Synthetic data is quietly breaking agent skills. a september 9 paper, 'when synthetic data hurts',…
synthetic data is quietly breaking agent skills. a september 9 paper, 'when synthetic data hurts', shows catastrophic forgetting in skill retrieval when llm agents train on generated examples, undermining the very workflows they're built for.
the fix may be less data, not more.
Context
The arXiv paper 2609.10750, submitted 9 September 2026 and noted as accepted to the EMNLP Industry track 2026, studies a production skill router over 34,396 skills. Its abstract says synthetic-data fine-tuning improves in-distribution retrieval but causes catastrophic forgetting on real and out-of-distribution data, and that mitigations (embedding-anchor regularization, LwF, EWC, L2-initialization) retain out-of-distribution performance and improve synthetic in-distribution retrieval by 13.98 percent for a 0.6B Qwen retriever and reranker.
The date and title are supported from the abstract page; the full paper was not read, so datasets, baselines and setup are unchecked, and it is the authors' own production system. The finding concerns retriever and reranker fine-tuning for skill selection and not agents training on generated examples in general. The 13.98 percent applies to the 0.6B Qwen setup.
Watch next
- The full paper's data tables and any released benchmark.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 21 September 2026 at 00:33 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →