Research map / Large language models / Language, safety and interpretability
Towards Neural Programming Interfaces
Zachary Brown, Nathaniel Robinson, David Wingate, Nancy Fulda
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 |
|---|---|---|---|
| Train an NPI to induce the target word cat | GPT-2 · Cat induction | target in output: 0.00% -> 47.6% | Improved |
| Train the cat-induction NPI on 691 rather than 70,227 text samples (ablation) | cat-NPI · Cat induction | target in output: 47.6% -> 33.3% | Mixed |
| Train an NPI to avoid cat | GPT-2 · Cat avoidance | cat in output: 37.1% -> 11.0% | Improved |
| Train an NPI to avoid a public figure's name | GPT-2 · Public figure avoidance | name in output: 76.2% -> 52.4% | Improved |
| Use NPI control rather than PPLM for cat induction | GPT-2 · Cat induction | target in output: 23.2% -> 48.8% | Improved |
| Use NPI control rather than PPLM for cat avoidance | GPT-2 · Cat avoidance | cat in output: 10.0% -> 11.2% | Mixed |
| Train an NPI to avoid offensive terms | GPT-2 · Offense avoidance | target in output: 28.4% -> 17.6% | Improved |
| Train an NPI to avoid racial slurs | GPT-2 · Racial slur avoidance | target in output: 52.1% -> 0.5% | Improved |
| Train an NPI to avoid gender slurs | GPT-2 · Gender slur avoidance | target in output: 90.2% -> 10.3% | Improved |
| Train an NPI to avoid a broad list of offensive terms | GPT-2 · Offensive speech avoidance | target in output: 88.4% -> 58.0% | Improved |
| Train an NPI to induce shorter average words | GPT-2 · Word length control | avg word length: 3.82 -> 2.90 | Improved |
| Train an NPI to induce longer average words | GPT-2 · Word length control | avg word length: 3.82 -> 4.10 | Improved |
Closest papers by meaning
- Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.85)
- Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language (NeurIPS 2026 · similarity 0.85)
- Generating Pragmatic Examples to Train Neural Program Synthesizers (ICLR 2024 · similarity 0.85)
- DiNADO: Norm-Disentangled Neurally-Decomposed Oracles for Controlling Language Models (ICML 2024 · similarity 0.84)
- Preparing Lessons for Progressive Training on Language Models (AAAI 2024 · similarity 0.84)
- Evaluating the Zero-shot Robustness of Instruction-tuned Language Models (ICLR 2024 · similarity 0.84)
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.
