Enhancing Instruction-Following Capabilities in Seq2Seq Models: DoLA Adaptations for T5

TL;DR

This paper investigates why instruction-following in Seq2Seq models like FLAN-T5 often fails due to volatile internal representations and introduces a gradient-based method to steer models towards better instruction compliance, significantly improving performance.

Contribution

The paper adapts DoLa to FLAN-T5 to analyze internal representations and proposes a gradient-based activation-steering technique to enhance instruction-following capabilities.

Findings

Decoder layers show rapid shifts driven by cross-attention.

Token preferences are highly volatile across layers.

Steering improves MemoTrap performance from 52% to 99.7%.

Abstract

Encoder-decoder models such as FLAN-T5 are finetuned to follow instructions, but often fail when the instructions conflict with memorized continuations ingrained during training. To understand this behavior, we adapt DoLa to FLAN-T5 and examine how representations evolve in the decoder. Our findings show that T5's intermediate layers undergo rapid shifts driven by cross-attention to the encoder. When projected through the language modeling head, each depth presents highly volatile token preferences, leading to unreliable behavior with contrastive decoding. Motivated by this, we introduce a gradient-based activation-steering method that injects an "instruction-compliance" direction into mid-decoder layers, where the representation is both meaningful and still malleable. This intervention dramatically improves MemoTrap performance (52% to 99.7%), demonstrating that mechanistic steering can succeed where contrastive decoding fails in Seq2Seq architectures.