What is temporal awareness?

Large language models constantly make decisions that play out over time. Answer now, or think longer? Optimize for this turn, or the whole conversation? Take the sure thing today, or the bigger payoff later?

Temporal awareness is how a model internally represents these time horizons — whether it’s leaning short-term or long-term — and whether that internal sense matches what it actually says out loud.

Internal vs. stated horizon

Here’s the key distinction this project is built around.

A model’s stated horizon is what it tells you: “I’d recommend the long-term plan.”
Its internal horizon is what its activations imply as it computes the answer.

These don’t always agree. A model might say something short-term and harmless while its internal state is oriented toward long-horizon planning — or vice versa. That gap is the safety-relevant signal. If we can measure it, we can notice when a model’s words and its internal “intentions” come apart.

Formally, the project grounds this in intertemporal preference: a value function that weighs a reward by how far away in time it is, and an “internal horizon” — the point where the model stops caring about the future. The question is whether that internal horizon matches the one the model states. See the research program.

Why it matters for AI safety

Three reasons this is more than a curiosity:

Agents now run for a long time. As models are deployed in long, autonomous loops — dozens of conversational turns, hundreds of documents — their behavior can drift. The project finds that safety behavior in particular can degrade over repetitive sessions, with some safeguards weakening faster than others.
Words are not enough to monitor. If we only watch what a model says, we miss internal shifts until they surface as bad outputs. Reading the internal time-horizon gives us an earlier signal.
It’s measurable today. A model’s temporal scope turns out to be linearly readable from a single layer of activations. That makes a cheap, real-time probe possible — the kind of concrete, deployable monitor that safety work needs.

This isn’t a lone bet

Reading a safety-relevant property straight from a model’s activations is already a working idea, not a hope. Independent researchers have used simple linear probes to spot hidden “sleeper-agent” backdoors, and to catch a 70-billion-parameter model being strategically deceptive — concealing insider trading, or quietly underperforming on a safety test — nearly perfectly (detection scores up to 0.999). We’re pointing that same well-tested playbook at a new target: a model’s sense of time.

What this project actually does

Detect temporal scope from internal representations (linear probes).
Steer a model’s temporal orientation by nudging its activations.
Locate the circuits that compute temporal decisions (activation patching).
Stress-test all of the above: when do these representations fail?

Early signs, honestly held

Two early findings hint at why this is worth doing — and one keeps us honest:

A model may “know” when it’s unsure. Its internal uncertainty shows up most strongly at the very first step of its reasoning, then fades as it keeps going — exactly when a long-running agent would most need that self-check.
“Planning” can be a mirage. When a model looks like it’s planning ahead, that can be one of two things: genuinely setting up a future step, or just leaving breadcrumbs a probe picks up after the fact. In one code-writing test, a dead-simple shortcut — the function’s name and inputs alone — predicted just as well as reading the model’s internals. So we now run a “baseline staircase” on every probe, to be sure we’re reading real understanding and not a trick.

See the verified numbers on What we score, poke at the geometry on Explore, then pick an issue and join in.