IC-575Four released LLMs (LLaMA-3.1-8B, Mistral-7B, Qwen2-7B, Yi-1.5-9B) can perform in-context learning on continuous vector representations projected into their embedding space, matching or outperforming few-shot ICL across text, time-series, graph, and fMRI tasks

Yufan Zhuang, Chandan Singh, Liyuan Liu, Jingbo Shang, Jianfeng Gao

SourceVector-ICL: In-context Learning with Continuous Vector Representations

The authors project continuous vectors from diverse encoders into the embedding space of four frozen LLMs using lightweight trained projectors, then use the resulting box tokens as in-context demonstrations. Across text reconstruction, function regression, text classification, summarization, molecule captioning, time-series classification, graph classification, and fMRI decoding, the LLMs learn from the projected vector context. With pretrained projectors, vector-ICL matches or outperforms few-shot ICL on 4 of 6 directly comparable tasks; with finetuned projectors it consistently surpasses both few-shot ICL and domain-specific baselines.

Evidence
correlational
Key metric
V-ICL (finetuned) accuracy: SST2 98.16, IMDB 97.28, Emotion 85.20, Financial Phrasebank 81.68 vs few-shot ICL 91.74, 93.50, 55.20, 71.78; V-ICL (finetuned) ROUGE-L: XSum 20.08, XLSum 20.49 vs few-shot ICL 19.53, 19.41
Caveat
Experiments did not cover all possible combinations of encoders, LLMs, modalities, and tasks; only single-token encoders were tested; the effect of instruction tuning on vector-context understanding was not explored.
Model
Llama 3.1 8B, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1, Qwen 2 7B, Yi Yi-1.5-9B
Datasets
IMDB [eval], Rotten Tomatoes [eval], SST-2 [eval], Financial Phrasebank [eval], XSum [eval], XLSum [eval], UCR Time Series Archive [eval], Wikitext-103 [train]
Related findings
IC-576, IC-577
Extraction
automatic-extraction