Research map / Large language models / Language, safety and interpretability
Composable Interventions for Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces a framework and metrics (Order-free Error, Order Sensitivity) plus a unified codebase to study how test-time interventions (knowledge editing, model compression, machine unlearning) compose, running 417 compositions on Llama-3 8B.
Problem
Test-time interventions on language models (editing, compression, unlearning) are typically developed independently, yet practitioners apply several in sequence; there is no standard way to measure how they interact or whether their effects depend on application order.
Approach
Defines composability for LM interventions via two metrics computed over hyperparameter sweeps: Order-free Error (best-case performance across both application orders) and Order Sensitivity (gap between the two orders), and applies them to pairwise compositions of editing, compression, and unlearning methods.
Key results
- Model compression degrades editing performance; GPTQ decays edit performance faster than AWQ, and AWQ wins 12 of the editing-compression comparisons versus GPTQ's 0.
- Order matters substantially: for pruning, compressing before editing is generally better, while for GPTQ editing should come first and for AWQ compression first.
- MMLU fails to separate composability: methods with near-identical MMLU scores differ greatly on edit success order sensitivity (e.g., MEMIT and Wanda near-perfect on MMLU but 0.24 Order Sensitivity on Edit Success).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply compression before editing rather than editing before compression | Knowledge editing with pruning · zsRE | Edit Success | Helped |
| Apply editing before GPTQ quantization | Knowledge editing with GPTQ · zsRE | Edit Success | Helped |
| Apply AWQ quantization before editing | Knowledge editing with AWQ · zsRE | Edit Success | Helped |
| Increase compression level | Knowledge editing methods · zsRE | Edit Success | Hurt |
| Apply unlearning to a compressed model | Machine unlearning methods · WMDP | Unlearning performance | Hurt |
| Use RMU instead of GA for composing editing and unlearning | Knowledge editing and unlearning · zsRE | Edit Success Order Sensitivity: .48 vs .01 | Helped |
| Use RMU instead of GD for composing editing and unlearning | Knowledge editing and unlearning · zsRE | Edit Success Order Sensitivity: .40 vs .01 | Helped |
| Use overlapping rather than different target layers (ablation) | RMU + MEMIT · zsRE | Edit Success Order Sensitivity: .01 -> .00 | No clear effect |
| Add a retain loss term to gradient-ascent unlearning | Gradient Difference · WMDP | Unlearning performance | Helped |
| Increase the retain-term weight in Gradient Difference | Gradient Difference · WMDP | Unlearning performance | Helped |
Surprising findings
- LoRA is surprisingly non-composable as an editor, with high order-free error under compression.
- Popular general-purpose utility metric MMLU is a poor proxy for composability.
Limitations
- Only ten methods and three intervention categories are studied; many popular methods are omitted.
- Results are primarily on Llama-3 8B, and generalization across model scale is unknown.
- WMDP is an imperfect evaluation of unlearning, and GA's poor performance may confound comparisons.
Experimental setup
Datasets: zsRE WMDP MMLU Wikitext
Benchmarks: zsRE WMDP MMLU
Metrics: edit success F1 edit generalization F1 strict edit locality F1 WMDP accuracy MMLU accuracy Order-free Error Order Sensitivity # wins
Baselines: Finetuning LoRA Gradient Ascent Gradient Difference SparseGPT Wanda GPTQ AWQ
Builds on: MEMIT SparseGPT Wanda GPTQ AWQ Representation Misdirection for Unlearning
Closest papers by meaning
- Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting (ICLR 2024 · similarity 0.87)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.86)
- Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation (NeurIPS 2025 · similarity 0.86)
- XIFBench: Evaluating Large Language Models on Multilingual Instruction Following (NeurIPS 2025 · similarity 0.86)
- Aioli: A Unified Optimization Framework for Language Model Data Mixing (ICLR 2025 · similarity 0.86)
- Evaluating the Zero-shot Robustness of Instruction-tuned Language Models (ICLR 2024 · similarity 0.86)
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.