Research map / Large language models / Language, safety and interpretability
Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth
Ido Amit, Ido Galil, Ran El-Yaniv
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 |
|---|---|---|---|
| add reasoning (CoT prompting or reasoning-specific training) | paired Instruct/Reasoning models · SALT | Precision: 78.2% | Improved |
| add reasoning (CoT prompting or reasoning-specific training) | paired Instruct/Reasoning models · SALT | ECE: 60.8% | Improved |
| add reasoning (CoT prompting or reasoning-specific training) | paired Instruct/Reasoning models · SALT | AUROC (ranking): -23.3% | Hurt |
| add reasoning, controlling for precision | paired Instruct/Reasoning models · SALT | calibration (ECE): p=0.567 | No clear effect |
| evaluate at atomic-unit instead of line-unit resolution | confidence ranking · SALT | Macro-AUROC: 0.45 to 0.7 (atomic Micro-AUROC) | Hurt |
| use logits-based instead of verbalized confidence functions | confidence functions · SALT | calibration and ranking: p < 0.05 | Improved |
| apply empirical binned calibration (vs Z-score+Sigmoid, Min-Max) | post-hoc calibration · SALT | ECE | Improved |
| apply empirical binned calibration (vs Z-score+Sigmoid, Min-Max) | post-hoc calibration · SALT | AUROC (ranking) | Hurt |
| corrupt/fix prefix atoms (global prefix correctness, no local control) | regenerated target atom · SALT DNA task | P(next correct) slope: +0.42 slope; +0.117 correction vs -0.231 corruption | Improved |
| corrupt/fix last-10 atoms (local prefix correctness, no global control) | regenerated target atom · SALT DNA task | P(next correct) slope: +0.29 slope; +0.063 correction vs -0.238 corruption | Improved |
| corrupting prefix versus correcting it | regenerated target atom · SALT DNA task | P(next correct): |corruption/correction| 2.0x (global), 3.8x (local) | Hurt |
| remove prefix-correctness controls from length-response analysis | answer-context length curve · SALT DNA task | drop in P(correct): -0.43 -> -0.50 | No clear effect |
Closest papers by meaning
- When Should LLMs Be Less Specific? Selective Abstraction for Reliable Long-Form Text Generation (NeurIPS 2026 · similarity 0.90)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.89)
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.89)
- SPANUQ: Span-Level Uncertainty Quantification for Large Language Model Generation (NeurIPS 2026 · similarity 0.89)
- Graph-based Uncertainty Metrics for Long-form Language Model Generations (NeurIPS 2024 · similarity 0.88)
- Improving Uncertainty Estimation through Semantically Diverse Language Generation (ICLR 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.
