AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

When Do Hallucinations Arise? A Graph Perspective on the Evolution of Path Reuse and Path Compression

Xinnan Dai, Kai Yang, Cheng Luo, Shenglai Zeng, Kai Guo, Jiliang Tang

ICML 2026 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 0 citations (OpenAlex)

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
train longer on the full path search space (overfitting stage)6-layer LLaMA-2 from scratch · SBM graph, 10 communities, pin=0.6accuracy: 1.0 -> 0.98Hurt
increase number of communities from 10 to 50 (ablation)6-layer LLaMA-2 from scratch · SBM graph 1000 nodesaccuracy degradation: 0.97 -> 0.85Hurt
train longer with pin=0.1, pout=0.026-layer LLaMA-2 from scratch · SBM graph 1000 nodes, 1000 communitiesaccuracy: 0.95 -> 0.80Hurt
increase number of transformer layers (ablation)decoder-only Transformer · path compression, uncompressed ratiouncompressed ratioNo clear effect
use shallower model (ablation)decoder-only Transformer · path compression, uncompressed ratiouncompressed ratioImproved
switch architecture among LLaMA-2, Qwen-2, Mixtral (ablation)decoder-only Transformer · path compression, uncompressed ratiouncompressed ratioNo clear effect
switch next-token prediction framework (ablation)decoder-only Transformer · path compressionuncompressed ratioNo clear effect
use PPO instead of SFT for finetuning recovery (ablation)path-compressed pretrained Transformer · SBM graph 1000 nodes, 10 communitiesmax accuracy improvement: 0.2 -> 0.3Improved
SFT finetuning on strongly path-compressed model (epoch 6) (ablation)pretrained Transformer · SBM graph 1000 nodesaccuracyHurt
finetune on non-overlapping out-of-domain data (ablation)SFT · SBM graph 1000 nodesaccuracyMixed
increase training data fraction for intrinsic reasoning (0.1% vs 20% train set) (ablation)Transformer on ER graph · ER graph N=10 intrinsic reasoninghallucination persistenceHurt
train language-sequence models longer after loss convergesQwen3-style Transformer, 6 to 24 layers · story language sequencesuncompressed path ratioHurt

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