AI Paper Atlas

Research map / Large language models / Agents, code and math

LLMs Struggle to Balance Reasoning and World Knowledge in Causal Narrative Understanding

Khurram Yamin, Shantanu Gupta, Gaurav Rohit Ghosal, Zachary Chase Lipton, Bryan Wilder

ICLR 2026 · Large language models › Agents, code and math · approach: Inductive and abstract reasoning evaluation

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
extract explicit causal graph then answer, instead of standard QAGPT-4o · CauseNet semi-synthetic, parametric conflictaccuracy (%): 67.2 -> 98.7Improved
extract explicit causal graph then answer, instead of standard QAGPT-4o · CauseNet real-world, parametric conflictaccuracy (%): 52.1 -> 93.2Improved
add chain-of-thought to standard promptingGPT-4o · CauseNet semi-synthetic, parametric conflictaccuracy (%): 67.2 -> 73.1Improved
add chain-of-thought to standard promptingGPT-4o · CauseNet real-world, parametric conflictaccuracy (%): 52.1 -> 57.6Improved
add chain-of-thought to standard promptingGPT-4o · CauseNet real-world, no conflictaccuracy (%): 90.9 -> 89.2No clear effect
narrative conflicts with parametric knowledge instead of agreeingGPT-4o standard prompting · CauseNet real-worldaccuracy (%): 90.9 -> 52.1Hurt
narrative conflicts with parametric knowledge instead of agreeingGPT-4o standard prompting · CauseNet semi-syntheticaccuracy (%): 99.8 -> 67.2Hurt
narrate events in reverse topological order instead of forwardGPT-4o standard, CoT, in-context · synthetic chainsaccuracy (%): ~50 % betterHurt
provide narrative together with extracted graph instead of graph aloneGPT-4o · synthetic reverse narrativesaccuracy (%)Hurt
complex graphs with forks and colliders instead of simple chainsGPT-4o · synthetic and CauseNet narrativesaccuracy (%)Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy