Swap paragraphs for words and Qwen denies injection — yet still guesses the word better
Sauers_ · x · 2026-11-19
Boundary-condition findings from Sauers' thread: when the injection target switches from paragraphs to single words, Qwen no longer wants to report that an injection occurred — yet still gets better at guessing the injected word.
Paragraphs work better than words, possibly because paragraph boundaries act as natural breakpoints; instructing focus before the paragraphs weakens the effect. Notably, attention lands on the word itself, not the label, while the prefill requires answering about the label — so it's not simply "mindlessly raising the logprobs of high-attention words."
More from Research
- Google's GUARD paper: frozen base model, self-editing agent harness gains +8.89 points — omarsar0 · 2026-11-19
- Kezek splits mono two-person conversations into per-speaker tracks for full-duplex dialogue training data — manifestro · 2026-11-19
- Elie Bakouch: current AI research benchmarks overfit and can't track RSI — eliebakouch · 2026-11-19
- Hutter Prize text compression jumps ~10% by overfitting small transformers for arithmetic coding — ricklamers · 2026-11-19
- HybridApexW optimizer: 2x faster than torch.compile AdamW+Muon baseline — omendUUde · 2026-11-19
- Sandia Labs' VASt model predicts high-entropy alloy properties from tens of samples — CurieuxExplorer · 2026-11-19