Saved in:
Bibliographic Details
Main Authors: Sun, Huey, Yong, Anabel, Gilly, Lorenzo, Jin, Felipe
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.03803
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915670432677888
author Sun, Huey
Yong, Anabel
Gilly, Lorenzo
Jin, Felipe
author_facet Sun, Huey
Yong, Anabel
Gilly, Lorenzo
Jin, Felipe
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2512_03803
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Enhancing Instruction-Following Capabilities in Seq2Seq Models: DoLA Adaptations for T5
Sun, Huey
Yong, Anabel
Gilly, Lorenzo
Jin, Felipe
Computation and Language
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.
title Enhancing Instruction-Following Capabilities in Seq2Seq Models: DoLA Adaptations for T5
topic Computation and Language
url https://arxiv.org/abs/2512.03803