Amazon's Plumbline finds silent bugs that a falling training loss hides
An Amazon team describes a set of checks that test each part of a language model training run against a trusted reference. Over two months, it found 61 problems across five new models, and about half raised no error at all.

AI illustrationWhat the study shows
Teams that adapt a released language model usually train it with open-source software, called training frameworks. These frameworks split the model across many graphics processors (GPUs). A fault can sit in the data handling, the saved files, the maths of each update or the network. Many such faults do not crash anything. The main health signal, the loss (a running score of how wrong the model is), keeps falling as normal.
The authors built Plumbline to test each part of a run on its own. Each check compares the framework's output with a reference the framework did not produce. Examples include the model maker's official chat template and original weights, or a slow, high-precision recalculation on an ordinary processor. Where an exact match is impossible, the check allows only a margin set above the measured random variation between runs.
The checks work at three levels. Some cover each stage, from preparing data to saving, resuming and exporting the model. Some test one training step under every supported setting. Others run short full training runs. A coding agent, an AI system that writes and runs code, chooses and runs the checks. A human reviewer approves its plans and confirms every problem.
The authors report results from September to October 2026. They scanned 5 newly released models on 3 frameworks, 13 combinations in all. The scan found 61 problems, and 32 were silent. The part-by-part checks caught 29 of those 32 first. At least 16 silent problems sat in runs that passed the short full-run checks. Of the 61, 42 are fixed or reported to the framework developers, and 8 were first raised by the authors.
The examples are concrete. In one model, loading the checkpoint (the saved weights) turned some weight tables on their side. The run started far worse than it should, yet its loss still fell and it passed the full-run checks. In another, the data setup never taught the model the token that ends a reply, so it was not trained to stop. In a third, a resumed run forgot how many update steps it had taken.
The same checks were used to trace each failure to its cause. The authors say all 61 were located, and 50 were traced down to a specific line of code. In one case, a crash on eight machines was explained using one machine for about half a node-hour, meaning one machine running for about thirty minutes. This is a preprint and has not been peer reviewed.
Our reading
In our reading, the central message is simple and useful: a falling loss is not proof that training is correct. The paper backs this with specific, checkable cases rather than general warnings. The transposed weights example is striking. The run started from a much worse point than a correct load would give, and the run-level checks still passed it, because they judge stability, not the value of the loss.
The design idea is sound engineering. Each check answers one narrow question against a reference that the code under test did not produce. That makes a failed check point to a likely location, and it lets the same checks be reused when hunting for a cause. A practical example: a team about to fine-tune a newly released model could run the data check first. It would learn, before booking any GPUs, that its setup drops the end-of-reply token.
The evidence is a count of problems the authors found in their own system, written by the system's builders. There is no test of how many faults it misses, no false alarm rate and no comparison with other tools. On our four-level readiness scale, we rate this I, Lab only. No public benchmark result exists and the code is not yet released. Amazon's production use is stated by its own authors without usage detail, so we do not count it as level IV evidence.
Read the rest of this story
Membership is free. You get every article in full and the Issue every Sunday.
We send a sign-in link, no password needed. Already a member? Sign inReadiness assessment
| Level | Status | Evidence |
|---|---|---|
| I Lab only | Met | The system, checks and rules are described in detail with worked examples. |
| II Benchmarked | Not yet | There is no public benchmark and no comparison with other tools. The evidence is the authors' own tally of problems found, plus a calibration of the step check on 3 models. This is arguably more than a lab demo, but it is not a benchmark result. Level 2 is proposed only loosely as the closest fit; strictly the criterion is not met. |
| III Open source | Not yet | The paper says 'We will release Plumbline as open source' but gives no link and no released code, weights or data. |
| IV In production | Not yet | The paper states Plumbline is used in production at Amazon, which names an organisation. It gives no documented usage details such as teams, volume or outcomes beyond the authors' own results, and it is self-reported by the vendor. We treat it as a claim, not independent proof. |
What would move it up
Releasing the code under an open licence would meet level 3. For level 2, the authors would need a public benchmark: for example, injected faults with measured detection and false-alarm rates, compared with existing tools. For level 4, they should document named users, the number of runs and the time saved, ideally with independent confirmation.
Checked automatically every week, last on 16 November.
The study
Ruochen Wen, Hongyang Wang, Siwei Li et al.. Plumbline: When the Loss Lies, Validating LLM Training Correctly. arXiv, 10 November 2026. doi:10.48550/arxiv.2611.13393
Notes
- Wen et al., Plumbline: When the Loss Lies, Validating LLM Training Correctly. arXiv preprint 2611.13393, 10 November 2026. Not peer reviewed. Source
- Code: the authors state they will release Plumbline as open source. No link or date was given at the time of writing.
- Exact figures from the paper: 61 problems, 32 silent; part-by-part (lifecycle and architecture) checks found 29 of the 32 silent problems first; full-run (convergence) checks found 21 problems first, 1 silent; at least 16 silent problems were in runs that passed convergence checks; 42 fixed or reported upstream, 8 first raised by the authors; 50 of 61 diagnoses reached the code line.
- Examples: gpt-oss-20b checkpoint loading transposed the down projections of all 32 experts, starting loss 17.08 against 1.48 for a correct load. GLM-5.3-Flash resume without restored optimizer step count gave loss differences up to 3.7%, against at most 0.02% between repeated uninterrupted runs. Step-check calibration on 3 models and 3 frameworks: defect-free runs at or below 3.44 times the noise floor, failing defective runs at least 4.35 times, with a bound of 4.
- Readiness: I, Lab only. No public benchmark, no released code, and production use at Amazon is self-reported without usage details.
Edited and signed by
Stefan PerdalEditor and publisher86 stories editedStefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.
How this was made
An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.
How we workNo corrections