📖 NLP Understanding¶

🔬 ICLR2026 · 2 paper notes

LANE: Label-Aware Noise Elimination for Fine-Grained Text Classification: LANE upgrades the classic margin indicator for identifying mislabeled samples to a Label-aware Margin. For negative margins, the penalty is reduced if the mislabeled category is semantically similar to the model's prediction (e.g., "Anger" labeled as "Fear") and increased if they are semantically distant (e.g., "Trust" labeled as "Fear"). Based on this, it applies dynamic weighting to each sample rather than hard deletion, consistently outperforming strong baselines like AUM and HMW across 10 text classification datasets.
What's the Plan? Metrics for Implicit Planning in LLMs and Their Application to Rhyme Generation and Question Answering: This paper proposes the mean activation difference steering method and accompanying quantitative metrics. Through two case studies—rhyme generation and question answering—it systematically proves across 23 open models (1B-32B) that representations of target tokens (rhymes/answers) are formed at early sequence positions (forward planning) and causally influence the generation of intermediate tokens (backward planning). Implicit planning emerges in models as small as 1B, suggesting it is a universal mechanism rather than exclusive to large-scale models.