AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

HInT: Hypergraph Infusion at the Structural Layers Improves Table Understanding

Wonjin Lee, Soomi Jeong, Kwang In Kim

ICML 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs

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
add hypergraph infusion at structural layersQwen2.5-3B (SFT) · WikiTQEM (%): 65.44 -> 67.39Improved
add hypergraph infusion at structural layersQwen2.5-3B (SFT) · HiTabEM (%): 70.12 -> 71.75Improved
add hypergraph infusion at structural layersQwen2.5-3B (SFT) · TabFactaccuracy (%): 85.25 -> 86.71Improved
add hypergraph infusion at structural layersQwen2.5-3B (SFT) · HybridQAEM (%): 39.92 -> 40.99Improved
add hypergraph infusion at structural layersQwen2.5-3B (SFT) · FeTaQABLEU (%): 26.11 -> 26.58Improved
add external (hyper)graph encoders (HeGTa, LLaSA) to a decoder-only LLMQwen2.5-3B (SFT) · WikiTQEM (%): 65.44 -> 62.52Hurt
add structural-head knowledge: remove top-40 structural heads vs random heads (ablation)Qwen2.5-3B · WikiTQEM (%): 47.51 -> 12.73 (structural) vs 32.29 (random)Improved
add three extra transformer layers (capacity-matched baseline) (ablation)Qwen2.5 (SFT) · WikiTQEM (%): 65.44 -> 64.43Hurt
inject structure at identified structural layers instead of random layers (ablation)HInT · WikiTQEM (%): 65.71 -> 67.39Improved
inject structure at identified structural layers instead of all layers (ablation)HInT · WikiTQEM (%): 66.37 -> 67.39Improved
increase structural layers from 0 to 12 (ablation)HInT · HiTabEM (%): 70.10 -> 71.84Improved
increase structural layers from 12 to 20 (ablation)HInT · HiTabEM (%): 71.84 -> 71.21Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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