Evaluation infrastructure for the multi-model era
Same prompt. Two models, names hidden. A person reads both and decides.
DualMind Labs builds the infrastructure for evaluating AI models and agents.
Free · sign in with Google, GitHub or email
Test questions leak into training sets. Scores inflate. Nobody knows by how much.
MMLU, GSM8K, HumanEval: saturated. At the ceiling, the gaps between models stop meaning much.
A measure that becomes a target stops being a measure.
Static test. Moving target.
Sources: Sainz et al. 2023 · Zhang et al. 2024 · Stanford AI Index 2025
No answer key to memorize. Just a person reading, then choosing.
Names stay hidden until the vote is cast. Reputation cannot lean on the scale.
In published arena research, crowd votes agreed with expert raters 72 to 83 percent of the time.
Anonymous. Side by side. Human.
Sources: Chiang et al. 2024
Free account / arena.dualmindlab.tech
Live standings built from blind, side-by-side decisions, not a frozen test set.
Pairwise human preference is the method behind some of the field's most cited rankings.
Human preference has biases too: length, polish. We name them rather than hide them.
Live / moves with every vote
Models on identical prompts. Agents on the same task. Workflows built to judge both.
Evaluation is infrastructure now, not the last slide of a launch deck.
Built in public, so the method can be checked, not just trusted.
Models / agents / structured evaluation
Open the Arena with a free account. Bring a question that matters to you.
Two minds. One question. A human decides.
Founded in 2025 in Gujarat, India, by Harsh J. Bhanushali and Rishi A. Bangoriya
Free · sign in with Google, GitHub or email
Pick a prompt. Two anonymous answers appear. Vote, then see what you chose.
Demo · answers written for this page
Model A
Two particles can be linked so that measuring one tells you about the other, however far apart they are. Picture gloves shipped in separate boxes: open one, see a left glove, and you know the other is right. The quantum twist is that neither was decided until you looked. It can't send messages faster than light, but it powers quantum computing and secure encryption.
Model B
Quantum entanglement is a fascinating and complex phenomenon that has puzzled scientists for decades. In simple terms, it is a special connection between particles. When particles become entangled, they share a kind of relationship, and what happens to one is connected to what happens to the other. It remains one of the most interesting areas of modern physics, with many potential applications in the future of technology.
Which response was better?
DualMind Labs is an AI research and product lab building infrastructure for evaluating and comparing AI models and agents, founded in 2025 in Gujarat, India.
In the Arena, names stay hidden until you vote. Here are ours.
The story so far
I · 2025
DualMind Labs founded in Gujarat, India.
II · 2025–26
DualMind Arena developed and launched.
III · 2026
Expanding into AI agent evaluation and infrastructure.
Built with everything we have.
A free arena at arena.dualmindlab.tech; sign in with Google, GitHub or email. In Battle mode two anonymous models answer the same prompt; you pick the better one, call a tie, or mark both bad, and the names appear after your vote. Side by Side lets you choose the two models, and Direct Chat lets you try one on its own. Every vote feeds a live leaderboard.
Benchmark questions leak into training data and inflate scores by an unknown amount. MMLU, GSM8K and HumanEval are saturated, so top scores no longer separate models. A score that becomes a target stops being a measure.
Published arena research found crowd votes agreed with expert raters 72 to 83 percent of the time. Human preference has known biases, such as favoring longer or prettier answers, which is why names stay hidden until after the vote.
Model evaluation on identical prompts, agent evaluation with multiple agents on the same task, and infrastructure for structured evaluation workflows.
DualMind Labs was founded in 2025 in Gujarat, India, by Harsh J. Bhanushali and Rishi A. Bangoriya. The work is built in public at github.com/orgs/dualmindlabs.