One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA

• URL: http://arxiv.org/abs/2606.10572v1
• Date: Tue, 09 Jun 2026 08:36:08 GMT
• Title: One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA
• Abstract summary: Latent Memory is a latent-space memory paradigm that replaces each raw text or image evidence item with a single high-dimensional latent token.<n>It achieves competitive QA performance compared to advanced RAG baselines while consuming 3x to 10x fewer generator tokens.<n>It can also deliver the strongest image-grounded QA performance on WebQA.
• Score: 21.546058294307244
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: External memory effectively grounds large language models (LLMs) and vision-language models (VLMs)-based question answering (QA) in relevant multimodal evidence. However, existing memory paradigms represent each memory item in raw text and image forms, so retrieval-based systems must pass the retrieved text or images to the generation LLMs/VLMs, resulting in high token consumption and storage pressure, making it unaffordable for resource-constrained applications. We propose Latent Memory, a latent-space memory paradigm that replaces each raw text or image evidence item with a single high-dimensional latent token produced by a small compressor LLM/VLM. Rather than retrieving raw evidence for generation, Latent Memory operates in a unified latent representation space: the query is embedded into this space to retrieve relevant latent tokens, and the retrieved latent tokens are directly prompted to a pretrained LLM or VLM for answer generation. To make each latent token simultaneously informative for reconstruction, retrieval, and generation, we train the compressor with reconstruction, contrastive, and distillation objectives in a unified end-to-end manner. Latent Memory is evaluated on seven text-only QA benchmarks (e.g., HotpotQA) and multimodal QA benchmarks, where it achieves competitive QA performance compared to advanced RAG baselines while consuming 3x to 10x fewer generator tokens. It can also deliver the strongest image-grounded QA performance on WebQA. Code is available at https://github.com/zz1358m/Latent-Memory-Master.

Related papers

• MEMO: Multimodal Evidence Memory Organization for Long-Horizon LLM Agents [24.912135848528845]
  Long-running LLM agents rely on external memory to store and reuse information beyond a single context window.<n>Key challenge in agent memory is not only to retrieve relevant records, but also to select necessary evidence under a given budget.<n>Existing memory readout methods mainly use textual or visual forms.
  arXiv  Detail & Related papers  (2026-09-07T13:24:40Z)
• Where to Look and What to Use: Retrieve-Localize-Generate for Long-Term Conversational Memory Question Answering [42.31517185994942]
  RAG enables large language models to answer questions by accessing external knowledge.<n>Existing methods suffer from two key challenges: fragmented evidence scattered across temporally distant sessions, and noisy content within retrieved sessions that triggers the lost-in-the-middle effect.<n>We propose MemLoc, a unified Retrieve-Localize-Generate framework for long-term conversational memory QA.
  arXiv  Detail & Related papers  (2026-09-07T06:32:12Z)
• Dual-Grained Agent Memory and Shapley Context Attribution for Multimodal Agentic Learner [55.80248281761638]
  multimodal large language models (MLLMs) deliver impressive perception yet still falter on scientific and mathematical reasoning.<n>We propose DG-Mem, a dual-grained agentic memory framework that augments a frozen MLLM with a non- externally stored memory built once from training-time rollouts and consulted read-only at test time.
  arXiv  Detail & Related papers  (2026-08-24T13:56:01Z)
• MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models [50.25006399944962]
  Memory is essential for large vision-language models to handle long, multimodal interactions.<n>MEMLENS is a benchmark for memory in multimodal multi-session conversations.<n>We evaluate 27 LVLMs and 7 memory-augmented agents.
  arXiv  Detail & Related papers  (2026-05-14T14:41:17Z)
• Beyond Short-Horizon: VQ-Memory for Robust Long-Horizon Manipulation in Non-Markovian Simulation Benchmarks [96.60530830276281]
  RuleSafe is a new articulated manipulation benchmark built upon a scalable LLM-aided simulation framework.<n>VQ-Memory is a compact and structured temporal representation that uses vector-quantized variational autoencoders.
  arXiv  Detail & Related papers  (2026-03-10T11:13:54Z)
• Panini: Continual Learning in Token Space via Structured Memory [4.979820180013486]
  Language models are increasingly used to reason over content they were not trained on.<n>A common approach is retrieval-augmented generation (RAG), which stores verbatim documents externally (as chunks) and retrieves only a relevant subset at inference time.<n>We propose a human-like non-parametric continual learning framework, where the base model remains fixed, and learning occurs by integrating each new experience into an external semantic memory state.
  arXiv  Detail & Related papers  (2026-02-16T19:58:03Z)
• Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory [89.65731902036669]
  Evo-Memory is a streaming benchmark and framework for evaluating self-evolving memory in large language model (LLM) agents.<n>We evaluate over ten representative memory modules and evaluate them across 10 diverse multi-turn goal-oriented and single-turn reasoning and QA datasets.
  arXiv  Detail & Related papers  (2025-11-25T21:08:07Z)
• Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs [28.807582003957005]
  We present a framework for evaluating the abilities of large language models (LLMs) for tasks that require long-term memory and thus long-context reasoning.<n>We then construct BEAM, a new benchmark comprising 100 conversations and 2,000 validated questions.<n>To enhance model performance, we propose LIGHT-a framework inspired by human cognition that equips LLMs with three complementary memory systems.
  arXiv  Detail & Related papers  (2025-10-31T07:29:52Z)
• From Single to Multi-Granularity: Toward Long-Term Memory Association and Selection of Conversational Agents [79.87304940020256]
  Large Language Models (LLMs) have been widely adopted in conversational agents.<n>MemGAS is a framework that enhances memory consolidation by constructing multi-granularity association, adaptive selection, and retrieval.<n> Experiments on four long-term memory benchmarks demonstrate that MemGAS outperforms state-of-the-art methods on both question answer and retrieval tasks.
  arXiv  Detail & Related papers  (2025-05-26T06:13:07Z)
• Rethinking Memory in AI: Taxonomy, Operations, Topics, and Future Directions [55.19217798774033]
  Memory is a fundamental component of AI systems, underpinning large language models (LLMs)-based agents.<n>In this survey, we first categorize memory representations into parametric and contextual forms.<n>We then introduce six fundamental memory operations: Consolidation, Updating, Indexing, Forgetting, Retrieval, and Compression.
  arXiv  Detail & Related papers  (2025-05-01T17:31:33Z)
• Towards Text-Image Interleaved Retrieval [49.96332254241075]
  We introduce the text-image interleaved retrieval (TIIR) task, where the query and document are interleaved text-image sequences.<n>We construct a TIIR benchmark based on naturally interleaved wikiHow tutorials, where a specific pipeline is designed to generate interleaved queries.<n>We propose a novel Matryoshka Multimodal Embedder (MME), which compresses the number of visual tokens at different granularity.
  arXiv  Detail & Related papers  (2025-02-18T12:00:47Z)
• ReWind: Understanding Long Videos with Instructed Learnable Memory [8.002949551539297]
  Vision-Language Models (VLMs) are crucial for applications requiring integrated understanding textual and visual information.<n>We introduce ReWind, a novel memory-based VLM designed for efficient long video understanding while preserving temporal fidelity.<n>We empirically demonstrate ReWind's superior performance in visual question answering (VQA) and temporal grounding tasks, surpassing previous methods on long video benchmarks.
  arXiv  Detail & Related papers  (2024-11-23T13:23:22Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.