4M-parameter topic monitor reads Gemma 3 activations, labels every token in real time
CatAstro_Piyush · x · 2026-11-13
MercityAI's first arXiv paper: a topic monitor under 4M parameters reads Gemma 3's activations and labels the topic of every token in real time. Trained only on prefill data, it tagged 90–98% of Gemma's generated tokens with a user-specified topic.
- Existing streaming monitors only flag safe vs harmful; topic guardrails wait for the finished response
- This is the first approach tracking topic token-by-token during generation
- Implication: intervene the moment output drifts off-topic instead of filtering after the fact
More from Research
- Single-constraint KL minimization reduces to 1-D bisection on the dual — caglar_ee · 2026-11-13
- Researcher pushes back on METR's 2x AI acceleration claim, citing 1.4x figure — HaydnBelfield · 2026-11-13
- Mercor launches OWE-Bench: 26 real engineering projects that stump coding agents — ben_burtenshaw · 2026-11-13
- HuggingFaceTB's smol-training-playbook trends on Hugging Face — HuggingFaceTB · 2026-11-13
- Blank Bio releases Ladon, an RNA foundation model for predicting patient outcomes — michaelrzhang · 2026-11-13
- GEM-4D: geometry-enhanced video world models lift robot sim success 22.7% to 75.1% — pliang279 · 2026-11-13