lang
0.140642
tie rate
0.089433
Appendix
Table 5: Regression coefficients for the best Δ rank achieved from strategic nomination (RMSE=1.85).
Figure 12: Best delta rank achieved when manipulating a constant 10% of battles on datasets of different sizes, on the heuristic voting strategy. When plotting dataset size (x) against best rank achieved with ≤ 10% battles modified (y), the slope is near-constant with p >> 0.05 (not significantly different from a line with slope 0). This indicates that the number of manipulations needed to achieve similar rank boosts scales linearly with dataset size.
Figure 13: Best delta rank achieved when manipulating a constant 10% of battles on datasets of different sizes, on the greedy voting strategy. When plotting dataset size (x) against best rank achieved with ≤ 10% battles modified (y), the slope is near-constant with p >> 0.05 (not significantly different from a line with slope 0). This indicates that the number of manipulations needed to achieve similar rank boosts scales linearly with dataset size.
Figure 14: Best delta rank achieved when manipulating a constant 10% of battles on datasets of different sizes, with the strategic prompting strategy. When plotting dataset size (x) against best rank achieved with ≤ 10% battles modified (y), the slope is very slightly positive (p < 0.05), meaning that the manipulation becomes slightly easier as dataset size increases.
Figure 15: Best delta rank achieved when manipulating a constant fraction of models on datasets of different sizes, using the strategic nomination strategy. When plotting dataset size (x) against best rank achieved (y) with ≤ 10 model copies inserted (number of battles per model copy thus scales with dataset size), the slope is slightly positive (p < 0.05), meaning the attack is somewhat easier as dataset size increases with the same proportion of models changed (though note that the proportion of models to total votes may vary by leaderboard).
Figure 16: Best delta rank achieved when manipulating a constant 10% of battles on datasets of different sizes, when combining the strategic nomination and prompting strategies in succession. When plotting dataset size (x) against best rank achieved with ≤ 10% battles modified (y), the slope is near-constant with p >> 0.05 (not significantly different from a line with slope 0). This indicates that the number of manipulations needed to achieve similar rank boosts scales linearly with dataset size.
Figure 17: Proof-of-concept that strategic prompting can use any definable, consistent subcategories: best delta rank and Elo achieved under strategic prompting, where prompt categories are based on whether the prompt is labeled as involving complexity, creativity, domain knowledge, and/or math. Rank boosts (up to 14 places) are comparable to those from strategic prompting using language for prompt subcategories.
Evaluation leaderboards such as LMArena play a central role in benchmarking large language models by aggregating pairwise human preferences into model rankings, yet the robustness of these rankings remains poorly understood. We present a unified perturbation framework for analyzing Bradley-Terry leaderboards under structured data modifications using influence-based approximations. Our framework studies three match-level perturbations -- Drop, Add, and Flip -- together with player removal, and evaluates their effects on top-k membership, global ranking consistency via Kendall's tau, and confidence-interval-based uncertainty. Across Chatbot Arena and six additional pairwise-comparison datasets, we show that modern leaderboards are non-robust across all three objectives: sub-1% targeted perturbations can change the top-ranked model, degrade Kendall's tau, and alter confidence intervals. Beyond robustness auditing, we show that the same influence scores enable efficient targeted perturbations, promoting or demoting specific models and reducing target-model uncertainty with fewer actions than previous manipulation and active-sampling baselines. By summarizing these effects with normalized dataset-level robustness scores, our framework provides a practical and helpful tool for auditing leaderboard stability and motivating more robust evaluation protocols.
Hosna Oyarhoseini, Jimmy Lin, Amir-Hossein Karimi
University of Waterloo, Waterloo, Ontario, Canada.
We propose a method for evaluating the robustness of widely used LLM ranking systems -- variants of a Bradley--Terry model -- to dropping a worst-case very small fraction of preference data. Our approach is computationally fast and easy to adopt. When we apply our method to matchups from popular LLM ranking platforms, including Chatbot Arena and derivatives, we find that the rankings of top-performing models can be remarkably sensitive to the removal of a small fraction of preferences; for instance, dropping just 0.003% of human preferences can change the top-ranked model on Chatbot Arena. Our robustness check identifies the specific preferences most responsible for such ranking flips, allowing for inspection of these influential preferences. We observe that the rankings derived from MT-bench preferences are notably more robust than those from Chatbot Arena, likely due to MT-bench's use of expert annotators and carefully constructed prompts. Finally, we find that neither rankings based on crowdsourced human evaluations nor those based on LLM-as-a-judge preferences are systematically more sensitive than the other.
Jenny Y. Huang, Yunyi Shen, Dennis Wei +1
Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology · MIT-IBM Watson AI Lab · IBM Research
Public LLM leaderboards guide model evaluation and adoption, yet the reported performance of top-ranked models can be systematically overstated. This leaderboard winner's curse arises because ranking noisy benchmark scores rewards sampling luck alongside genuine ability. The same selection effect can distort scores throughout the leaderboard, a phenomenon we call the rank-induced bias. We develop an m-out-of-N bootstrap to correct for this bias. It assumes a fixed model pool with independent, identically distributed bounded tasks while allowing arbitrary within-task correlation among models. Under exact population tiers separated by positive gaps, we prove that the bootstrap reduces Θ(N−1/2) ranking bias to expected residual bias o(N−1/2). Across 14 benchmarks from 5 benchmark families, the rank-1 adjustments range from approximately 0.09 to 3.86 percentage points. In our SWE-bench Verified snapshot, the tied leaders receive an adjustment of 0.90 percentage points. Such overstatement is large enough to make the apparent winner statistically fragile. Thus, we use the same bootstrap to further quantify the intervals of plausible rank positions for each model and group models into tiers. We establish sufficient conditions for consistent recovery of population tiers. On SWE-bench Verified, this procedure places 13 of 173 models in the leading tier. We combine the score adjustment and rank intervals in the Rank Uncertainty and Leaderboard Evaluation Report (RULER). We validate its finite-sample performance using sample splitting and calibrated simulations, and advocate publishing RULER alongside raw rankings so readers can assess both reported performance and the precision of the ordering. RULER also provides guidance for benchmark design that accounts for selection effects and ranking uncertainty.
Sikun Xu, Raphael Thomadsen, Dennis J. Zhang +1
Olin Business School, Washington University in St. Louis St. Louis, MO 63130 · W. P. Carey School of Business, Arizona State University Tempe, AZ 85287