ReasonBench v2: test items found in public training data; scores under review
Maintainers say about one in eight test items appear verbatim in a widely used web corpus. Leaderboard scores are frozen while affected items are removed.
Latest development, 12 Mar, 07:00
ReasonBench corrections reshuffle the top of the leaderboard
Two models swap places at the top after the cleaned set is applied; the gap between them is within the benchmark's stated margin.
- Heat at demo clock
- 30.1
- Independent sources
- 5
- Reports
- 6
- First seen
- Last report
- In briefings
- Wed 11 Mar, Thu 12 Mar
Timeline and provenance
Every report on this event, oldest first. All times UTC.
-
ReasonBench v2: test items found in public training data; scores under review
Kestrel Benchmarks Notes Official, tier 1
Maintainers say about one in eight test items appear verbatim in a widely used web corpus. Leaderboard scores are frozen while affected items are removed.
Demo report: no original link exists.
-
A popular reasoning benchmark may be contaminated, its maintainers say
The Latent Weekly Press, tier 2
Kestrel Benchmarks has paused ReasonBench scoring after finding test questions in public training data, raising questions about recent leaderboard gains.
Demo report: no original link exists.
-
ReasonBench contamination: what it means for the leaderboard
Builders Forum top threads Community, tier 3
A thread collecting reactions; several posters ask which published model scores rely on the affected items.
Demo report: no original link exists.
-
Statement on ReasonBench re-scoring
Meridian Models Newsroom Official, tier 1
Meridian Models says it will re-run its published ReasonBench results once Kestrel Benchmarks releases the cleaned test set, and will correct its model card if scores change.
Demo report: no original link exists.
-
ReasonBench v2.1 published with 412 items removed; leaderboard corrected
Kestrel Benchmarks Notes Official, tier 1
The cleaned test set drops 412 contaminated items. Corrected scores are published for every model that was on the leaderboard.
Demo report: no original link exists.
-
ReasonBench corrections reshuffle the top of the leaderboard
Circuit Review Press, tier 2
Two models swap places at the top after the cleaned set is applied; the gap between them is within the benchmark's stated margin.
Demo report: no original link exists.
How the heat adds up
10 × 0.5age / 24 h per independent source, newest report only, within 48 h.
| Source | Report | Age | Weight | Points | Why |
|---|---|---|---|---|---|
| Kestrel Benchmarks Notes | 10 Mar, 09:00 | 2.0 days | – | 0.00 | older than 48 h |
| The Latent Weekly | 10 Mar, 12:00 | 45 h | 0.273 | 2.73 | counted |
| Builders Forum top threads | 10 Mar, 13:30 | 43.5 h | 0.285 | 2.85 | counted |
| Meridian Models Newsroom | 11 Mar, 15:00 | 18 h | 0.595 | 5.95 | counted |
| Kestrel Benchmarks Notes | 12 Mar, 06:00 | 3 h | 0.917 | 9.17 | counted |
| Circuit Review | 12 Mar, 07:00 | 2 h | 0.944 | 9.44 | counted |
| Total | 30.1 | 5 sources counted | |||
6 hours earlier, at 12 Mar, 03:00 UTC, the heat was 16.7, so the trend is rising (+13.4). How trends are labelled
Why these reports are one event
Joined when similarity reaches 0.35: 0.6 × shared entities + 0.4 × shared wording, and at least one shared entity.
| Report | Decision | Evidence |
|---|---|---|
| Kestrel Benchmarks Notes, 10 Mar, 09:00 | new | Started this event: no open event scored above the threshold. |
| The Latent Weekly, 10 Mar, 12:00 | joined | Score 0.682 against the Kestrel Benchmarks Notes report: entities 1.00, wording 0.21. Shared: Kestrel Benchmarks, ReasonBench. |
| Builders Forum top threads, 10 Mar, 13:30 | joined | Score 0.362 against the Kestrel Benchmarks Notes report: entities 0.50, wording 0.16. Shared: ReasonBench. |
| Meridian Models Newsroom, 11 Mar, 15:00 | joined | Score 0.457 against the The Latent Weekly report: entities 0.67, wording 0.14. Shared: Kestrel Benchmarks, ReasonBench. |
| Kestrel Benchmarks Notes, 12 Mar, 06:00 | joined | Score 0.680 against the Kestrel Benchmarks Notes report: entities 1.00, wording 0.20. Shared: Kestrel Benchmarks, ReasonBench. Closest other event: Introducing Orchard 3: open-weight models in three sizes at 0.189. |
| Circuit Review, 12 Mar, 07:00 | joined | Score 0.657 against the Kestrel Benchmarks Notes report: entities 1.00, wording 0.14. Shared: Kestrel Benchmarks, ReasonBench. Closest other event: Introducing Orchard 3: open-weight models in three sizes at 0.174. |
Named in this event
ReasonBench, Kestrel Benchmarks, Meridian Models