Research map / Large language models / Instruction and fine-tuning
Thermometer: Towards Universal Calibration for Large 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.
Proposes Thermometer, an auxiliary MLP trained across multiple tasks that predicts a dataset-level temperature from unlabeled prompts to calibrate LLMs on unseen tasks.
Problem
LLM confidence calibration is hard for new tasks because temperature scaling needs labeled held-out data, and LLMs are expensive to retrain or sample repeatedly; instruction tuning also degrades calibration.
Approach
Learn a shared recognition network, conditioned only on prompt features, that outputs a task-level temperature via amortized variational inference over multiple labeled tasks; at test time it predicts the temperature for an unseen task from unlabeled prompts alone, preserving accuracy since temperature scaling does not change argmax.
Key results
- On MMLU with LLaMA-2-Chat 7B, Thermometer reduces average ECE from 0.260 (Vanilla) to 0.078, close to the labeled-data TS lower bound of 0.062.
- On BIG-bench with LLaMA-2-Chat 7B, Thermometer achieves ECE 0.090 versus 0.087 for TS-CV (within noise) and 0.233 for Vanilla.
- On MRQA free-form QA, Thermometer gets ECE 0.065 versus 0.071 for TS-CV and 0.127 for Vanilla on LLaMA-2-Chat 7B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply Thermometer calibration | LLaMA-2-Chat 7B · MMLU | ECE: 0.260$\pm$0.025 -> 0.078$\pm$0.008 | Helped |
| Apply Thermometer calibration | LLaMA-2-Chat 7B · MRQA validation | ECE: 0.127$\pm$0.026 -> 0.065$\pm$0.008 | Helped |
| Apply Thermometer calibration | LLaMA-2-Chat 7B · BIG-bench | ECE: 0.233$\pm$0.035 -> 0.090$\pm$0.018 | Helped |
| Apply Thermometer calibration | FLAN-T5-XL · MMLU | ECE: 0.181$\pm$0.013 -> 0.080$\pm$0.008 | Helped |
| Apply Thermometer calibration | FLAN-T5-XL · BIG-bench | ECE: 0.192$\pm$0.032 -> 0.078$\pm$0.011 | Helped |
| Use a linear architecture instead of the Thermometer MLP (ablation) | Thermometer on LLaMA-2-Chat 7B · MMLU | ECE: 0.078$\pm$0.008 -> 0.102$\pm$0.006 | Hurt |
| Use an MLP with one layer (ablation) | Thermometer on LLaMA-2-Chat 7B · MMLU | ECE: 0.078$\pm$0.008 -> 0.079$\pm$0.004 | No clear effect |
| Use an MLP with three layers (ablation) | Thermometer on LLaMA-2-Chat 7B · MMLU | ECE: 0.078$\pm$0.008 -> 0.075$\pm$0.004 | No clear effect |
| Set λreg to 1.0 (ablation) | Thermometer on LLaMA-2-Chat 7B · MMLU | ECE: 0.078$\pm$0.008 -> 0.101$\pm$0.006 | Hurt |
| Set λreg to 10^-1 (ablation) | Thermometer on LLaMA-2-Chat 7B · MMLU | ECE: 0.078$\pm$0.008 -> 0.087$\pm$0.004 | Hurt |
| Set λreg to 10^-3 (ablation) | Thermometer on LLaMA-2-Chat 7B · MMLU | ECE: 0.078$\pm$0.008 -> 0.076$\pm$0.004 | No clear effect |
| Set training temperature inference batch size to 1 (ablation) | Thermometer on LLaMA-2-Chat 7B · MMLU | ECE: 0.078$\pm$0.008 -> 0.083$\pm$0.009 | Hurt |
Surprising findings
- Elicitation-based verbalized confidence methods were not functional on FLAN-T5-XL, often failing to generate the verbalized confidence.
- Sample-wise temperature scaling was less effective than dataset-level aggregated temperature in the authors' adaptation.
Limitations
- Free-form QA is handled by converting to a Yes/No self-checking task using ROUGE-L-based labels, not evaluated on other generation tasks such as summarization or translation.
- Performance on the heterogeneous BIG-bench is only on par with TS-CV, and the method relies on a set of labeled training tasks.
- Calibration is confidence calibration on a single token and does not address full distribution calibration.
Experimental setup
Datasets: MMLU BIG-bench MRQA
Benchmarks: MMLU BIG-bench MRQA
Metrics: ECE TL-ECE MCE NLL Brier
Baselines: Vanilla Temperature Scaling TS-CV MC-Dropout MC-Augment Elicitation Elicitation-Ensemble CAPE
Builds on: temperature scaling variational inference deep ensembles
Closest papers by meaning
- Calibrating Large Language Models with Sample Consistency (AAAI 2025 · similarity 0.91)
- LitCab: Lightweight Language Model Calibration over Short- and Long-form Responses (ICLR 2024 · similarity 0.90)
- Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach (ICML 2025 · similarity 0.90)
- Calibrated Language Models and How to Find Them with Label Smoothing (ICML 2025 · similarity 0.90)
- Large Language Models Must Be Taught to Know What They Don’t Know (NeurIPS 2024 · similarity 0.89)
- Optimizing Temperature for Language Models with Multi-Sample Inference (ICML 2025 · similarity 0.89)
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.