Research map / Large language models / Language, safety and interpretability
Escaping Mode Collapse in LLM Generation via Geometric Regulation
Xin Du, Kumiko Tanaka-Ishii
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 |
|---|---|---|---|
| apply RMR value-cache damping instead of standard decoding | Qwen3-4B-Base · HEIDEGGER, temperature 0.7 | non-collapse rate: 8% -> 56% | Improved |
| apply RMR instead of standard decoding at low entropy | Qwen3-4B-Base · HEIDEGGER, entropy-locked 1.0 | non-collapse rate: 5% -> 33% | Improved |
| use typical sampling instead of standard decoding | Qwen3-4B-Base · HEIDEGGER, low randomness | non-collapse rate | No clear effect |
| regulate a random value-cache subspace instead of the persistent subspace (ablation) | RMR · HEIDEGGER, non-collapse rate | non-collapse rate | Hurt |
| add RMR to decoding at 1000 tokens | Qwen3-4B-Base · 1000-token generations, temperature-locked | non-collapse rate: 84% -> 99% | Improved |
| add RMR to decoding at 4000 tokens | Qwen3-4B-Base · 4000-token generations, temperature 1.0 | non-collapse rate: 6% -> 99% | Improved |
| add RMR to decoding on non-collapse samples | Qwen3-4B-Base · LLM-as-a-Judge coherence, syntax, progress | judge scores: Deepseek-R1-14B Coherence 8.26 -> 8.27 | No clear effect |
| add weak damping of the history-dependent variable m_t (ablation) | minimal state-dependent IFS model · low temperature toy model | state-space accessibility (dimension) | Improved |
Closest papers by meaning
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.87)
- LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs (ICLR 2025 · similarity 0.85)
- Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs (ICLR 2025 · similarity 0.85)
- Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.85)
- The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation (ICLR 2025 · similarity 0.85)
- A Theory of Time-Sensitive Language Generation: Sparse Hallucination Beats Mode Collapse (NeurIPS 2026 · similarity 0.85)
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.
