Tag
提出CogMem认知记忆架构，基于PEC²F图谱将对话记忆从被动检索转为主动重构，在LoCoMo和LongMemEval上取得多跳、时序和知识更新任务的良好表现。
Exa introduces ATLAS, a new benchmark for evaluating AI agents on search-intensive, real-world research tasks with verified golden answers and a refreshable automated pipeline. Key findings: no sub-$1 agent run achieved row F1 over 0.5, top agents still miss ~1/3 of golden results, and Exa's search backend defines the cost-performance Pareto frontier.
The paper presents TopoPlanner, a topology-consistent planning framework that lifts tool dependency graphs into cellular workflow complexes to guide LLM tool planning, enabling handling of verification-correction loops, branch merging, and reusable intermediate states beyond sequential or DAG-like workflows, with consistent gains over baselines on four benchmarks.
Perplexity 发布了 PPLX-EMBED-V2-LATE，提供多向量、多模态嵌入，保留 token 级向量并在共享嵌入空间中实现更丰富的文本与图像检索，在 ViDoRe(V3) 等基准上表现领先。
Simon Willison comments on Google's release of EmbeddingGemma 2 under the Apache 2.0 license, arguing that open weights are essential for embedding models so users can avoid being locked into proprietary hosted services that could later be discontinued.
EmbeddingGemma 2 enables fully offline multimodal RAG on-device, running text search in ~191MB RAM and text+image+video+audio embeddings in ~567MB on a Pixel 11 Pro. Open weights are released under the Apache 2.0 license.
This paper introduces 'knowledge accessibility' in LLMs and proposes an accessibility center hypothesis, showing that more accessible queries lie closer to a center in the model's representation space, revealing a knowledge boundary that helps decide when to use query rewriting, chain-of-thought, or retrieval.
APDMem introduces a hierarchical long-term memory architecture for LLM assistants that applies agent-controlled progressive disclosure over four memory layers, achieving strong LongMemEval performance while accessing only 8% of conversation data.
该预印本系统比较了在不同推理信息假设（无化学式、已知化学式、峰值带子式注释）下的最近邻基线方法，用于从MS/MS光谱预测分子指纹，评估了五种谱相似度在NPLIB1与MassSpecGym数据集上的表现。
A systematic study comparing fine-tuning versus RAG-based approaches for LM-based world modeling across five environments, finding fine-tuning often wins (15/20 settings) but RAG is more data-efficient, and proposing a hybrid system combining parametric dynamics with retrieval over an actively maintained memory store that outperforms both.
Qdrant publishes a case study on Lucy, which built a source-of-truth retrieval layer over 2.8M+ SEC and Korea DART filings using hybrid lanes and in-query payload filters, achieving 94.7% Recall@10 on FinanceBench and +0.160 answer accuracy with the answer model held fixed.
An article examining how agent memory systems should distinguish source observations, accepted decisions, derived summaries, and operational checkpoints in a human-readable layer, asking how to draw boundaries between files, relational stores, and vector indexes to prevent stale summaries from outliving their sources.
Kapa.ai introduces the Company Knowledge Bench, a 1,000-case evaluation set built from real production data for benchmarking retrieval on messy internal company knowledge. Their results show frontier models with grep alone (0.61) match tuned hybrid-search pipelines but take 5x longer, while their agentic retriever Kapa Deep reaches 0.65 in about five seconds.
MemFit is a long-term memory system for conversational agents that stores turns verbatim in an append-only store and uses LLM-free insertion and multi-path retrieval to drastically cut memory construction cost and latency while achieving state-of-the-art results on LoCoMo, MemGallery, and LongMemEval-S.
This arXiv preprint develops a theory of context sufficiency for generative AI personalization, proposing a Context-Sufficiency Frontier that distinguishes four states (insufficiency, sufficiency, saturation, interference) and showing via a full-factorial experiment with a generative recommender at a home-furnishing retailer that relevant context improves appropriateness while irrelevant context harms it and destabilizes retrieval.
论文提出 Legal Research Bench（LRB），一个由专家撰写的 413 题开放式美国法律研究基准，并评测 13 个前沿模型的智能体；最强的 Claude Opus 4.8 在全通过率上仅达 42.9%，表明法律研究智能体仍远未达到可靠水平。
This paper builds a streaming-recall benchmark that separates an agent's query-independent retention decision from its query-aware selection decision, showing that much of a prior evaluation's reported memory advantage was actually an access-regime artifact and that recall under bounded memory is upper-bounded by retention rate.
Garry Tan announced that GBrain now supports Jev, improving its remembering/dream cycle while keeping retrieval behavior unchanged.
A benchmark comparing 18 RAG pipeline variants against a simple agent loop on Google's FRAMES multi-hop questions finds the best pipeline hits 78.9% while an agentic loop that can re-search reaches 92.7%, with surprising results showing reranking often hurts accuracy. The authors also note models sometimes fill gaps from memory despite instructions to stick to retrieved documents.
TRACE introduces target-aware retrieval, attributed evidence localization, and contract-constrained extraction to close the 'grounding contract gap' in scientific LitTraceQA, scoring 0.760613 on the official 71-question test set with strong paper F1 (0.9728) but weaker table extraction performance.