Research map / Large language models / Language, safety and interpretability
How Stable is the Next Token? A Geometric View of LLM Prediction Stability
Deyuan Liu, Zecheng Wang, Zhanyue Qin, Zhiying Tu, Dianhui Chu, Dianbo Sui
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 |
|---|---|---|---|
| dTCB-guided prompt enhancement instead of baseline prompt | MMLU ambiguous questions · MMLU AQ set | PDR (%): 30 -> 10 | Improved |
| dTCB-guided prompt enhancement instead of baseline prompt | MMLU ambiguous questions · MMLU AQ set | worst-case accuracy: 0.15 -> 0.30 | Improved |
| dTCB-guided prompt enhancement instead of baseline prompt | MMLU ambiguous questions · MMLU AQ set | average dTCB: 1983.0 -> 2734.0 | Improved |
| dTCB-guided prompt enhancement instead of baseline prompt | GSM8K ambiguous questions · GSM8K AQ set | accuracy: 0.35 -> 0.65 | Improved |
| dTCB-guided prompt enhancement instead of baseline prompt | GSM8K very confident questions · GSM8K VCQ set | accuracy: 0.85 -> 0.92 | Improved |
| co-optimize prompts for dTCB instead of perplexity alone | prompt optimization · MMLU & GSM8K | worst-case accuracy: 0.40 -> 0.65 | Improved |
| co-optimize prompts for dTCB instead of perplexity alone | prompt optimization · MMLU & GSM8K | average accuracy: 0.70 -> 0.72 | No clear effect |
| co-optimize prompts for dTCB instead of perplexity alone | prompt optimization · MMLU & GSM8K | average PPL: 1.9 -> 2.4 | Hurt |
| add misleading clarification '(7 days)' to the question | gsm8k_811 with algebraic ICLs · GSM8K gsm8k_811 | accuracy: 100.0 -> 0.00 | Hurt |
| add misleading clarification '(7 days)' to the question | gsm8k_811 with algebraic ICLs · GSM8K gsm8k_811 | dTCB: 8.20 -> 46.97 | Improved |
| remove ICLs and use algebraic decomposition instruction zero-shot | gsm8k_811 clarified question · GSM8K gsm8k_811 | dTCB: 46.97 -> 49450.23 | Improved |
| cluster competitor embeddings toward the top-1 token with probabilities fixed (ablation) | MMLU prompts · MMLU | fraction where dTCB ordering held: 90% | Improved |
Closest papers by meaning
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.91)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 2025 · similarity 0.88)
- Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models (NeurIPS 2024 · similarity 0.88)
- Brittlebench: Quantifying LLM robustness via prompt sensitivity (NeurIPS 2026 · similarity 0.88)
- SETLEXSEM CHALLENGE: Using Set Operations to Evaluate the Lexical and Semantic Robustness of Language Models (NeurIPS 2024 · similarity 0.88)
- Prescriptive Scaling Reveals the Evolution of Language Model Capabilities (ICML 2026 · similarity 0.87)
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.
