WiC is Not WSD: LLMs err by overthinking fine-grained word senses, AACL 2026 paper finds
Bollegala · x · 2026-11-08
A paper accepted to AACL 2026 argues that LLMs' struggles with Word-in-Context (WiC) stem not just from comparing contextual word uses, but from the lack of an explicit sense inventory defining the right granularity.
- Evaluating open LLMs on WiC vs traditional WSD, providing candidate senses improves WiC performance in all settings
- Human evaluation shows many apparent errors reflect label ambiguity or mismatches between model and annotator sense boundaries, not comprehension failures
- Key finding: LLMs tend to overthink sense distinctions, making errors based on overly fine-grained splits
More from Research
- MuJoCo 4.0 ships with speculative contacts, permanent soft-body deformation and MuJoCo Studio — ZoubinGhahrama1 · 2026-11-08
- mitsuhiko: djs' JS "heap dumps" may beat real dumps for durable coding agents — mitsuhiko · 2026-11-08
- Salesforce paper: RL-trained experts merged via OPD save 42% FLOPs vs external-teacher distillation — rohanpaul_ai · 2026-11-08
- Albert Gu: Bitter Lesson isn't 'just scale Transformers' — algorithmic innovation matters — sewon__min · 2026-11-08
- Autoware open-sources an off-road driving simulator built on NVIDIA Isaac Sim — rsasaki0109 · 2026-11-08
- The meta-question as agent swarms scale: can models beat gradient descent at its own game? — JensHonack · 2026-11-08