Research map / Large language models / Instruction and fine-tuning
Log Probability Tracking of LLM APIs
Timothee Chauvin, Erwan Le Merrer, Francois Taiani, Gilles Tredan
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 |
|---|---|---|---|
| use a longer prompt (33 tokens instead of 1.5 tokens) (ablation) | Logprob tracking · TinyChange | ROC AUC: about 1% AUC difference, baseline around 91% AUC | No clear effect |
| use logprob tracking instead of MET | change detection · TinyChange | ROC AUC: 0.670 -> 0.915 | Improved |
| use logprob tracking instead of MMLU-ALG | change detection · TinyChange | ROC AUC: 0.878 -> 0.915 | Improved |
| use logprob tracking instead of MET (cost) | change detection · TinyChange | cost per year: $146 -> $0.14 | Improved |
| use logprob tracking instead of MET (sensitivity, weight pruning) | change detection · TinyChange weight pruning | threshold pruned fraction for AUC 0.9: 2^-1 -> 2^-10 (or below) | Improved |
| use logprob tracking instead of MMLU-ALG (sensitivity, weight pruning) | change detection · TinyChange weight pruning | threshold pruned fraction for AUC 0.9: 2^-4 -> 2^-10 (or below) | Improved |
| detect LoRA fine-tuning instead of full fine-tuning | Logprob tracking · TinyChange | ROC AUC: L2 distance 1.3 x 10^-5 -> 4.3 x 10^-9 (1 step) | Hurt |
Closest papers by meaning
- Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking (NeurIPS 2026 · similarity 0.83)
- ContractBench: Can LLM Agents Preserve Observation Contracts? (NeurIPS 2026 · similarity 0.83)
- Monitoring Violations of Differential Privacy over Time (NeurIPS 2026 · similarity 0.81)
- BaxBench: Can LLMs Generate Correct and Secure Backends? (ICML 2025 · similarity 0.81)
- From Judgment to Interference: Early Stopping LLM Harmful Outputs via Streaming Content Monitoring (NeurIPS 2025 · similarity 0.81)
- Optimal Aggregation of LLM and PRM Signals for Efficient Test-Time Scaling (ICLR 2026 · similarity 0.81)
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.
