📖 NLP Understanding¶

🧪 ICML2026 · 2 paper notes

Controlling the Risk of Corrupted Contexts for Language Models via Early-Exiting: This paper formalizes the problem where "user-provided corrupted contexts degrade LLM performance" as a risk control task. By using zero-shot performance as a "safety baseline," combining dynamic early-exit (predicting at intermediate layers to avoid late-layer overthinking of 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 guarantees risk \(\leq\) user-specified \(\epsilon\) while achieving \(> 50\%\) computational acceleration across 9 tasks.
Causal Fine-Tuning under Latent Confounded Shift: This paper proposes Causal Fine-Tuning (CFT): an SCM-inspired decomposition of "high-level stable features \(C\) + low-level confounding-sensitive features \(\Phi\)" is embedded into standard BERT fine-tuning. By utilizing a front-door style do-calculus adjustment for prediction, it significantly outperforms single-domain generalization baselines such as SFT/SWA/WISE under text spurious correlation injection attacks.