Tag
Microsoft's M365 Research introduces FOCUS, a training-free test-time method that compresses LLM agent interaction histories by retaining only the units that causally shape future decisions, cutting peak context by up to 48% and boosting task success by up to 8.9 points over full-history execution.
A researcher shares an experimental plan for identifying causal dependencies between capability dimensions in a 31B model using contrastive targeted SFT and circuit tracing, seeking feedback on methodology and related work.