arXiv/Article/25/11/Word Salad Chopper
Properties
- logseq-entity
- Logseq/Entity/Article
- date-created
- 2025/11/01
Word Salad Chopper: Reasoning Models Waste A Ton Of Decoding Budget On Useless Repetitions, Self-Knowingly
- arXiv:2511.00536 · Wenya Xie, Shaochen (Henry) Zhong, Hoang Anh Duy Le, Zhaozhuo Xu, Jianwen Xie, Zirui Liu
Summary
- Large reasoning models waste significant decoding budget through meaningless token repetitions — and their hidden states reveal the model is self-aware of the loop.
- A single-layer linear classifier over newline-token hidden states detects word salad during generation with minimal overhead.
- Removing detected word salad tokens and regenerating yields substantial length savings with minimal quality loss; the authors argue this filtering should be standard practice.
Key Quote
Examples in Appendix G
Case 1: Semantic Loop from Unresolved Ambiguity (MATH-500).
- The model begins with valid reasoning but then becomes trapped in a semantic loop — repeating the same confusion without resolution:
‘‘But when I added step-by-step, I got 9997.nn’’
‘‘But wait, 6270 + 3737 is 10,007, so why is the step-by-step adding 3000, 700, 30, and 7 giving me 9997?nn’’
‘‘But why does the step-by-step addition give me 9997?nn’’ (chopped here)
‘‘Wait, so 6270 + 3737 is 10,007...nn’’- WSC detects early signs of degeneration and chops at the third chunk, followed by a regeneration prompt. The regenerated continuation quickly resolves the problem with correct reasoning within a 4k budget.
Case 2: Endless Enumeration without Convergence (MATH-500).
- The model attempts a brute-force enumeration without reaching a conclusion:
‘‘For k=1: ...’’
‘‘k=12: ...’’
‘‘k=14: ...’’ (chopped here)
‘‘k=27: ...’’- Here, WSC intervenes at chunk 318 to prevent further unbounded enumeration, ensuring the continuation remains within budget. This illustrates WSC’s ability to detect degeneration early and prevent catastrophic repetition.
Context
- Defines "word salad" specifically as the token bloat variety — see Word/Salad for the full taxonomy.