Research map / Large language models / Language, safety and interpretability
Trained on Tokens, Calibrated on Concepts: The Emergence of Semantic Calibration in LLMs
Preetum Nakkiran, Arwen Bradley, Adam Golinski, Eugene Ndiaye, Michael Kirchhof, Sinead Williamson
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 |
|---|---|---|---|
| none (standard pretraining, immediate answer) | base LLMs (Qwen, Llama, Mistral, Gemini) · GSM8K, OpenMathInstruct-2, TriviaQA, SimpleQA | semantic calibration error | Improved |
| increase model size from small to large (ablation) | base LLMs · 4 QA datasets, concise/sentence | semantic calibration error | No clear effect |
| change response style from concise to sentence (ablation) | base LLMs · 4 QA datasets | semantic calibration error | No clear effect |
| post-train with instruction DPO (ablation) | Mistral-7B-v0.1 · semantic calibration (sentence style) | calibration error | Hurt |
| post-train with instruction SFT only (ablation) | Mistral-7B-v0.1 · semantic calibration (sentence style) | calibration error | No clear effect |
| post-train with RLHF/DPO/RLVR instruct models (ablation) | instruct LLMs · 4 QA datasets | semantic calibration error | Hurt |
| allow chain-of-thought reasoning before answering (ablation) | base LLMs · GSM8K, OpenMathInstruct-2 | semantic calibration error | Hurt |
| increase model capability (accuracy) (ablation) | base LLMs in predicted-calibrated settings · 4 QA datasets | calibration error | No clear effect |
| lower LoRA KL gap in predicting own semantic distribution (rank-8 LoRA) (ablation) | Qwen2.5 base 0.5B-14B · GSM8K | calibration error | Improved |
Closest papers by meaning
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.94)
- Multicalibration for Confidence Scoring in LLMs (ICML 2024 · similarity 0.91)
- Calibrating LLMs with Semantic-level Reward (NeurIPS 2026 · similarity 0.91)
- QA-Calibration of Language Model Confidence Scores (ICLR 2025 · similarity 0.91)
- Calibrating Large Language Models with Sample Consistency (AAAI 2025 · similarity 0.90)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.90)
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.
