Sparse Activation Editing for Reliable Instruction Following in Narratives

2025/05/22 by Runcong Zhao, Zhao, Runcong, Cao, Chengyu +12 · 2 citations
Computer Science · #Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #FOS: Computer and information sciences #Human-Computer Interaction (cs.HC) #Multimodal Machine Learning Applications #Text Readability and Simplification #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2505.16505

openalex publication_date 2025/05/22 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Complex narrative contexts often challenge language models' ability to follow instructions, and existing benchmarks fail to capture these difficulties. To address this, we propose Concise-SAE, a training-free framework that improves instruction following by identifying and editing instruction-relevant neurons using only natural language instructions, without requiring labelled data. To thoroughly evaluate our method, we introduce FreeInstruct, a diverse and realistic benchmark of 1,212 examples that highlights the challenges of instruction following in narrative-rich settings. While initially motivated by complex narratives, Concise-SAE demonstrates state-of-the-art instruction adherence across varied tasks without compromising generation quality.

Citations

• RoleMRC: A Fine-Grained Composite Benchmark for Role-Playing and Instruction-Following
• SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models
• Low-Rank Adaptation for Foundation Models: A Comprehensive Review
• An Automated Explainable Educational Assessment System Built on LLMs
• LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
• Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders
• Steering Knowledge Selection Behaviours in LLMs via SAE-Based Representation Engineering
• Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2
• WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs
• PLAYER*: Enhancing LLM-based Multi-Agent Communication and Interaction in Murder Mystery Games
• Learning to Edit: Aligning LLMs with Knowledge Editing
• In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering
• Representation Engineering: A Top-Down Approach to AI Transparency
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• Steering Language Models With Activation Engineering
• InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback
• Locating and Editing Factual Associations in GPT
• Language Models are Few-Shot Learners
• Null It Out: Guarding Protected Attributes by Iterative Nullspace\n Projection
• A Tutorial on Bayesian Optimization
• The CMA Evolution Strategy: A Tutorial

Cited by

• Enhancing LLM Steering through Sparse Autoencoder-Based Vector Refinement
• Iterative Multilingual Spectral Attribute Erasure

Related

• Instruction Tuning for Story Understanding and Generation with Weak Supervision
• SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models
• MUFFIN: Curating Multi-Faceted Instructions for Improving Instruction-Following
• Non-instructional Fine-tuning: Enabling Instruction-Following Capabilities in Pre-trained Language Models without Instruction-Following Data
• Spotlight Your Instructions: Instruction-following with Dynamic Attention Steering