Method Drift›LLM reasoning / chain-of-thought
DualHSIC
DualHSIC: HSIC-Bottleneck and Alignment for Continual LearningLLM reasoning / chain-of-thought · first seen Apr 30, 2023
superseded — cited as a baseline and beaten by newer methods
1 papers critique it · 1 beat it on benchmarks
What papers say
Verbatim critique sentences, each from a paper that cites DualHSIC as a baseline.
“In contrast to STAR, OCM and DualHSIC do not consider the change in model outputs in the local parameter neighbourhood”
— STAR: Stability-Inducing Weight Perturbation for Continual Learning
Beaten on benchmarks
Head-to-head results where a newer method reports beating DualHSIC. Values are copied from the source paper's tables — verify against the cited paper.
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+ER-ACE beats DualHSIC · accuracy [ER-ACE + buffer size 100]
60.69 vs 57.03
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+ER-ACE beats DualHSIC · accuracy [ER-ACE + buffer size 200]
67.58 vs 64.05
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+ER-ACE beats DualHSIC · accuracy [ER-ACE + buffer size 500]
75.44 vs 72.35
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+ER-ACE beats DualHSIC · accuracy [ER-ACE + buffer size 2000]
51.67 vs 49.94
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+ER-ACE beats DualHSIC · accuracy [ER-ACE + buffer size 1000]
21.06 vs 20.75
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+DER++ beats DualHSIC · accuracy [DER++ + buffer size 100]
61.76 vs 58.90
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+DER++ beats DualHSIC · accuracy [DER++ + buffer size 200]
68.60 vs 67.11
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+DER++ beats DualHSIC · accuracy [DER++ + buffer size 500]
76.52 vs 74.34
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+DER++ beats DualHSIC · accuracy [DER++ + buffer size 1000]
22.4 vs 21.73
- STAR: Stability-Inducing Weight Perturbation for Continual Learning
STAR+DER++ beats DualHSIC · accuracy [DER++ + buffer size 2000]
28.19 vs 27.42