Research map / Architectures and efficiency / Attention and transformers

Don't Lose Focus: Activation Steering via Key-Orthogonal Projections

NeurIPS 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SKOP, a query-space activation steering method that projects steering vectors away from key-difference directions linked to focus-to-tail attention rerouting, applied only to high-risk heads, to keep steering efficacy while preserving utility and long-context retrieval.

Problem

Activation steering often degrades reasoning and retrieval utility; the paper argues this is driven by steering-induced attention rerouting away from contextually important tokens.

Approach

Identify attention rerouting as the mechanism behind the steering-utility trade-off in query-space steering, then project steering vectors onto the complement of the top key-difference eigenspace (focus vs tail tokens) computed on a utility calibration set, applied only to top-risk heads ranked by a Rayleigh quotient.

Key results

• On LLaMA-3.1-8B-Instruct, SKOP achieves the best average rank (2.69) across steering and utility benchmarks, with IFBench 25.0, ARC 65.0, HellaSwag 65.2, GSM8K 66.6 versus 75.2 baseline and 14.0-60.5 for steering baselines.
• SKOP reduces utility degradation by 5-7x while retaining over 95% of vanilla query-space steering efficacy (per abstract).
• On RULER NIAH at 16K context, SKOP retrieval accuracy is 88.2% versus 72.6% for DISCO-Q and 55.7% for Comm Steer.

Surprising findings

• Full key-invariant projection eliminates steering effects, revealing a fundamental tension between steering efficacy and attention invariance.
• Risk-ranked heads drive stronger steering but also larger utility drops than discriminative heads.

Limitations

• Analysis and method restricted to query-space steering; extension to residual-stream steering left as future work.
• Requires a small utility calibration set and incurs a small loss in steering efficacy relative to vanilla steering.
• Evaluated on only two instruction-tuned models and a limited set of steering behaviours.

Experimental setup

Datasets: GSM8K Alpaca PIQA NarrativeQA TruthfulQA Model-Written Evaluation (power-seeking, wealth-seeking, corrigibility) IFBench ARC-Challenge HellaSwag RULER

Benchmarks: TruthfulQA Model-Written Evaluation power-seeking Model-Written Evaluation wealth-seeking Model-Written Evaluation corrigibility IFBench ARC-Challenge HellaSwag GSM8K RULER NIAH

Metrics: LLM judge score (1-4) True*Info accuracy formatting compliance average rank

Baselines: CAA ITI DISCO-Q Comm Steer Angular Steer CAST SADI LoRA

Builds on: CAA DISCO-Q null-space steering Comm Steer

Closest papers by meaning

• Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs (NeurIPS 2026 · similarity 0.88)
• When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.87)
• DISCO: Disentangled Communication Steering for Large Language Models (NeurIPS 2025 · similarity 0.87)
• Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.86)
• Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads (NeurIPS 2026 · similarity 0.85)
• Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning (NeurIPS 2025 · similarity 0.85)