Related threads

These studies share the project’s methods, but they measure other things. They aren’t evidence about how a time horizon is represented, so they’re kept separate from the main argument. Most come from the SPAR final report and haven’t been separately published.

Apparent planning in generated code

Across 11 models, probes predicted a function’s return type with 80–91% accuracy, which looked like evidence of planning ahead. A classifier using only function and parameter names scored 87–95%, matching or beating the probes. Acrostic tests were also near chance. This is why the “no surface baseline yet” label appears on the temporal probes.

These models go up to 7B parameters, so larger models remain untested.

Behavior under repetition

Repetition affects models differently. In the compiled results, the refusal rate in Qwen3-8B falls from 0.75 to 0.50 over 20 repetitions, while Llama-3.1-8B-Instruct stays at 0.75. Injection resistance improves with repetition in the two models tested. Rates move in steps of 0.125, which implies few prompts per cell.

A draft study looked for probe activation ahead of behavioral change and found no gap across 17 runs. Representation and behavior shifted together from the first repetition.

Sycophancy under social pressure

Social pressure flips answers in 10–62.5% of cases, but true surrender occurs in 1.5–15.5%. The hidden-state probe is underpowered, with 11 positive examples where about 1,000 are needed.

Uncertainty across reasoning steps

Uncertainty encoding is strongest at the first reasoning step and weakens after it. It varies significantly across datasets and only weakly predicts future states.

Output entropy in long contexts

Output entropy collapses as context fills, without a matching gain in accuracy. In-context learning partly offsets this on some task patterns.