📖 NLP Understanding

🧪 ICML2026 · 2 paper notes

📌 Same area in other venues: 💬 ACL2026 (33) · 🤖 AAAI2026 (2) · 🧠 NeurIPS2025 (2) · 📹 ICCV2025 (1)

Causal Fine-Tuning under Latent Confounded Shift

This paper proposes Causal Fine-Tuning (CFT): embedding an SCM-inspired decomposition of "high-level stable features \(C\) + low-level confounder-sensitive features \(\Phi\)" into standard BERT fine-tuning. Predictions are made via a front-door style do-calculus adjustment formula, significantly outperforming single-domain generalization baselines like SFT, SWA, and WISE under spurious correlation injection attacks in text.

Controlling the Risk of Corrupted Contexts for Language Models via Early-Exiting

This paper formalizes the issue of "performance degradation in LLMs caused by user-provided corrupted contexts" as a risk control problem. By using zero-shot performance as a "safety baseline" and combining dynamic early-exiting (exiting at intermediate layers to avoid "overthinking" harmful contexts) with a context-aware loss and an improved Learn-then-Test framework (preserving negative loss values via risk transformation rather than clipping), this method ensures risk \(\leq\) user-specified \(\epsilon\) across 9 tasks while achieving over 50% computational speedup.