We find out: the agent's policy state is enough to estimate the future action outcome! we'll present #ECCV2026 in Sweden. With our plain #LLM (dubbed DILLO), the agent finds it out 14x faster. Performance improves up to 15pp. This is #SafeAI for #agents arxiv.org/abs/2603.23149
Describe-Then-Act: Proactive Agent Steering via Distilled Language-Action World Models
Deploying safety-critical agents requires anticipating the consequences of actions before they are executed. While world models offer a paradigm for this proactive foresight, current approaches relyin...
arxiv.org
Fabio Galasso@galasso.bsky.social · 5mo ago#Worldmodels are great! --with those, #agents can simulate the outcome of their actions, understand whether they will be successful, and safe. BUT world models are expensive, and not compatible with real-time agent's decision making.
