It scores causal mechanism reconstruction rather than final outputs.
Many users praise and thank contributors for CausaLab's tests of LLM agents on causal discovery, while one criticizes the models as flawed causal parrots that output wrong graphs despite accuracy scores.
Fun things we kept seeing:
right answer, wrong mechanism (great prediction, wrong causal graph) give the agent the perfect experiments and its structure recovery gets worse — discovery is in choosing them, not the data it stops experimenting way too early
Paper link: https://arxiv.org/abs/2605.26029
In spirit, this echoes the principle of interactive evaluation my friends recently worked on :) @keyang_xuan @p_song1 @lupantech @pengrui_han
@dylan_works_ Pretty cool! Congrats
Takeaway: prediction accuracy hides whether a model really understands. The principle for anyone building causal/scientific agents — score the mechanism, not just the answer; let it run its own experiments; demand transfer; make it verify before committing.
Importantly, many thanks to the inputs from causal experts which helped shape our work @XiangchenSong @chenyuen0103!
Most importantly, shout out for my faithful coauthor @junlin45300
@dylan_works_
@zhengyaojiang Thanks for your interest!
@dylan_works_ Science is sound, vibration, elements opening, matter. Its all here
@dylan_works_ @keyang_xuan @p_song1 @lupantech 🥳🥳
@dylan_works_ 92% accuracy, 0.47 F1. we're building causal parrots. right number, wrong graph, and they quit halfway through the budget. textbook.