Tag
OrDA is a framework that disentangles user access habits from genuine content interests in homepage marketing block recommendations using orthogonal regularization and causal intervention, achieving a 5.64% UCTR improvement on Zhima's homepage.
This paper introduces Retroactive Advantage Correction (RAC), a closed-form bias correction method for delay-aware RLHF that handles asynchronous reward signals by queuing and reinjecting delayed rewards with a V-trace-style clipped residual update.
This paper introduces SLC (State-space Logit Correction), which corrects per-item logit bias in knowledge tracing models using empirical-Bayes shrinkage via a Kalman smoother, improving AUC beyond global calibration techniques.
This paper introduces PRECISE, an extension of Prediction-Powered Inference that combines a small set of human labels with a large set of LLM judgments to produce unbiased and variance-reduced estimates of ranking evaluation metrics like Precision@K. The method is validated on the ESCI benchmark and in a production A/B test, where it correctly identified the best system variant using only 100 human labels, confirmed by a +407 bps sales improvement.
This paper identifies a bias in attention weights caused by quantizing keys in KV-cache compression for chunk-wise autoregressive video diffusion, and proposes a per-attention-score correction that removes the bias with negligible overhead, recovering near-BF16 video quality at INT2 quantization.
This paper presents PushCen-ADFL, a communication-efficient asynchronous decentralized federated learning framework that uses centroid-based messaging and bias-correction to improve accuracy and reduce communication overhead under heterogeneous conditions.