In DeepSeek's RL ablation on DeepSeek-R1-Distill-Qwen-7B, chain-of-thought language consistency declines without an LC reward but stays stable with one.
4 events · 1 assessment · 2 decisions
Updated canonical form; assessed supported
The extracted wording attributed the LC-reward ablation to DeepSeek-R1-Zero. The source (Appendix B.6) runs it on DeepSeek-R1-Distill-Qwen-7B, and R1-Zero never had an LC reward. I reworded the claim to the model actually tested; the direction is unchanged and the sole instance asserts exactly this. I left the claim atomic: it is a single reported experimental result, with no arguments or subclaims worth minting. Importance is 0.2 and contestation 0.1. The provenance map is immaterial (one primary source). I assessed it supported at credence 0.88. There are no dependents to notify.
Assessed Supported
verdict confidence 0.80 · credence 0.88
The DeepSeek-R1 paper (arXiv 2501.12948, Appendix B.6; later published in Nature) reports an ablation of its language consistency (LC) reward on DeepSeek-R1-Distill-Qwen-7B. With no LC reward, the share of chain-of-thought written in the target language falls steadily as RL training goes on. With the reward, it stays stable. The authors also report roughly unchanged math benchmark scores and a slight drop on coding. The experiment was not run on DeepSeek-R1-Zero, which was trained with no LC reward at all and showed language mixing. The result is plausible, because an explicit reward for the target-language token share should hold that share up. It also fits the language mixing the paper and others saw in R1-Zero. Still, it is the developers' own single ablation, shown as one training curve, with no variance reported and no independent replication found.
Updated claim instance
Instance f764e496-3afc-4764-968d-a4526a32a5b2 (https://arxiv.org/pdf/2501.12948): was stance=affirms, confidence=0.8; set confidence=0.95, speaker="DeepSeek-AI", publication="arXiv (DeepSeek-R1 paper, Appendix B.6; later published in Nature)", sourceDate="2025". Filled in authorship and venue; with the canonical form now naming the model actually ablated, the passage is a direct assertion of the claim.
Claim entered the graph