Steering Instruction Hierarchies at Inference Time
Quick Answer
V-Steer is a novel inference method that enhances instruction hierarchy adherence in LLMs, achieving up to 92% accuracy on role conflict benchmarks across models ranging from 7B to 70B parameters.
Quick Take
By modifying cached value vectors, it significantly outperforms prompt-only baselines and matches state-of-the-art training methods with minimal overhead.
Key Points
- V-Steer boosts privileged inputs in by editing cached value vectors.
- Achieves 92% accuracy on controlled role conflict benchmarks.
- Outperforms prompt-only baselines and rivals state-of-the-art training methods.
- Compatible with fused attention backends, adding minimal prefill overhead.
- Applicable across models from 7B to 70B parameters.
DeepSignal Analysis
What happened
V-Steer is a new method designed to enhance adherence to instruction hierarchies in large language models (LLMs) during inference. It modifies cached value vectors to improve the accuracy of role conflict benchmarks, achieving up to 92% accuracy across models with parameters ranging from 7B to 70B.
Key evidence
- V-Steer modifies cached value vectors at prompt positions to restore the influence of higher priority inputs in LLMs.
- The method significantly improves primary constraint accuracy from under 18% to 92% on controlled role conflict benchmarks.
- V-Steer outperforms prompt-only baselines and matches or exceeds state-of-the-art training methods across three of four scales of LLMs.
Why it matters
The ability to maintain instruction hierarchies is crucial for the safe deployment of LLMs, as it ensures that higher priority instructions override conflicting lower priority ones. V-Steer addresses a common failure in current models, potentially enhancing their reliability in real-world applications. This method's training-free nature and minimal overhead could facilitate broader adoption in various LLM applications.
What to watch
Paper Resources
Source Excerpt
Instruction hierarchies are a core safety assumption of language model deployment: higher priority inputs, such as system prompts, should override conflicting lower priority inputs from users or tools. Yet frontier often violate this hierarchy. We introduce V-Steer, a training-free inference time method that restores privileged influence by editing cached value vectors at prompt positions. Using direct logit attribution on the first next token prediction, V-Steer identifies heads where lowe
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.