Radical Centrism or Leaderboardism
From the AI Research Compass wiki
Radical Centrism or Leaderboardism: Every camp has a point and none deserves loyalty; use whatever method currently tops the leaderboard.
Beliefs
Radical Centrism or Leaderboardism is the dead centre of the chart, earned by answering every debate with “it depends”. Language or world? Both, weighted by the task. One model or a harness? Whichever ships. New paradigm or more scale? Ask again after the next benchmark release. In its sincere form this is principled empiricism: ideology is a prior, evaluation is the posterior, and the leaderboard is where priors go to be updated. In its less flattering form it is a refusal to hold any view that a benchmark has not yet confirmed.
Its central position and Steady shade mean exactly that: no lean on either axis, no bold paradigm bet, no strong feelings about openness or speed. The double name is the joke and the point. The same person can be a careful pragmatist who weighs evidence from every camp, or someone whose research agenda is set by whatever topped a public ranking last week.
Arguments
- Holistic evaluation across many scenarios and metrics shows that no single approach dominates everywhere, which vindicates case-by-case choices.
- Public preference-based arenas let methods compete on equal terms, giving a shared, checkable reference point instead of competing narratives.
- History punishes dogma: approaches dismissed by one camp, from neural networks to retrieval, have repeatedly come back to top the rankings.
Critiques
- Benchmarks saturate, leak into training data and get gamed, so chasing the leaderboard often optimises the measurement rather than the capability.
- Leaderboards only reward what they can measure today, which systematically disadvantages long-term bets that have not paid off yet.
- Refusing to commit to any hypothesis is not neutrality; it is outsourcing the research agenda to whoever designs the benchmarks.
Representative works
- Holistic Evaluation of Language Models2022
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference2024
Relationships
Friends
How to get it
Land on this cell in the full or pro test: x ≈ +0.00, y ≈ +0.00, paradigm ≈ +0.00.