Benjamin Han<p>6/</p><p>In a more recent work [5], the authors tested LLMs on *pure* <a href="https://sigmoid.social/tags/causalInference" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>causalInference</span></a> tasks, where all variables are now symbolic (screenshot 1). They constructed systematically a dataset starting by picking variables, to generating all possible <a href="https://sigmoid.social/tags/causalGraphs" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>causalGraphs</span></a>, to finally mapping all possible statistical <a href="https://sigmoid.social/tags/correlations" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>correlations</span></a>. They then “verbalize” these graphs into problems for LLMs to solve for a given causation hypothesis (screenshot 2).</p><p><a href="https://sigmoid.social/tags/Paper" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Paper</span></a> <a href="https://sigmoid.social/tags/NLP" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>NLP</span></a> <a href="https://sigmoid.social/tags/NLProc" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>NLProc</span></a> <a href="https://sigmoid.social/tags/Causation" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Causation</span></a> <a href="https://sigmoid.social/tags/CausalReasoning" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>CausalReasoning</span></a> <a href="https://sigmoid.social/tags/research" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>research</span></a></p>