SCOUT is a frozen‑encoder approach for sim‑to‑real text‑based person retrieval that predicts cross‑modal embeddings instead of fine‑tuning cross‑encoders. It uses a trainable predictor to map patch tokens from a frozen video encoder (V‑JEPA) into the embedding space of a frozen text encoder (EmbeddingGemma), guided by a bidirectional InfoNCE objective. The method achieves state‑of‑the‑art results on the AI City Challenge 2026 Track 4, with a full retrieve‑fuse‑rerank pipeline reaching 84.25 mAP@10 and a single frozen model alone scoring 60.63, while training costs are modest (≈95 GPU‑hours).
arXiv:2606. 00910v1 Announce Type: cross Abstract: Composed Video Retrieval (CoVR) seeks the target video that results from applying a free-form textual modification to a reference video.
The paper uncovers a systematic mismatch in modern text embeddings: while sentiment can be extracted with high accuracy from frozen encoders, the embeddings’ cosine similarity metric barely reflects polarity differences. Across eleven models from 2021–2026, a linear probe achieves up to 91.3% accuracy, yet cosine similarity margins are only 0.003–0.064, leading to poor retrieval performance (e.g., Qwen3-Embedding-4B’s 64.3 MAP). Introducing a lightweight supervised‑contrastive head (Sentiment‑Injected Sentence Embeddings) dramatically improves probe accuracy, retrieval MAP, and polarity margins, but at the cost of downstream semantic similarity. "whyItMatters":"The findings reveal that widely used retrieval and clustering systems may ignore valuable sentiment information encoded in embeddings, impacting their effectiveness."
The paper introduces a multi‑model retrieval and reranking system for text‑based person anomaly search that bridges the synthetic‑to‑real domain gap. It adapts the Qwen3‑VL‑Embedding‑8B model to synthetic descriptions via contrastive learning and domain randomization, fuses scores from six Qwen3‑VL checkpoints and a frozen SigLIP2 model, and then reranks candidates with a Qwen3‑VL‑Reranker‑8B cross‑encoder. A synthetic distractor‑matched validation protocol is used to tune hyper‑parameters, yielding a 94.78% mAP and 90.85% R@1 on the AI City Challenge 2026 Track 4 test set, placing eighth on the public leaderboard.
Text-based person anomaly search requires a model to retrieve individuals from a large image gallery using natural-language descriptions of both appearance and behavior. This task is challenging becau...
arXiv:2607. 22771v1 Announce Type: cross Abstract: Picking the frozen image encoder for a 3D~CT vision--language model (VLM), together with the token-compression scheme on top of it, is a search over many candidates.
VEDJE is a new video-text retrieval method that compresses video features into a reusable cache while keeping them distinct for each frame. It introduces feature-change prediction as an auxiliary training signal, enabling improved retrieval from the compressed cache without extra cost at query time. Experiments on MSR-VTT, MSVD, DiDeMo, and ActivityNet show that VEDJE boosts R@1 in both retrieval directions, achieving 59.8% text-to-video R@1 on MSR-VTT with a fine-tuned VideoCLIP-XL first stage and maintaining recall when the per-video cache is reduced to 12 KiB.
arXiv:2606. 15134v1 Announce Type: cross Abstract: Vision encoders for retrieval are typically trained with class-label supervision: each training pair reduces to a scalar that uniformly pushes the embedding apart or pulls it together, as if every visual attribute either differed or matched.
arXiv:2605. 11374v5 Announce Type: replace Abstract: Test-time compute is widely believed to benefit only large reasoning models, leaving small models with nothing to gain.
arXiv:2607. 28986v1 Announce Type: cross Abstract: Zero-shot image captioning (ZIC) describes images without paired image-caption supervision during captioner training, relying on text-only corpora and frozen pretrained image-text scorers.
The paper investigates the conflict between image‑based (I2I) and text‑based (T2I) person re‑identification objectives when trained on a shared vision encoder. It shows that optimizing both simultaneously degrades performance because I2I contracts intra‑identity distances while T2I preserves them. A two‑stage pipeline—first training the encoder for I2I invariance and then freezing it while aligning the text encoder—improves T2I results, suggesting a path toward unified ReID systems.
The paper introduces Iris, a framework that extends typed decision‑making models to multimodal settings where candidates can be images, text, or hybrids. Iris preserves key properties such as candidate‑order invariance, scalability, and cache reuse, and achieves high accuracy on a visual‑candidate benchmark while reducing latency through efficient encoding strategies. The authors also demonstrate significant speedups on GPU systems and outperform larger baseline models on MetaEncoder benchmarks.
arXiv:2611.05599v1 Announce Type: new Abstract: Grounding a text query in an image needs a matrix scoring every term against every patch. Vision-language models compute it with an aligned text encode...