Research map / Large language models / Instruction and fine-tuning
A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMs
Trenton Chang, Tobias Schnabel, Adith Swaminathan, Jenna Wiens
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 |
|---|---|---|---|
| scale model size from 8B to 70B | Llama3.1 · steerability probe | miscalibration: 0.667 -> 0.455 | Improved |
| larger and newer models | Llama3 family · steerability probe | orthogonality (side effects): 0.718 best (Llama-3.3-70B) | No clear effect |
| request anti-correlated rather than correlated changes | Llama3.3-70B · steerability probe | steering error: 0.404 -> 0.535 | Hurt |
| remove negative prompt and add specific instructions (ablation) | Llama3.1-8B · steerability probe | miscalibration: 0.667 -> 0.444 | Improved |
| prompt engineering variants (ablation) | Llama3.1-8B · steerability probe | orthogonality | No clear effect |
| use best-of-4 sampling (ablation) | Llama3.1-8B · steerability probe | steering error: 0.429 -> 0.404 | Improved |
| double N in best-of-N (ablation) | Llama3.1-8B · steerability probe | steering error: at most 0.031 per doubling | Improved |
| RL fine-tune with steering-error reward | Llama3.1-8B · 2D steerability probe | steering error: 0.300 -> 0.119 | Improved |
| RL fine-tune with steering-error reward | Llama3.1-8B · 2D steerability probe | orthogonality: 0.147 -> 0.160 | Hurt |
| RL reward on miscalibration only instead of full steering error (ablation) | Llama3.1-8B · 2D steerability probe | orthogonality: 0.160 -> 0.366 | Hurt |
| RL reward on orthogonality only instead of full steering error (ablation) | Llama3.1-8B · 2D steerability probe | miscalibration: 0.294 -> 1.463 | Hurt |
| RL fine-tune on 1D instructions | Llama3.1-8B · 2D anti-correlated requests | orthogonality gap vs correlated: 0.114 -> 0.005 | Improved |
Closest papers by meaning
- SteerConf: Steering LLMs for Confidence Elicitation (NeurIPS 2025 · similarity 0.90)
- Evaluating Large Language Models at Evaluating Instruction Following (ICLR 2024 · similarity 0.89)
- MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback (ICLR 2024 · similarity 0.89)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.88)
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs (ICLR 2026 · similarity 0.88)
- EvaLearn: Quantifying the Learning Capability and Efficiency of LLMs via Sequential Problem Solving (NeurIPS 2025 · 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.
