Retrieval pipelines, vector search, chunking and reranking: how models are grounded in a corpus instead of their weights.
arXiv:2608.20942v1 Announce Type: new Abstract: Motivated by the classical Chan-Vese model and the ability of deep priors to capture complex spatial structures, we develop a segmentation model that l...
The paper introduces the Generative Embedding Benchmark (GEB), which evaluates how much content from an embedding can be recovered by a decoder that only has access to the frozen embedding and a question, without the original image or intermediate features. GEB uses a curated visual‑question‑answering dataset with 1,800 development and 900 test items covering natural images, scene text, and visual documents. Experiments on seven public embedding models show that visual‑only scores range from 28.25 to 33.21, while joint image‑question encoding boosts scores up to 65.56, revealing that generative readout uncovers information bottlenecks not captured by traditional separability‑based benchmarks.
The paper investigates a failure mode in Graph-JEPA, a joint‑embedding predictive model trained on a large scientific‑reasoning graph. Despite achieving high linear‑probe accuracy and effective rank, the learned representation contains almost no usable instance information, as shown by retrieval metrics. The authors diagnose the issue to variance allocation in the objective, propose a repair that restores near‑perfect information recovery, and demonstrate that the problem persists even after repair, highlighting limitations in the evaluation metrics used.
arXiv:2608.20801v1 Announce Type: cross Abstract: While Large Language Models (LLMs) have significantly advanced reranking in recommendation, effectively leveraging item-side information remains chal...
PhotoBench is a new benchmark built from authentic personal photo albums that moves beyond simple visual matching to focus on personalized, intent-driven retrieval. It incorporates a multi-source profiling framework that combines visual semantics, spatial‑temporal metadata, social identity, and temporal events to generate complex queries reflecting users’ life trajectories. Evaluation on PhotoBench reveals two key limitations: a modality gap where unified embedding models fail on non‑visual constraints, and a source fusion paradox where agentic systems struggle with tool orchestration.
arXiv:2607.15218v2 Announce Type: replace Abstract: Large language models (LLMs) increasingly serve as high-level planners for embodied agents, where linguistically benign instructions can become uns...
The paper shows that large language models (LLMs) naturally organize their hidden state manifolds into small‑world networks, enabling efficient multi‑hop reasoning. By converting similarity matrices into unweighted graphs, the authors trace connectivity between distant semantic anchors and find a sharp topological phase transition: deep reasoning layers compress conceptual distances into paths bounded by six semantic hops, while early syntactic layers remain fragmented. The framework is applied to zero‑shot hallucination detection in Retrieval‑Augmented Generation, revealing that factual generations preserve a ~3‑hop structure, whereas hallucinations collapse the topology.
The paper evaluates whether zero‑shot large language models (LLMs) can replace fine‑tuned natural language understanding (NLU) classifiers for intent detection in conversational systems. Experiments on ATIS and CLINC150 show that fine‑tuned RoBERTa outperforms Claude Haiku zero‑shot when ample in‑domain labels are available, but the LLM matches the supervised model on the broader 150‑intent CLINC150 schema. The authors identify three production‑relevant scenarios where LLMs excel—out‑of‑scope detection, robustness to ASR noise, and dynamic per‑deployment schemas—and present a decision framework to guide practitioners. whyItMatters:"The study provides concrete evidence and a practical framework for deciding when to deploy LLMs versus fine‑tuned models in real‑world conversational applications."
PrimeAgentOrchestrator (PAO) is a system that spawns new instances of Claude Code, Anthropic’s terminal-based coding agent, pre‑loaded with relevant memories from a user’s personal databases. PAO queries a PostgreSQL entity‑observation database and a Cloudflare Worker semantic search index in parallel, fuses the results, and injects the compiled briefing into the agent’s filesystem to exploit its auto‑read configuration. The paper reports four months of deployment, detailing three generations of context delivery mechanisms, the failure modes that prompted each redesign, and the engineering tradeoffs involved in integrating heterogeneous memory systems instead of building a unified one.
MedRAGChecker is a claim-level verification framework designed for biomedical retrieval‑augmented generation (RAG). It decomposes generated answers into atomic claims and assesses each claim’s support by combining evidence‑grounded natural language inference with biomedical knowledge‑graph consistency signals. The aggregated claim decisions provide diagnostics that distinguish retrieval and generation failures, such as faithfulness, under‑evidence, contradiction, and safety‑critical errors, and the system is distilled into compact models for scalable evaluation.
The paper proposes a method for automated research‑idea generation that preserves the typed structure of scientific papers by modeling each paper as a small category with typed research entities as objects and asserted relations as morphisms. It introduces a three‑layer algorithm—categorical signature clustering, a functor‑preservation gate, and a six‑axis LLM plausibility judge—to identify cross‑domain analogies that maintain relation chains. Experiments on tens of thousands of papers show the categorical gate filters candidates at a 17:1 ratio while keeping a falsifier rate above 83%, and it logs rejected candidates with detailed rationale.
Nexus introduces a depth‑adaptive KV‑cache splicing and retrieval‑decoupled tool routing mechanism for agentic large language models that reduces the time‑to‑first‑token (TTFT) by decoupling tool routing from the expensive schema re‑encoding step. It uses an INT8 semantic lookaside buffer to select tools via retrieval and generates arguments from a compressed textual signature, maintaining about 89% routing accuracy even as the tool registry scales to 250 tools. Additionally, Nexus can splice compiled schema KV blocks into the live context, repairing the seam with a depth‑adaptive suffix redecode when rotary position embedding drift exceeds a threshold, ensuring output fidelity while achieving up to 1.7× TTFT speedup at moderate depth.
The paper introduces CoVR‑R, a reason‑aware composed video retrieval system that, given a reference video and an edit instruction, retrieves a target video that satisfies the edit. It employs a zero‑shot reason‑then‑retrieve pipeline using Qwen3.5‑27B to generate structured descriptions and dense embeddings for gallery videos, and performs edit reasoning on the query to produce a target‑video description used as the query embedding. The method combines dense retrieval with a TF‑IDF branch over generated texts, fusing the rankings with split‑specific weights, achieving state‑of‑the‑art retrieval metrics on both validation and blind test splits.
The paper introduces AgenticRAG-FP, an interventional benchmark designed to attribute causal failures in agentic retrieval‑augmented generation (RAG) systems. By injecting a certified fault at a specified hop and re‑executing the downstream trajectory, the benchmark evaluates whether post‑hoc diagnostics can correctly identify the fault’s location. Experiments on MuSiQue questions show that coverage‑based diagnosis performs well at hop 1 but poorly at later hops, while counterfactual probes reveal varying diagnostic success depending on propagation depth.
arXiv:2608. 20810v1 Announce Type: cross Abstract: Multimodal information systems increasingly route generated visual content back through the same vision-language index that informed its production, so the output must remain retrievable by the queries it was meant to serve.
This paper introduces an automated one‑shot bird call classification pipeline tailored for rare species that lack large training datasets. By leveraging embedding spaces from large bird classification networks and a cosine‑similarity classifier, the system incorporates filtering and denoising steps to detect calls with minimal data. The approach was validated on simulated Xeno‑Canto recordings and on the critically endangered tooth‑billed pigeon, achieving 1.0 recall and 0.95 accuracy. "whyItMatters":"The system enables conservationists to monitor endangered species with only a few recordings, filling a gap left by existing large‑scale classifiers."
The paper introduces Trustworthy RAG, an evaluation agent designed to detect misinformation and knowledge poisoning in Retrieval-Augmented Generation systems. It combines natural language inference verification, a five-signal poison detector, and a weighted Trust Index to assess the reliability of retrieved content. Experiments on multiple LLMs show high accuracy and precision, with the agent effectively blocking unsafe advice in a secure-coding assistant scenario.
arXiv:2601.21948v2 Announce Type: replace Abstract: Neural visual decoding is a central problem in brain-computer interface research, aiming to reconstruct human visual perception and to elucidate th...
arXiv:2608.20969v1 Announce Type: new Abstract: Multimodal clinical AI is limited by weakly aligned inputs and the absence of domain-specific interpretable representations, particularly when learning...
arXiv:2608.21023v1 Announce Type: new Abstract: Word alignment has traditionally been studied between sentences, but many cross-lingual tasks increasingly require correspondences across full document...