Research map / Large language models / Language, safety and interpretability
Controlling the Risk of Corrupted Contexts for Language Models via Early-Exiting
Andrea Wynn, Metod Jazbec, Charith Peris, Rinat Khaziev, Anqi Liu, Daniel Khashabi, Eric Nalisnick
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| early-exit at an intermediate layer instead of using the full model under harmful context | Llama-style LLMs with ICL · ICL classification tasks with incorrect demonstrations | accuracy | Improved |
| early exit with a suitable threshold also on helpful context | early-exit LLM · ICL tasks with correct demonstrations | accuracy | Mixed |
| replace loss clipping with domain-preserving risk transformation in LTT (ablation) | risk-controlled early exit · 8 ICL tasks (epsilon 0.05) | layers evaluated: 53% fewer layers on average | Improved |
| replace loss clipping with domain-preserving risk transformation in LTT (ablation) | risk-controlled early exit · SQuAD v2.0 | layers evaluated: up to 81.8% fewer layers | Improved |
| replace loss clipping with risk transformation (ablation) | risk-controlled early exit · ICL tasks and SQuAD v2.0 | risk control validity | No clear effect |
| add zero-shot fallback when no exit is confident | early-exit model · SQuAD v2.0 | fallback rate on misleading vs answerable questions: 3.0% vs 0.9% | Improved |
| fall back to zero-shot on unsafe context rather than full-model prediction (as CALM does) | early-exit model · SQuAD v2.0 unanswerable | question-level correctness: zero-shot better on 71% of questions | Improved |
| restrict early exit to the second half of layers (ablation) | early-exit design · all tasks and models | safety/accuracy | Improved |
Closest papers by meaning
- Learning and Forgetting Unsafe Examples in Large Language Models (ICML 2024 · similarity 0.89)
- Prompt Risk Control: A Rigorous Framework for Responsible Deployment of Large Language Models (ICLR 2024 · similarity 0.89)
- CASE-Bench: Context-Aware SafEty Benchmark for Large Language Models (ICML 2025 · similarity 0.88)
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.88)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.88)
- UpSafe℃: Upcycling for Controllable Safety in Large Language Models (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
