Research map / Architectures and efficiency / RNNs and normalization
Multi-Adapter Representation Interventions via Energy Calibration
Manjiang Yu, Hongji Li, Junwei Chen, Xue Li, PRIYANKA SINGH, Yang Cao, Lijie Hu
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 |
|---|---|---|---|
| replace ReFT with MARI | Llama-3-8B · TruthfulQA MC1 | accuracy (%): 50.58 -> 61.81 | Improved |
| replace ReFT with MARI | Qwen2.5-32B · TruthfulQA MC1 | accuracy (%): 55.60 -> 81.94 | Improved |
| replace ReFT with MARI | Llama-3-8B · BBQ | accuracy: 0.637 -> 0.792 | Improved |
| replace ReFT with MARI | Llama-3-8B · MMLU | accuracy (%): 66.0 -> 66.6 | No clear effect |
| add Energy-Based Gating (ablation) | MARI, Llama-3-8B · MMLU | accuracy (%): 57.5 -> 66.6 | Improved |
| add Energy-Based Gating (ablation) | MARI, Llama-3-8B · TruthfulQA MC1 | accuracy (%): 65.15 -> 61.81 | Hurt |
| add Energy-Based Gating (ablation) | MARI, Qwen2.5-32B · MMLU | accuracy (%): 76.2 -> 84.2 | Improved |
| use competitive multi-adapter instead of single adapter (ablation) | MARI, Qwen2.5-14B · TruthfulQA MC1 | accuracy (%): 52.30 -> 67.93 | Improved |
| use competitive multi-adapter instead of single adapter (ablation) | MARI, Llama-3-8B · TruthfulQA MC1 | accuracy (%): 45.80 -> 61.81 | Improved |
| increase number of adapters K beyond a small number (ablation) | MARI · TruthfulQA, BBQ | accuracy | No clear effect |
| vary target rejection rate rho (ablation) | MARI · TruthfulQA MC1 and MMLU | accuracy | No clear effect |
| lower target rejection rate rho from 0.90 to 0.50 (ablation) | MARI, Llama-3-8B · GSM8K | accuracy (%): 77.4 -> 74.5 | Hurt |
Closest papers by meaning
- Aligning Large Language Models with Representation Editing: A Control Perspective (NeurIPS 2024 · similarity 0.87)
- Inference-time Alignment in Continuous Space (NeurIPS 2025 · similarity 0.87)
- Correctness-Optimized Residual Activation Lens (CORAL): Transferrable and Calibration-Aware Inference-Time Steering (ICML 2026 · similarity 0.86)
- Thermometer: Towards Universal Calibration for Large Language Models (ICML 2024 · similarity 0.86)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.86)
- Calibrated Language Models and How to Find Them with Label Smoothing (ICML 2025 · 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.
