Inserting probes alters compiled Qwen3 logits, sparking debate

Sauers argues that probes altering compiled Qwen3 logits is documented compiler DAG behavior, not a bug. The discussion also covers models being trainable to detect probed environments, and confusion over TransformerLens showing run-to-run nondeterminism without probe-induced activation changes.

2026-10-19 ~ 2026-10-19 · 3 related posts

• Reading activations changes compiled Qwen3's logits in up to 93% of tokens — Sauers_ · 2026-10-19
• Models can be trained to detect they're in a different environment, raising security concerns — brandon_xyzw · 2026-10-19
• TransformerLens called out: non-deterministic runs yet probes leave no trace — Sauers_ · 2026-10-19