---
title: "Creating Steering Vectors via Input... | SI Paper Desk"
url: https://thesimind.si/papers/2611.17481/
description: "A steering vector's effect on a language model can often be copied by a text suffix added to the prompt, so no access to model internals is needed when..."
---

Number 10 of 12 · November 16, 2026

# Creating Steering Vectors via Input Optimisation

A steering vector's effect on a language model can often be copied by a text suffix added to the prompt, so no access to model internals is needed when generating.

Batuhan Akar · Submitted November 12, 2026 · arXiv 2611.17481v1 · cs.CL

[Read the paper on arXiv](https://arxiv.org/abs/2611.17481) · [The PDF](https://arxiv.org/pdf/2611.17481) · [The day's 12](https://thesimind.si/papers/day/2026-11-16/)

## What they found

The author introduces ActGCG, a change to the Greedy Coordinate Gradient method. It searches for a suffix whose effect on the model's internal activations matches the effect of adding a steering vector. The author also adds universal suffixes trained across many prompts, a step that rewrites the suffix into human-readable text midway through the search, and a soft-suffix version that drops the limit to real tokens. For steering a model to answer in German, a discrete suffix reproduces the steering vector's behaviour almost perfectly and transfers well across prompts and models. For anti-refusal steering (jailbreaking), the match is weaker. It works under a phrase-avoidance metric but only partly under the stricter HarmBench classifier.

## Why it matters

Activation steering is a way to change how a model behaves by adding a vector to its hidden state. It normally needs white-box access, meaning you must be able to edit the model's insides at every step. This work suggests that some of the same control can be reached through the input alone. The abstract says how well this works depends on the concept being steered.

## How it changes the race

This is a bachelor's thesis result, so treat it as an early step. If it holds, steering effects could be moved into plain prompts, which would work with models where internal access is not available. The abstract does not say how this compares with other methods.

## Safety and control

The paper links steering to jailbreaking: a suffix can partly reproduce an anti-refusal steering vector without any access to model internals. That suggests safety training could be attacked through the input alone. The match was only partial under the HarmBench classifier.

### The abstract, by the authors

Activation steering influences a language model's behaviour by adding a steering vector to its residual stream during generation, but it requires white-box access to modify the hidden states at every generation step. We introduce ActGCG, a modification of Greedy Coordinate Gradient which optimises a suffix to match the target activations obtained by adding a steering vector to the residual stream. We further introduce universal suffixes trained across many prompts at once, a model-translation step that rewrites the suffix into human-readable text midway through optimisation, and a soft-suffix variant to remove the discrete token space constraint. We study two behaviours: steering a model to answer in German, and anti-refusal steering for jailbreaking. For German, a discrete suffix reproduces the steering vector's behaviour almost perfectly and transfers well across prompts and models. For anti-refusal, the equivalence is weaker, succeeding under a phrase-avoidance metric but only partially under the stricter HarmBench classifier. Our results show that a steering vector's behaviour is reachable from the input, with no access to model internals at generation time, up to a loss that depends on the concept being steered.

Quoted from arXiv:2611.17481. The research and its words belong to its authors.

Desk score **70** out of 100

Why it was picked

- Interpretability (in the title)
- Safety
- Top group: ETH Zurich
- Language models
- Cross-listed in 2 fields

Topics

- [Interpretability](https://thesimind.si/papers/?tag=interpretability)
- [Safety](https://thesimind.si/papers/?tag=safety)
- [Security](https://thesimind.si/papers/?tag=security)
- [Language models](https://thesimind.si/papers/?tag=language-models)
- [Alignment](https://thesimind.si/papers/?tag=alignment)

**arXiv ID:** 2611.17481v1
**Submitted:** November 12, 2026
**Fields:** cs.CL (Computation and Language), cs.LG (Machine Learning)
**Level:** Intermediate
**Comments:** Bachelor's Thesis, ETH Zurich

Fingerprint, SHA-256

`bd80d97c0162c88f61f26de2e94ea28135455e3217c32ef5a88c7765e9f43e73`

On Solana: [Proof of Mind seal, 3mXUMW...GSim](https://solscan.io/tx/3mXUMW5MkdFiyPBNPYd5DfwHhma5EdjhUFEw9jZaDuuPqgszJtoDYttSrD9kyWUFdMCzn7RnFNd4nQhwAUgrGSim)

### The exact text fingerprinted

{"arxiv":"2611.17481","version":1,"title":"Creating Steering Vectors via Input Optimisation","authors":["Batuhan Akar"],"day":"2026-11-16","rank":10,"score":70,"status":"written","summary":{"takeaway":"A steering vector's effect on a language model can often be copied by a text suffix added to the prompt, so no access to model internals is needed when generating.","found":"The author introduces ActGCG, a change to the Greedy Coordinate Gradient method. It searches for a suffix whose effect on the model's internal activations matches the effect of adding a steering vector. The author also adds universal suffixes trained across many prompts, a step that rewrites the suffix into human-readable text midway through the search, and a soft-suffix version that drops the limit to real tokens. For steering a model to answer in German, a discrete suffix reproduces the steering vector's behaviour almost perfectly and transfers well across prompts and models. For anti-refusal steering (jailbreaking), the match is weaker. It works under a phrase-avoidance metric but only partly under the stricter HarmBench classifier.","matters":"Activation steering is a way to change how a model behaves by adding a vector to its hidden state. It normally needs white-box access, meaning you must be able to edit the model's insides at every step. This work suggests that some of the same control can be reached through the input alone. The abstract says how well this works depends on the concept being steered.","race":"This is a bachelor's thesis result, so treat it as an early step. If it holds, steering effects could be moved into plain prompts, which would work with models where internal access is not available. The abstract does not say how this compares with other methods.","safety":"The paper links steering to jailbreaking: a suffix can partly reproduce an anti-refusal steering vector without any access to model internals. That suggests safety training could be attacked through the input alone. The match was only partial under the HarmBench classifier.","level":"intermediate","tags":["interpretability","safety","security","language-models","alignment"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Batuhan Akar

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 12, 2026 · Number 8TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual ModulationTERM is a defense that runs while a model answers. It watches how the model's attention shifts from its own starting point and scales down the model's internal updates when something looks off.SecuritySafetyAlignment](https://thesimind.si/papers/2611.13577/) · [Nov 10, 2026 · Paper of the DayMETA-IR: Meta-Introspective Reasoning for Adaptive Safety Alignment in LLMMETA-IR trains language models to check their own answers for risk and to adapt that checking to new jailbreak tricks, and the authors report better robustness than all baselines.SafetyAlignmentSecurity](https://thesimind.si/papers/2611.10089/) · [Nov 14, 2026 · Paper of the DayData-Efficient LLM Safety Alignment Guided by Interpretable FeaturesA small, carefully chosen set of 2.4K training examples, picked using interpretable features inside a model, beat much larger safety datasets on safety and over-refusal tests.AlignmentInterpretabilitySafety](https://thesimind.si/papers/2611.16183/)

[Number 9](https://thesimind.si/papers/2611.17395/) · [All 12 for Nov 16, 2026](https://thesimind.si/papers/day/2026-11-16/) · [Number 11](https://thesimind.si/papers/2611.17461/)
