エージェント記憶圧縮におけるKVキャッシュ直接操作の可能性
エージェントの記憶は「要約」ではなく「KVキャッシュの直接操作」で圧縮できるのか、135Mモデルで試した
- エージェントの会話履歴増大に伴うKVキャッシュのコスト増大が課題となっている。
- テキスト要約ではなくKVキャッシュを直接操作するアプローチを論文で検証する。
- 135Mモデルでの実験を通じて効率的なメモリ圧縮手法の可能性を探る。
DAILY FEED · 2026-W43
137件の記事をAIが要約 · 毎朝 7:00 JST 更新
50件を表示
エージェントの記憶は「要約」ではなく「KVキャッシュの直接操作」で圧縮できるのか、135Mモデルで試した
Self-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented Large Language Models
ReflectPO: Reflection-Guided Policy Optimization for Long-Horizon LLM Agents
UniGuard: Multi-View Guardrail Learning for Cross-Interface Safety Generalization
IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems
NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required
REQLOXの評価設計――未確定label・反復安定性・Sourceの追加価値をどう測るか
AI-Assisted Paper on Z-Library Cites a TorrentFreak Article That Doesn't Exist
【KaniVolcanoEngine】LLMによるゲーム開発の比較:Bevyと自作Rustエンジン
Invisibilization as a Source of Probability Bias in Language Models
A Reality Check of Inference-Time Context Augmentation Methods
Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters
RetiJev: Fast and Source-Grounded Ophthalmic Decision Models
Verbalizing Activation Differences with Natural Language Autoencoders
Silicon Samples Match the Answers but Not the Respondents
TRIAD: Benchmarking Omni-Modal Ambiguity in Multimodal Large Language Models
TutorCaseBench: A Case-Grounded Benchmark for Evaluating LLM Tutors in Supporting Students with Challenging Behaviors
What the Specification Did Not Say: Specification-Oracle Gaps in Test-Graded Coding-Agent Benchmarks
IFC-MCQ: Automatically Generating Multiple-Choice Benchmarks for Evaluating LLM Agents on IFC Data Retrieval
Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals
AQL-Lite: Enabling Efficient Knowledge Acquisition for Arabic LLMs via Data-Centric Continual Pretraining
Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable
CiteAnchor: Source-Aware Grounding for Hallucinated Citation Detection and Resolution
Protegrity-PII Technical Report
MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments
SIM-UT: A User Simulation based Evaluation Framework for Medical Triage Chatbots
Rethinking the Evaluation of LLM Fact-Checkers
Meaning Contracts: Local Semantic Geometry in Large Language Models
ULME: Predictive Coupling for Mutually Reinforcing Embedding and Language Modeling
From Clinical Judgment to Checkable Rules: Clinician-Led Guardrails for Mental-Health Chatbots
Direct cueing of word identity through word duration: evidence from Mandarin homophones
WikiOriginQA: Automated Construction of a QA Dataset with Knowledge Origins for Analyzing Cultural Bias
Spectral Smoothing and Spiking Perturbations Reconfigure Capability Profiles and Selective Task Performance in Large Language Models
Compiling Autoregressive LLM Control into Policy-Exact Parallel Inference for Hierarchical RAG
Where meanings converge and diverge: Revisiting linguistic relativity across 81 languages
LitPivot: Developing research ideas through dynamic contextualization within the literature landscape
“あいまいな指示”を入力されたAIは安全なコーディングができない? 香港科技大などが調査
Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
Beyond tree cover: mapping the world’s forest types
Interpretive Variation and Discourse Structure in Detecting Enthymemes in Controversial Political Discourse
baseVultra: Compact Nonlinear Recurrence with Reallocated Capacity
Lighthouse: On Building a Trustworthy Digital Assistant for Exploring History with Civil Rights Archives
Climbing the Ladder: Automating Data Pipelines and Reproducing Baselines in Low-Resource Settings
TauIndianBankBench: Procedure-Aware, Action-Oriented Evaluation of Tool-Using Conversational Agents
FusionEval: Benchmarking Model Fusion for Large Language Models
Measuring Research Progress and Linguistic Inclusion in Low-Resource Languages
Synthetic-to-Real Code-Switched ASR: A Spontaneous Nepali-English Podcast Benchmark
From Perpetrators' Digital Footprints to Behavioral Communities: An Ontology-Driven Framework for Bengali Cyberbullying Analysis--- FULL VERSION
Once a Tree, Always a Tree? Cross-lingual Transfer of Hierarchical Generalization in Language Models
A practical guide to AI image generation, from diffusion to agents