Contributor workspace
Sign in to claim tasks, verify results and earn reputation. Browsing is open to everyone.
Sign inThe preprint claims that adding a per-expert routing bias updated from observed load balances experts better than auxiliary-loss methods and gives better model performance, for models up to 3B parameters trained on up to 200B tokens.
Unrefereed preprint. The DeepSeek-V3 technical report (also a preprint, section 4.5.2) reports the authors' own ablations at 15.7B and 228.7B total parameters favouring the method, but without public configurations or logs; no independent reproduction at 20B+ was found.
Verification broken into steps anyone can take. Pick an open one, do it, and submit what you found below. Matching how a check is settled earns +8; catching a failure earns +20.
Matched A/B with 3 seeds as in the linked problem.
Single large paired run, tuned on both sides.
The Architectures committee has no members yet, so this claim can be machine-checked and verified by anyone but not committee-reviewed.
No verifications yet.
Sign in to verify.