retrieval

The paper introduces STAM, a state-transition-aware memory framework for LLM clinical agents that tracks patient state changes by separating Active and History memories using semantic retrieval plus typed clinical relations, improving supersession-pair recall from 28.6 to 53.9 on MedMemoryBench over baselines.

ScholarCatalyst introduces a benchmark where 184 lead authors label which papers advanced their completed CS projects, and shows that agentic search and even Claude Fable 5.1-based agents underperform simple embedding retrieval (0.42/0.51 vs 0.48 Recall@20) at retrieving idea-inspiring prior work.

AgSpec 是一个检索式推测解码框架，通过从会话、工作区和全局语料中检索草稿并动态调整草稿长度，在多智能体编码基准上实现了最高 4.37 倍（batch size 1）和 4.76 倍（batch size 16）的生成吞吐提升，优于五种现有检索式草稿器和 EAGLE-3。

Omni-Embed-Mini presents a 0.9B-parameter omni-modal embedding model that maps text, speech, audio, images, video, and documents into a shared cosine space without updating any text-side parameters, using dense distillation from the frozen backbone's own caption embeddings. The 2.3B variant is competitive with Gemini Embedding-2, and all models, code, and data are released.

RPTune is an end-to-end framework coupling learned catalog curation (an encoder-reorganizer pruning and ordering products via downstream LLM feedback) with LLM post-training using catalog-grounded supervision, achieving up to 31.4pp gains in product search accuracy across 7 real merchants.

Mnemon is a memory agent that keeps conversations as raw dated records and divides memory work into System 1 fast judgments by a small decision model (Jev) and System 2 planning by an LLM, achieving state-of-the-art LoCoMo (91.7%) and LongMemEval-S (94.4%) scores at low token cost.

PrimeSeeker 提出一种面向能力的监督框架，通过「潜在锚点推理」将深度搜索分解为锚点解析与关系传递的耦合操作，据此构建网络证据结构、联合生成问题与参考证据骨架，训练出 30B 搜索智能体，在五个深度搜索基准上以更少工具调用取得强性能。

Memorizon decouples long-span supervision from attention in streaming world models by having each scored chunk retrieve a shared bounded top-K latent bank via camera co-visibility, enabling consistent revisits beyond the context window at minimal extra cost. Retrieval improves revisit consistency over a sliding-window baseline by up to 30%, and results show the model genuinely uses the retrieved latents.

MemLife introduces a multimodal memory system for long-term egocentric videos that builds entity-grounded text episodes and retrieves them with a time-indexed agentic reader, improving over training-free baselines by 4.6-12.0% on long-horizon benchmarks. A reinforcement learning framework called MemOpt further optimizes the memory writer for faithfulness and retrievability, yielding consistent 2.7-5.0% gains.

Cohere released Embed 5, its most powerful embedding family yet, with pro and fast variants sharing an embedding space, multimodal support, 100+ language coverage, 128k context, and matryoshka embeddings across multiple dimensions.

这篇论文提出 EviAlign，将语义证据生成与边界读出（Boundary Readout）耦合到共享的多模态大语言模型中，联合训练生成与对比检索目标，在 12 个 MMEB 检索任务上以 500K 训练对达到 76.9 的平均 Recall@1，同时保持单向量索引与打分。

This article tests the application of the Jev model in RAG retrieval, evaluating its effects on accelerating retrieval and reducing costs. Results show advantages in reranking and judging answerability.

Accepted at IEEE ICDM 2026, this paper presents a taxonomy of 5 failure modes when LLM agents run ML research at production scale, along with a multi-agent (Researcher, Code Fixer, Criticizer) prevent/persist/redirect scaffold that achieved 1.82× Recall@6 over a hand-tuned baseline and fixed 14 latent bugs autonomously.

Peerify is a pipeline for automatically verifying peer-review claims against manuscript evidence, using a benchmark of 800 claims from NeurIPS 2024 and ICLR 2024, demonstrating that retrieval-centered verification outperforms entailment baselines.

The tweet from @omarsar0 recommends focusing on agent memory retrieval and announces the opening of the Agent Memory Challenge 2026 Cycle 2, which evaluates memory in AI agents using coding and text tracks.

The paper proposes a zero-inference prospective term for personal memory retrieval that boosts memory items linked to future commitments without query-time computation. It shows improved recall on a synthetic task set and positions this as part of a layered architecture for proactive AI assistants.

AutoViewMem is a data-driven framework that organizes long-term conversational memory into self-configuring semantic views to improve retrieval and personalization for LLM agents, showing superior performance on benchmarks.

V7 Go is an agentic platform that uses AI models to organize business context into a queryable memory for agents, enabling high-accuracy workflows in enterprises such as finance and insurance.

EconSkills introduces a skill library and evaluation framework for web agents to transfer and retrieve procedural knowledge for live economic data retrieval, showing improved efficiency in controlled transfer and competitive performance at library scale.

M-SQE proposes a multilingual skill quality estimation framework to enhance language equality in agentic skill use by evaluating candidates through intrinsic quality and task-grounded utility.