Right Reset: Chunking by Prefix Removal
Quick Answer
This paper shows that The Right Reset (RR) method enhances prefix-removal probing in causal language models, achieving a 47.7% recovery rate of original records compared to 25.9% with BGE embeddings.
Quick Take
This technique minimizes local output disruption across six models and shows that context dependence can signal boundaries effectively.
Key Points
- RR recovers 47.7% of original records as clean units.
- BGE embedding achieves only 25.9% recovery without task-specific training.
- RR minimizes local output disruption across six language models.
- Observed-token likelihood-ratio readout is competitive in some architectures.
- Context dependence provides a boundary signal when surface structure is weak.
DeepSignal Analysis
What happened
The Right Reset (RR) method improves prefix-removal probing in causal language models, achieving a 47.7% recovery rate of original records. This is significantly higher than the 25.9% recovery rate achieved using BGE embeddings. The method also reduces local output disruption across six different models.
Key evidence
- Right Reset achieves a 47.7% recovery rate of original records, compared to 25.9% with BGE embeddings.
- RR minimizes local output disruption across six language models, indicating its effectiveness.
- The technique shows that context dependence can signal boundaries effectively, particularly when surface structure is weak.
Why it matters
The findings suggest that the Right Reset method could enhance the performance of causal language models in tasks requiring accurate record recovery. By demonstrating a significant improvement over existing methods, this technique may influence future research and applications in natural language processing. The ability to effectively signal boundaries using context dependence could lead to more robust models in various language tasks.
What to watch
Paper Resources
📖 Reader Mode
~2 min readAbstract:Removing the left context from a causal language model reveals a useful kind of boundary: an edge where the model processes the same right-hand tokens with little change. We turn this observation into prefix-removal probing and introduce Right Reset (RR), which measures preservation of the right-hand hidden-state trajectory. A dynamic program converts RR edge scores into variable-length chunks. On flattened text formed by concatenating topically similar records after deleting their separators and layout, RR recovers 47.7% of the original records as clean units, versus 25.9% for a BGE embedding boundary, the strongest tested conventional baseline without task-specific model training. The gain persists after rendering and OCR. Passive scores from the same Qwen3-4B layer and direct prompting of a same-scale instruction model perform substantially worse on flattened records. Across six language models, RR-selected cuts also undergo consistently less local output disruption than unselected candidate edges. An observed-token likelihood-ratio readout is competitive in some architectures, indicating that the central contribution is the intervention: context dependence itself can provide a boundary signal when surface structure is weak.
| Comments: | 12 pages, 2 figures, 4 tables. Code, data, and reproduction materials: this https URL |
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2608.04330 [cs.CL] |
| (or arXiv:2608.04330v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2608.04330 arXiv-issued DOI via DataCite |
Submission history
From: Mike Vegeto [view email]
[v1]
Wed, 5 Aug 2026 01:17:12 UTC (67 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.