🔮 Future Oracles

Read what a language model is about to say from a single activation vector. Click a token: each oracle takes the residual-stream vector at that token (one layer, one position) from its own Qwen3 model and writes the next k tokens. It is scored against what the model itself would go on to say from there.

Textbox

Examples

Oracles

Qwen3-0.6B · L19 Qwen3-1.7B · L19 Qwen3-8B · L8–32 (multi-layer) · hard labels Qwen3-8B · L8–32 (multi-layer) · distilled Qwen3-8B · L24

k · tokens to read

1 9

Controls

none: marker embedding untouched · shuffled: another token's vector

none shuffled

Layer (multi-layer oracles)

Click a token above. Each oracle reads the residual-stream vector at that token from its own model and writes the tokens the model is about to produce.

5 oracles loaded · new: Qwen3-0.6B · L19, Qwen3-1.7B · L19, Qwen3-8B · L8–32 (multi-layer) · hard labels, Qwen3-8B · L8–32 (multi-layer) · distilled, Qwen3-8B · L24 · adapters, data and evals: syvb/rl-future-lens-qwen3-8b · API endpoint /probe