FINAL-Bench
A public benchmarking framework and competition infrastructure for GPU-accelerated quantum simulation. It defines evaluation standards, problem sets, submission protocols, and leaderboard mechanisms, and is used to assess the reasoning capabilities of AI systems on physics simulation tasks.
Timeline (last 90 days)
Counting the number of first-place positions held, the top organization is FINAL-Bench (11 first places), second is inclusionAI (4), and third is zai-org (4).
FINAL-Bench has a technical partnership with the leaderboard maintainer QuantID, and its parent company VIDRAFT appears in the rankings, so this daily report is not an independent evaluation.
FINAL-Bench's 11 first places come from its Darwin-180B-RSI series models, covering areas such as MMLU-Pro (88.12), gpqa (94.44), MMMU_Pro (79.48), the LEXam series, and hmmt_feb_2026 (a score of 100).
The Open Materials Challenge requires submitting lithium electrolyte crystals, with season 1 open until 2026-11-30.
The Open Discovery Challenge requires submitting drug candidate molecules (SMILES), with validation based on target and anti-target structure prediction, ADMET, and novelty; seasons 2 through 4 are open.
The Open Superconductor Challenge requires submitting d-wave pairing estimates for each two-dimensional material, with season 1 open until 2026-12-31.
The Open Quantum Challenge requires submitting Hubbard energies or surface-code decoders, with two seasons open until 2027.
LEADBOARD contains 21 leaderboards that require property prediction on held-out molecules, with labels never made public.
Hidden validators mitigate reward hacking by cutting the feedback loop, but cannot eliminate it entirely, since contestants can still indirectly infer scoring preferences from public rules and training data.
All FINAL-Bench challenge scores are computational predictions rather than laboratory measurements; a high score only represents a hypothesis worth testing, not a verified scientific fact.
10 more timeline events
All Facts (17)
FINAL-Bench's 11 first places come from its Darwin-180B-RSI series models, covering areas such as MMLU-Pro (88.12), gpqa (94.44), MMMU_Pro (79.48), the LEXam series, and hmmt_feb_2026 (a score of 100).
50%UnverifiedCounting the number of first-place positions held, the top organization is FINAL-Bench (11 first places), second is inclusionAI (4), and third is zai-org (4).
50%UnverifiedFINAL-Bench has a technical partnership with the leaderboard maintainer QuantID, and its parent company VIDRAFT appears in the rankings, so this daily report is not an independent evaluation.
50%UnverifiedThe Open Discovery Challenge requires submitting drug candidate molecules (SMILES), with validation based on target and anti-target structure prediction, ADMET, and novelty; seasons 2 through 4 are open.
50%UnverifiedThe Open Materials Challenge requires submitting lithium electrolyte crystals, with season 1 open until 2026-11-30.
50%UnverifiedThe Open Superconductor Challenge requires submitting d-wave pairing estimates for each two-dimensional material, with season 1 open until 2026-12-31.
50%UnverifiedFINAL-Bench launched five open science challenges, covering the four areas of drug discovery, materials science, superconducting physics, and quantum computing.
50%UnverifiedFINAL-Bench's grader is invisible to submitters, the answer key always stays on the organizers' hardware, and the rules, weights, and submission limits are all public.
50%UnverifiedLEADBOARD contains 21 leaderboards that require property prediction on held-out molecules, with labels never made public.
50%UnverifiedHidden validators mitigate reward hacking by cutting the feedback loop, but cannot eliminate it entirely, since contestants can still indirectly infer scoring preferences from public rules and training data.
50%UnverifiedAll FINAL-Bench challenge scores are computational predictions rather than laboratory measurements; a high score only represents a hypothesis worth testing, not a verified scientific fact.
50%UnverifiedThe validator is trained on historical data, may extrapolate poorly to genuinely novel candidates, and suffers from distribution shift, so high-scoring entries may fail in experiments.
50%UnverifiedThe Open Quantum Challenge requires submitting Hubbard energies or surface-code decoders, with two seasons open until 2027.
50%UnverifiedIn the count of first-place finishes by organization, FINAL-Bench ranks first with 11, while inclusionAI and zai-org tie for second with 4 each.
50%UnverifiedThe publisher QuantID is a technology partner of VIDRAFT, and FINAL-Bench is an organization under VIDRAFT, so the leaderboard involves a conflict-of-interest disclosure, and the original text states that this is not an independent evaluation.
50%UnverifiedFINAL-Bench is the framework and competition infrastructure of this benchmark, responsible for defining evaluation criteria, the problem set, and the leaderboard mechanism, with a role similar to an evaluation protocol layer.
50%UnverifiedDecide Live is built by VIDRAFT / FINAL-Bench and is open-sourced under the Apache-2.0 license.
50%