Tsinghua Unveils Jev-LDE: Edit ICL Examples Once, Boost 1-Shot Accuracy

Researchers at Tsinghua have proposed Jev-LDE, a way to improve in-context learning not by adding examples to the prompt but by editing the ones already there. The edit is made once. After it, the model in 1-shot mode, with a single demonstration, shows noticeably higher average accuracy. Jiqizhixin reported this on October 14, 2026. The available excerpt of the article gives no specific figures for the gain, so we neither cite nor assess them.

What problem does Jev-LDE solve?

The standard ICL setup works like this. The system pulls the few closest demonstrations from an example library, the Top-K, and places them in the prompt alongside the current question. The authors point to a weakness in this scheme: the retrieved Top-K does not necessarily suit the specific target model. Embedding similarity does not guarantee that an example will help solve the task.

There is also a second problem, a financial one. Every extra example increases the context and the cost of inference, and context is paid for on every request. Irrelevant or repetitive demonstrations can also confuse the model.

This is usually treated on the retrieval side: swap the retriever, rerank candidates, tune K. Judging by the title of the work, Jev-LDE shifts the focus to the example itself. If the demonstration is improved, one example can replace a whole set.

What is known about the method

There are few facts, and that is worth saying plainly. Here is what can be stated based on the headline and the introductory section:

• Authors: a Tsinghua team
• Method: Jev-LDE, editing demonstrations for ICL
• Edit frequency: once, not on every request
• Evaluation mode: 1-shot
• Result: higher average accuracy; the size of the gain is not disclosed in the available text
• Source: Jiqizhixin, 14.10.2026

Several open questions determine how valuable the work is. What does the acronym stand for? Who edits the examples: the target model, another LLM, or a search algorithm? On which models and datasets was it measured? Do the edited examples transfer between models, or is the edit tied to one? Without answers, it is impossible to say whether this is a general technique or a tuning for a specific benchmark.

Why this matters in practice

The economics of the idea are attractive. A one-time edit of an example is amortized across the entire stream of requests, and a short prompt reduces both latency and the token bill. For high-traffic products, where a few-shot prompt goes to the model millions of times, this is a significant line item.

Caution is needed in one place. Average accuracy hides variation across tasks: on some sets the gain may be large, on others zero or negative. And everything depends on the baseline. If 1-shot is compared against a raw random example, the improvement looks more impressive than against a well-tuned few-shot setup with reranking.

What to check before adoption

Wait for the full paper and the code. Then look at what the result is compared against: Top-K, reranking, or manual example preparation. Check the cost of the edit itself and how the method behaves on your models and your types of tasks. If the edited demonstrations transfer between models, that is a strong argument for a pilot. If not, the edit will have to be repeated every time you switch models, and the savings will shrink.