arXiv · 17 Nov 2026
From Token Changes to Task Failures: Bit Flips in a Shared KV-Cache Block
Yuji Yamamoto, Satoshi Matsuura
Citations by database
Each database indexes a different set of journals, preprints and books, so they are shown side by side rather than blended.
- OpenCitations0
Citations per year
The current year is still counting.
- 2026 so far0
Abstract
In LLM serving systems with prefix caching, each KV block of a shared system prompt is held in GPU memory as a single physical copy, which every request with that prompt reads. Our earlier work measured change and showed that one bit flip in such a block changes many later responses without making them visibly degenerate. This paper extends that work by measuring how often the changed responses are wrong. Using software fault injection into vLLM serving Qwen3-8B, we flip the most significant mantissa bit or the sign bit of a value in such a block. We score responses on four tasks: code generation, short-answer questions, multiple-choice reasoning, and refusal under a policy. The fraction of changes that turned a correct response wrong depended strongly on the task: 0.023 on MBPP+ and 0.59 on held-out TriviaQA at the mantissa bit, and 0.019 and 0.60 at the sign bit. Harm fell on requests whose sampled answers are sometimes wrong, so the damage of a flip depends on how common such requests are. Harmed responses looked like ordinary errors, so a flip should be detected by checking cached blocks rather than by monitoring outputs.
What the data shows
- No database records a citation yet, as is usual for a paper this new.
Graph
This paper in the middle, its authors and their most-cited papers below, the papers it cites to the left and the papers citing it to the right. The timeline lays every paper out by year.
Select a node for its details. Drag to pan, scroll to zoom, drag a node to move it.
Start with
Cited by 0 papers
Who cites this paper, from open citation data, with Crossref's record of each; the 100 newest when there are more.
References (0)
What it cites, from open citation data, with Crossref's record of each.