Improved language model with reduced hallucination

Abstract

41 20 26 20 06 62 30 J an 2 02 6 A B S T R A C T 2 0 2 6 2 0 0 6 6 2 3 0 J a n 2 0 2 6 4 1 C START 200 RECEIVE AN INPUT FOR A LANGUAGE MODEL HAVING A NUMBER OF DECODER LAYERS 202 PREDICT, BY THE LANGUAGE MODEL, A CHUNK OF TOKENS BASED ON THE INPUT DETERMINE, FROM THE CORRESPONDING ATTENTION MAP WEIGHTS OF EACH OF THE ATTENTION r 204 HEADS OF THE DECODER LAYERS, A LOOKBACK RATIO VECTOR FOR THE CHUNK OF TOKENS PREDICT, USING THE LOOKBACK RATIO VECTOR, A HALLUCINATION SCORE FOR THE CHUNK OF r 206 TOKENS DETERMINE THAT THE HALLUCINATION SCORE FAILS TO SATISFY A PREDEFINED THRESHOLD, / 1., r 208 SUCH THAT THE CHUNK OF TOKENS REPRESENTS A HALLUCINATION OF THE LANGUAGE MODEL DETERMINE, RESPONSIVE TO THE CHUNK OF TOKENS REPRESENTING THE HALLUCINATION, AN EDIT r 210 DIRECTION FOR THE CORRESPONDING ATTENTION MAP WEIGHTS FOR THE CHUNK OF TOKENS MODIFY, TO GENERATE A POST-PROCESS LANGUAGE MODEL, THE LANGUAGE MODEL BY r ADJUSTING, USING THE EDIT DIRECTION, THE CORRESPONDING ATTENTION MAP WEIGHTS OF AT 212 LEAST SOME OF THE ATTENTION HEADS C END FIG. 2 2 / 6 START 200 204 206 208 212 20 26 20 06 62 30 J an 2 02 6 2 0 2 6 2 0 0 6 6 2 3 0 J a n 2 0 2 6 R E C E I V E A N I N P U T F O R A L A N G U A G E M O D E L H A V I N G A N U M B E R O F D E C O D E R L A Y E R S 2 0 2 2 0 4 2 0 6 2 0 8 D I R E C T I O N F O R T H E C O R R E S P O N D I N G A T T E N T I O N M A P W E I G H T S F O R T H E C H U N K O F T O K E N S 2 1 2 L E A S T S O M E O F T H E A T T E N T I O N H E A D S E N D

Images