Quiet day · no new POST verdicts in 24h

We tell you which ML papers
are wrong.

By running the code. Publicly. On every paper on arXiv.

“Wrong” is a technical term — our reproduction did not match this claim's reported numbers. See methodology →

21k
papers indexed
186
reproductions run
0 public WRONG since the validator landed·0 of 54 verdicts claim-validated·0/10 anchors healthy

Latest verdicts

Most recent current POST verdicts, newest first.

See the Wall of Wrong →
PENDING
2 OLMo 2 Furious
· arXiv 2024 · cs.CL
PENDING
RoBERTa: A Robustly Optimized BERT Pretraining Approach
· arXiv preprint · cs.CL
PENDING
Mamba: Linear-Time Sequence Modeling with Selective State Spaces
· COLM 2024 · cs.LG
PENDING
DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter
· NeurIPS 2019 EMC^2 Workshop · cs.CL
PENDING
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
· NAACL 2019 · cs.CL
PENDING
Stable LM 2 1.6B Technical Report
· arXiv 2024 · cs.CL
PENDING
OLMoE: Open Mixture-of-Experts Language Models
· arXiv 2024 · cs.CL
PENDING
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
· arXiv 2025 · cs.CL
PENDING
SmolLM2: When Smol Goes Big — Data-Centric Training of a Small Language Model
· arXiv 2025 · cs.CL
PENDING
Qwen2.5 Technical Report
· arXiv 2024 · cs.CL

We run your code

Every reproduction starts from the official repo, runs in a clean Modal sandbox, and logs every command. No re-implementation.

Authors get the last word

Right of reply isn't a footer — it's a sidebar. Every verdict links to the author's response in the same view.

Evidence in one click

Each WRONG label is one click from the diff, the logs, and the reproduction job. Always.

The Wall of Wrong

A chronological feed of every paper that didn't reproduce on our reproduction job, sortable by venue, lab, and confidence. The signature page of paperiswrong.

Open the Wall →