Tag
该论文提出 CrossProfile 诊断基准，研究多智能体讨论压缩为记忆时的选择策略，发现按稀有性或相关性选择均会失败，且决策模型存在从众偏差，完整转录反而降低准确率。
ECHOES 提出一种通过递归历史压缩与特权在线蒸馏来增强长时程 GUI 智能体状态保持能力的方法，无需在部署时保留历史截图即可取得有竞争力的性能，并发布了细粒度诊断基准 ECHOESBMK。
MemoWM 提出一种以世界模型为条件的记忆分配框架，通过共享预测压缩智能体长期经验存储，在五个长时程智能体记忆基准上达到42.42%平均准确率，同时将经验特定存储降低53.9%。研究还揭示了共享模型容量与重复存储成本之间的权衡关系。
MemFold 提出一种通过行为优化固定预算软记忆的方法，将查询条件化的文本记忆压缩为 K 个连续向量，并用组相对奖励与置信度门控的 on-policy 蒸馏信号训练读者模型，在 PersonaMem-32K/128K 上取得最佳准确率并向 PrefEval、LongMemEval 迁移。
WorldPlay2 is an interactive world model that integrates a factorized hybrid control interface with compressed memory and stable distillation to enable real-time responsiveness and long-horizon consistency in control tasks.
This paper introduces Geometry Guided Evidence Preserving Memory (GEM), a training-free compressor for long-horizon agents that optimizes for preserved task evidence over geometric coverage, reducing token usage while maintaining performance.
D-Quant is a flexible KV cache quantization framework using driftable entropy coding to convert variable-length representations into fixed-size bitstreams, enabling efficient parallel dequantization in attention kernels and achieving up to 7x memory reduction and 3.5x throughput improvement while preserving near BF16 performance.
AgentZip is a memory compression system designed for AI-agent sandboxes that reduces memory usage by up to 8.7× by exploiting redundancy across sandboxes and aligning compression with LLM waiting periods.
ReCache is a framework for efficient KV cache reuse and compression in tool-augmented LLM agents, achieving significant speedup and memory reduction while maintaining performance.
This paper investigates how epistemic stance (qualifiers, attributions) survives memory compression in AI agent memory systems. It finds that making the stance explicit as a labelled field improves retention significantly, while merely lengthening the text does not.
This paper shows that a language model with a lossy memory that retains a wrong conclusion but drops the evidence produces confident incorrect answers, whereas an empty memory leads to abstention. The authors propose a source-first compression policy that preserves recomputable sources instead of conclusions to maintain correctability, and demonstrate the mechanism across multiple models and dialogue systems.
Introduces KV-Compression Aware Training (KV-CAT), a method that encourages transformers to learn compressible key-value caches during training, improving memory efficiency for long-context tasks without sacrificing performance.
MemRefine is an LLM-guided framework for compressing long-term agent memory under fixed storage budgets, using similarity for candidate pairing and an LLM judge for factual deletion/merge decisions, outperforming rule-based baselines on benchmarks.
FlashMemory-DeepSeek-V4 proposes a novel inference paradigm called Lookahead Sparse Attention (LSA), which uses a neural memory indexer to actively predict future context needs, compressing physical KV cache usage to 13.5% of full context baseline while improving average accuracy by 0.6%. This method adopts a decoupled training strategy that allows independent training of the indexer without loading the base model, significantly reducing training cost.
OmniMem introduces a modality-aware memory allocation and perturbation-aware selection strategy for streaming audio-visual LLMs, achieving 2-4% absolute accuracy gains over compression baselines on long-video benchmarks.
AURA-Mem proposes a constant-size memory for robot policies using a learned gate that writes only when current observations would change the next action. It matches baseline accuracy with significantly fewer writes and constant VRAM, addressing the memory bottleneck for long-horizon robot tasks.
VideoMLA replaces per-head KV caches in video diffusion models with a shared low-rank latent and decoupled 3D-RoPE positional keys, reducing per-token KV memory by 92.7% and improving throughput by 1.23x on a B200 while maintaining quality on VBench benchmarks.
This paper proposes a unified framework called Efficiency Frontier, which treats large model context management as a deployment optimization problem, jointly modeling task performance, token overhead, and preprocessing reuse. On 5,000 HotpotQA instances, deployment optimization saves 25% of token usage, while memory compression is more than half the cost of full context in high-precision scenarios.
Introduces The Efficiency Frontier, a unified framework for optimizing cost and performance in LLM context management, achieving approximately 25% reduction in effective token usage at comparable performance on HotpotQA.
WorldKV is a training-free framework that retrieves and compresses key-value cache chunks to maintain long-term consistency in video diffusion world generation, achieving higher throughput while matching full-memory fidelity.