Contextual embedding models are typically trained and evaluated on the assumption that a relevant document contains one relevant chunk, but in real use a retriever has to find the answer across several chunks and also surface the evidence needed to verify it. The authors introduce pplx-embed-v2-context, a family of contextual embedding models trained by distilling graded chunk-level relevance from a query-aware context compression model, which removes the need for gold-chunk labels and makes the training target independent of the chunking. Two benchmarks are introduced alongside it: Context-bench, a controlled diagnostic, and Q2D-Chunk, built from real-world search queries and released as a public leaderboard. The models ship in 9B and 0.6B sizes with compatible embedding spaces, trained on a vision-language backbone, and the 9B model sets a new state of the art on both new benchmarks and on ConTEB. The visual retrieval capabilities of the 9B model also transfer from its vision-language base, giving state-of-the-art results among dense retrievers on visual ViDoRe-V3, improving further with granular image chunking. The limitation is that chunking granularity still has to be chosen by the user, and extending chunking to image inputs is proposed rather than shipped.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.