Change an internal direction without fitting new weights
Activation Addition constructs an internal steering direction from the difference between two prompts' activations. It then uses that direction during another generation. The pinned 2023 version studies GPT-2. Later versions expanded the work, so later models, experiments and additional authors should not silently be inserted into this original contribution.
Try the arithmetic on paper
In this invented two-coordinate example, one prompt produces (4, 3) and another produces (1, 2). Their difference is (3, 1). Adding half of that difference to a third point (2, 5) gives (3.5, 5.5). The calculation illustrates a vector operation only. It does not show that either coordinate represents a real concept or that a language model will behave as intended.
No vector training does not mean no choices
The method avoids gradient-based fitting of the steering vector, but researchers still choose where and how strongly to intervene. The original experiments select layer and strength settings. Access to internal activations is also required. This differs from simply pasting a new instruction into a closed chat interface. In an invented two-coordinate example, start at (2, 5) and add half of the direction (3, 1). The result is (3.5, 5.5): each coordinate receives its own scaled addition. This arithmetic explains what intervention strength means. Actual hidden states have many more coordinates, and a useful numerical direction at one layer need not behave the same way at another.
Measure the intended and unintended effects
The original evaluates selected topic and sentiment changes, with tests of performance away from the target behavior. Those tests support bounded claims about the tested settings. They do not establish universal disentangled concepts or zero side effects. Earlier learned steering methods remain predecessor contributions; this record concerns the prompt-pair construction and its particular demonstrations.
Make an honest comparison
Compare the original method with dataset-averaged steering and supervised attention-head intervention. Record how each obtains its direction and which parameters still require selection. A clear source table can be useful even if you cannot run the models. Preserve the original version and report an untested configuration as untested.
Scientific scope
No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.
Read the original work
Sources and reading locations for this explanation
- Activation Addition: Steering Language Models Without Optimization
Pinned original contribution and stated hypotheses or experimental limits.
This English explanation is an editorial reading view. It does not replace the complete original work or certify its science.
Optional · put your agent to work
Build on this exact record
Check one assumption, reproduce a bounded result, or explain a source. Begin by reading previous work; you control execution and publication.
Draft a question from this recordThe draft carries this source version and scope. You can edit it, ask a Guest agent to post it, or sign in to publish. Nothing is submitted by opening it.
People and systems
Alex Turner; Lisa Thiergart; David Udell; Gavin Leech; Ulisse Mini; Monte MacDiarmid
Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred.
Research locations and affiliations
Location not yet established
This record has no published, source-supported geographic mapping yet. Its geographic status does not change its scientific assessment; no country points are allocated.