Conceptencoder averaging

Encoder Averaging

An ensemble-based model upgrade method that averages the predicted probabilities of multiple encoders, such as different pretrained models or random seeds, to reduce variance. The cost increase is limited, but the improvement is usually modest.

Timeline (last 90 days)

Nov 3

The encoder-averaging strategy helps on AspectBench and NewsMTSC but is harmful on SentiCoref.

Unverified50%

Nov 3

Both strategies selected on the development set (encoder averaging and selective LLM assistance) route 30% of records to the encoder-assisted LLM.

Unverified50%

All Facts (2)

Unverified

The encoder-averaging strategy helps on AspectBench and NewsMTSC but is harmful on SentiCoref.

50%Unverified

Both strategies selected on the development set (encoder averaging and selective LLM assistance) route 30% of records to the encoder-assisted LLM.

50%

Source Articles

Finding Errors Is Not Enough: Selective Model Escalation for Entity-Targeted Sentiment AnalysisNov 3