Two AACL-IJCNLP papers probe how fine-tuning reshapes LLM internals and why WiC stays hard
Bollegala · x · 2026-09-11
A Liverpool NLP team announced two papers and a tutorial accepted at AACL-IJCNLP:
- Fine-tuning internals: Causally important components for a task concentrate in specific layers, largely uncorrelated with the layers undergoing the biggest representational changes during fine-tuning.
- WiC is Not WSD: LLMs struggle on Word-in-Context mainly because no explicit sense inventory fixes semantic granularity; providing candidate senses improves performance across settings, and human evaluation shows many "errors" are over-fine-grained sense distinctions.
- The team will also give a tutorial on Diffusion Language Models.
More from Research
- Fruit flies trained to write: neural signals turn insect legs into a font generator — dejavucoder · 2026-09-11
- Stokes' theorem formalized in Lean 4 with true pullbacks and d²=0 proof — basedjensen · 2026-09-11
- Missing Benchmark for Agent Runtimes, Not Just Models, Reddit Thread Argues — Balance- · 2026-09-11
- PISA data preprocessed for AI agents: open dataset with AGENTS.md for instant analysis — pelayoarbues · 2026-09-11
- X-AuT: Progressive Audio-Encoder Compression for Speech LLMs via Cross-Scale Distillation — XPENG-AI · 2026-09-11
- Beihang's UniH3 Unifies Hierarchical Homogeneity and Heterogeneity for Medical Image Restoration — Beihang · 2026-09-11