A team is choosing a model, and the tie-breaker is the Arena. Two candidates score within a few points on the static benchmarks the team already trusts, so they open the public human-preference leaderboard — the one where anonymous models answer the same prompt and a crowd votes which reply is better — and pick the one sitting a few rungs higher. It feels like the most honest signal available: no curated task set to overfit, no contractor writing the questions, just real people preferring real answers, aggregated into an Elo number. A market of human taste, and the market has spoken.
The Elo number is real, in the sense that the votes happened and the arithmetic is correct. What it is not is a measurement of model quality the way the team reads it. A static benchmark is a fixed instrument you can at least characterize — its size, its noise floor, its blind spots. A live preference leaderboard is something else: a mechanism, with rules, and rules have degrees of freedom the better-resourced player gets to exploit. Who gets to test privately and publish only their best run. Who gets sampled more and removed less. Whether a thousand purchased votes can move a rank. Whether a longer, better-formatted answer wins for being longer rather than better.
This post is about the Arena specifically, and the failure is not the one the other eval posts cover. It is not that the score is a point estimate that won’t survive production, nor that the delta sits inside a static set’s sampling noise. It is that the leaderboard is a contested mechanism, and an entity with more access and a strategy can move its rank without building a better model. The rank reflects the game as much as the player.
Selective disclosure turns a private test bench into a public best-of
Start with the cheapest exploit, because it requires no attack at all — only access and a willingness to use it.
The Leaderboard Illusion (arXiv 2504.20879) audits Chatbot Arena and finds, first, an undisclosed private-testing regime. A handful of providers can run multiple variants of a model before any public release, see each variant’s score, and publish only the best one — retracting the rest as if they never happened. The paper’s language is direct: “undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired,” and “the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results.” At the extreme they “identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release.”
Sit with the mechanism, because it is the same one that makes a clean contamination audit lie after RL post-training: a selection effect dressed as a measurement. Sample one model once and its Arena score is a noisy estimate of where it sits. Sample twenty-seven variants and publish the max, and the number is no longer an estimate of any model’s quality — it is the upper tail of a distribution of noisy draws, and the more draws you take the higher that tail reaches, whether or not any variant was actually better. The provider who tested one variant and the one who tested twenty-seven are not reporting the same kind of number, yet the leaderboard prints both in the same column with the same units. The rank gap between them can be manufactured entirely out of attempts.
This is not cheating in the sense of breaking a rule. It is playing a degree of freedom the rules leave open, available only to providers granted private-testing access — which is the next problem.
The data is handed out lopsided, and data is a competitive advantage
The second finding in The Leaderboard Illusion is structural rather than strategic, and it compounds the first. The Arena does not sample all models equally. Proprietary closed models “are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives.” Both policies push Arena data — the prompts real users send and the votes they cast — toward a few providers, and the concentration is stark: “Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data.”
Two providers with roughly 40% of the data between them; eighty-three open-weight models splitting under a third. Were that data inert — just a tally feeding the public number — the asymmetry would be a fairness complaint, not a measurement problem. But it is not inert: access to the Arena distribution is itself worth points on the Arena. The paper estimates that “even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates,” and draws the conclusion the whole post is circling: “Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality.”
That is the loop that makes the rank a measure of access. A provider sampled more often sees more of the live prompt distribution, can tune toward it, and climbs — not because the model generalizes better, but because it has been fitted to the very distribution the leaderboard scores on. The climb then justifies more sampling. The model that got more games to play looks better at the game, and “the game” and “general capability” have quietly come apart. The audit is no toy simulation, either: it combines data sources “encompassing 2M battles, auditing 42 providers and 243 models across a fixed time period (January 2024 - April 2025).”
A thousand votes can move the board
Selective disclosure and data asymmetry are exploits of access. The next one is an exploit of the voting mechanism itself, available to anyone willing to spend a little effort, and it is the part most people assume is impossible because the battles are anonymous.
Anonymity is the assumed defense: you cannot vote strategically for model X if you cannot tell which blurred answer came from X. Exploring and Mitigating Adversarial Manipulation of Voting-Based Leaderboards (arXiv 2501.07493) removes that assumption. Despite anonymized battles, an adversary “can determine which model was used to generate a given reply with more than 95% accuracy” — deanonymizing the model from the texture of its own output. Once you know which side is which, you can vote with intent, and the budget is small: an attacker can “alter the leaderboard (to promote their favorite model or demote competitors) at the cost of roughly a thousand votes.” A rounding error against the value of a top-five placement in a launch announcement.
The deanonymization is the load-bearing step, and it generalizes for a reason: models have stylistic fingerprints, and a classifier learns them. The same property that lets a watermark detector flag generated text — and that, as it happens, does not survive a paraphrase — lets an attacker fingerprint the model behind an anonymous Arena reply with better than 95% accuracy. Anonymity was the entire defense, and it leaks.
A second paper makes the budget look even worse by exploiting the scoring rule rather than fighting it. Improving Your Model Ranking on Chatbot Arena by Vote Rigging (arXiv 2501.17858) starts from the obvious obstacle: your target model is rarely on screen. “On average only about 1% of new battles will involve” the model you want to promote, so voting only in your target’s own battles barely registers. The paper’s “omnipresent” strategy sidesteps this through the Elo update rule — it moves a target’s ranking “even if” it “is not directly involved in the battle,” by voting in other battles in the directions that rearrange the surrounding Elo field. The result: “omnipresent rigging strategies could improve model rankings by rigging only hundreds of new votes.” Hundreds, not thousands, by attacking the math of the ranking — validated against “around 1.7 million historical votes from the Chatbot Arena Notebook,” real data, not a simulation.
Even the honest votes reward presentation over substance
Suppose every vote is genuine — no rigging, no deanonymization, no private best-of. Is the honest aggregate then a clean read on quality? Not quite, and the platform’s own operators showed why.
The LMSYS analysis Does style matter? Disentangling style and substance in Chatbot Arena regresses out answer length and markdown formatting from Arena outcomes and reports plainly that “Style indeed has a strong effect on models’ performance in the leaderboard.” Fit coefficients to the style features and the ranking turns out to measure presentation as much as content: answer length carried a coefficient of 0.249, with markdown lists at 0.031, headers at 0.024, and bold at 0.019 — and “Length was the dominant style factor. All other markdown effects are second order.” Voters, in aggregate, prefer the longer, better-formatted answer, holding substance roughly constant.
The consequence is that the uncontrolled leaderboard is partly a formatting contest, and controlling for style materially reorders it. With style control applied, “GPT-4o-mini” drops from rank 6 to rank 11 and “Grok-2-mini” falls from rank 6 to rank 18, while “Claude 3 Opus” climbs from rank 16 to rank 10 and “Claude 3.5 Sonnet” rises to tie for #1 on hard prompts. Those are double-digit rank moves produced by removing a degree of freedom a provider can optimize directly. Train a model to write longer and lard its answers with headers and bold without making it one bit more correct, and the leaderboard rewards you for it — until someone regresses the style back out, at which point your rank was partly borrowed.
Why the mechanism is gameable at all
It helps to step back from any single exploit and ask why an arena leaderboard has so many open seams at all. Gaming the Arena: AI Model Evaluation and the Viral Capture of Attention (arXiv 2512.15252) offers the framing: the rise of gladiatorial model arenas is powered by a “‘viral’ desire to capture attention both in, and outside of, the AI community,” and “arena gaming” describes a phenomenon through which “model developers hope to capture attention” rather than only demonstrate capability. The leaderboard is not a neutral instrument providers happen to compete on; it is a contested attention surface, and that attention is “critical to the scaling and commercialization of AI products.”
That reframes every exploit above as rational rather than aberrant. If a top placement converts directly into attention, funding, and adoption, then testing twenty-seven variants to publish the best, lobbying for more sampling, or buying a thousand votes are not deviations from the game — they are the game, played by participants who understand the prize is the ranking itself, not the quality it was meant to track. A measurement that pays out in attention gets optimized for attention. Goodhart’s law with a leaderboard attached.
What this argument does not claim
The case overstates if you let it, so three limits.
First, the Arena is not worthless. A live human-preference signal over real prompts measures something the static benchmarks miss; the failures here are about reading a rank as a clean quality ordering, not about the signal carrying no information. The mistake is the precision implied by “rank 4 beats rank 7,” not the act of consulting human preference.
Second, the platform is not standing still, and the honest reading credits the defenses. The vote-manipulation work was disclosed to Chatbot Arena, which deployed mitigations — “bot protection with Cloudflare, malicious user detection, rate limiting,” plus “reCAPTCHA and login” measures that “substantially increases the cost of such attacks.” But read what that concedes: it raises cost, “rather than making them impossible.” The vulnerability is structural to crowdsourced voting; defenses move the price of an attack, they do not close the seam. Likewise LMSYS built style control and publishes a style-controlled board — the operators are the ones surfacing the problem. A fair scorecard credits that while noting the default number many people quote is still the uncontrolled one.
Third, the numbers belong to their audits. The 19.2% and 20.4% data shares, the 27 variants, the up-to-112% gain, the ~1,000-vote and hundreds-of-votes budgets, the >95% deanonymization, the style coefficients — each is a measurement from a specific study of a specific platform over a specific window, not a universal constant. The mechanism generalizes to any access-asymmetric, crowd-voted leaderboard; the exact magnitudes are estimates with their own assumptions. Cite the seam, not the decimal.
Reading a leaderboard like a mechanism, not a measurement
The practices below read the rank with its conditions attached, the way the agent-benchmark post insists a score be quoted with its conditions.
Prefer the style-controlled board to the raw one. Where the operator publishes a style-controlled ranking, read that one — it removes the largest known degree of freedom a provider can optimize without improving substance. A rank that survives style control is worth more than one that does not.
Discount the freshly-launched chart-topper. A model debuting near the top right after a launch is exactly where selective disclosure and launch-window sampling bite hardest. Treat a brand-new #1 as the upper tail of an unknown number of private draws until broader sampling has washed the selection effect out.
Do not break a near-tie on rank alone. Two models within a handful of Elo points are, given the access asymmetries and the movability of the board, not reliably ordered. A small static delta can be sampling noise; a small Arena delta can also be access noise. Break the tie on your own task, not on their rung.
Ask who had the data. A provider sampled at many times the rate of its rivals has been fitted to the live distribution the board scores on — a reason its rank may not transfer to your distribution. The up-to-112% finding is the warning that Arena-specific gains are a real and separable thing.
Treat the Arena as one input, never the gate. A live preference leaderboard is a useful, gameable, attention-driven signal. Put it next to your own evaluation — your task, your distribution — and let it inform rather than decide. The rank is evidence about a contested mechanism; your eval is evidence about your use.
The rank is a move in a game you are not playing
Before a leaderboard rank is allowed to settle a model decision:
Reading list
The Elo number was honest arithmetic over real votes, and the team that broke its tie on it read the arithmetic correctly — rank 4 does sit above rank 7. Their error was treating a mechanism as an instrument. A live preference leaderboard hands more games, more data, and more private attempts to whoever already has access, prices a board-moving vote campaign at about a thousand votes, and rewards the longer answer for being longer. The rank was set partly by who got to play more games, and a number set by the game is not a clean measurement of the player.