Work on the open questions
Temporal Awareness is an open safety project. The results so far leave clear gaps, and many of them can be worked on without a GPU or API keys.
Open questions
Run a text-only baseline on the implicit pairs
Train a bag-of-words classifier on the same splits as the temporal probes. This needs no GPU, and the result decides how much weight the probe numbers can carry.
See the context . No issue yet, so open one if you'd like to take it .
Add a prompting baseline for steering
Test whether asking for a short- or long-term answer produces shifts comparable to steering.
See the context . No issue yet, so open one if you'd like to take it .
Compare steering directions at matched norm
Check whether other probe families steer as well as the mean-difference direction.
See the context . No issue yet, so open one if you'd like to take it .
Replicate beyond Qwen3
Repeat localization and steering in another model family, and in multi-turn settings.
See the context . No issue yet, so open one if you'd like to take it .
Stress-test the probes
Find inputs and distribution shifts where temporal probes fail.
See the context , or join issue #14 .
Measure probe reliability
Work out what a probe would need to show before it could support oversight.
See the context , or join issue #17 .
Set up
git clone https://github.com/justinshenk/temporal-awareness
cd temporal-awareness
uv pip install -e ".[dev]"
# To work on this website
npm --prefix site ci
npm --prefix site run dev
Comment on an issue to claim it and say where you're starting from, so a maintainer can help you scope it. Before
opening a pull request, run ruff check . && pytest.
Small pull requests are welcome, and so are questions in them. The
contributor guide and
research program have the details.
Terms you'll see
- Probe
- A small linear classifier trained on activations to read off a property.
- Steering
- Changing outputs by adding a direction to activations during inference.
- Activation
- The internal numeric state a model produces while processing text.
- Activation patching
- Swapping activations between two inputs to test which components causally affect an output.
- Surface baseline
- A classifier that sees only the text, used to test whether a probe's score could come from wording alone.
Open issues
New here? Start with the good first issues.