Models#

Eight implementations, each on its own page: the problem it attacks, the method with its objective written out, the training loop step by step, a map from the method to the code, and the keys that run it. Read Select-then-predict first if the architecture is new to you, since every page below names a difficulty stated there.

All eight answer one question, which is how to stop a selector and a predictor trained together from settling on an uninformative highlight they both agree about.

Model	What it changes	Venue	Cost
FR: folded rationalization	One encoder shared by selector and predictor, so neither can drift into a representation the other does not hold.	NeurIPS 2022	One encoder fewer
MGR: multi-generator rationalization	Several generators against one shared predictor, so no single generator dictates the equilibrium.	ACL 2023	One encoder per generator
DR: decoupled rationalization	The predictor’s learning rate scaled by what the selection kept, which restrains it while the selection is poor.	KDD 2023	None, it changes no loss
MCD: d-separation for causal self-explanation	A second prediction from the full input, and a highlight trained to make the two agree.	NeurIPS 2023	Two phases per batch
G-RAT: guidance-based rationalization	An attention classifier over the full input, supervising the selection and matched in distribution.	AAAI 2024	A third encoder, pretrained
DAR: discriminatively aligned rationalization	A frozen aligner that only ever read full text, scoring the highlight it is handed.	ICDE 2024	A pretraining loop
MRD: maximizing the remaining discrepancy	The question reversed: what the complement can still say, maximised rather than minimised.	NeurIPS 2024	Two phases per batch
GenSPP: interlocking-free rationalization through genetic search	No gradient descent on the generator at all, and a genetic search in its place.	ACL 2025	A predictor per candidate

GroundedSPP: select-then-predict over a knowledge base answers a different question. It grounds a select-then-predict model in a corpus’s knowledge base, for corpora that explain their labels with free text rather than with spans.

Writing a method of your own rather than running one of these is Writing your own method, which builds a small architecture end to end and names what to override for what.

Every architecture is registered for a GRU backbone and for a Transformer one. The algorithms mention neither, since a backbone is anything implementing encode, pool and output_size, so swapping one for the other is a key rather than a code change.

Reading order#

The order above is chronological, and it is also the order the pages read best in. FR: folded rationalization is the shortest intervention and the best place to see the base architecture with nothing added to it, while MGR: multi-generator rationalization and DR: decoupled rationalization change the optimizer rather than the objective and are the two cheapest departures from it. The phased pair MCD: d-separation for causal self-explanation and MRD: maximizing the remaining discrepancy are worth reading together, since the second reverses the question the first asks and shares its training loop. GenSPP: interlocking-free rationalization through genetic search is last because it abandons the assumption every other page shares, which is that the generator is trained by gradient descent at all.